TL;DR

• I built my own AI-assisted knowledge management system — I call it AKM — using Obsidian, Claude, and a custom 3x3 grid. It’s not a Second Brain; I’m still doing the thinking.

• This article walks through why I built it, how it’s structured, what Claude actually does versus what I do myself, and what’s still missing from it. You can build it yourself using this guide.

• Premium members get the full system: four copy-paste Claude prompts that build your own 3x3 grid, your own starting themes, your own vault structure, and a scheduled check that keeps it all honest over time — plus an installable Claude Skill that files your writing into it automatically from then on.

• None of it requires copying my exact grid, my themes, or my setup. Every prompt walks you through defining your own, based on what you actually write about.

• You’ll need Claude Cowork (Pro plan or higher) for the vault-building, filing, and scheduled-check steps, since those write real files and folders on your computer.

Three decades ago, as I started working at a law firm in the IT department, there was a lot of talk about Knowledge Management. From Lotus Notes to Microsoft Exchange, and many other options.

I’ve seen how a global organization, Siemens, shared best practices and other info globally. It was amazing! Can you imagine hundreds of offices from all over the world united by a common repository of information? Thousands of knowledge workers looking for what made some marketing campaigns succeed, others looking for engineers with extremely specific skills around the organization.

But after some years of experience, I noticed that finding the right information was always painful. It was a constant issue to motivate the teams to keep updating the Knowledge Base.

Jumping 30 years to the present, we finally have a unique opportunity to solve most of the issues of Knowledge Management using AI.

I don’t want to call it a Second Brain

There is a lot written about Second Brains. The term is most closely tied to Tiago Forte, whose book Building a Second Brain frames it as a system for capturing, organizing, distilling, and expressing knowledge outside your own head, so nothing gets lost and your thinking compounds over time. Nick Milo pushed the idea further into note-linking with his “Linking Your Thinking” approach. Both trace back to Niklas Luhmann’s Zettelkasten, the slip-box method he used to produce a lifetime of scholarly work by letting notes connect to each other instead of sitting isolated in folders.

When I read “second brain,” I imagine a machine thinking for me, using my ideas, writings, and instructions. But I’m sincerely trying to do the thinking myself.

There are two Substackers I enjoy reading, Jenny and Nate, and they have different ideas about Second Brains. As I understand it, Jenny treats it as infrastructure, a shared vault that multiple AI tools read and write to automatically, complete with dashboards, scheduled jobs, and skills that update themselves. Nate does something different: he argues the classic second brain fails because it asks too much of you in the moment, so his fix is a single frictionless capture point where AI does the interpreting, classifying, and routing for you, with almost no filtering on your end. Both are impressive systems. Both also ask me to trust a lot of AI judgment I have no easy way to check.

I call it Assisted Knowledge Management, or AKM. And in this article, I’ll show you how I did it for myself and what is missing in my system.

What did I want to achieve?

Initially, the idea was to use this Assisted Knowledge Management (AKM) system for my Substack articles and notes. I wanted to include all the notes and posts I ever made.

I also wanted to solve a problem I had with Notion. Claude can use the Notion MCP to directly interact with it. I had a few Claude Skills running to update Notion databases and sometimes make queries, like for example: What were the 5 latest Substack notes I made about Digital Sovereignty? The results were a disaster: slow, unreliable, and the process would easily consume my Claude credits very fast.

Something else I wanted to do was to find modern standards that could be used to manage my knowledge in the age of AI. I found Google Open Knowledge Format (OKF) and also Karpathy’s Wiki gist.

Both of them complement each other and would help me make it easy to share information with others if desired.

Karpathy’s gist describes a workflow: use an LLM to build and maintain a growing personal wiki, one markdown file per idea, letting the model do the drafting and cross-referencing as your knowledge grows.

Google’s OKF takes that same pattern and formalizes it into an actual specification, defining the exact file structure, YAML (a data format built for storing structured information, not for styling) frontmatter, and conventions needed so a wiki built this way can be read, shared, and reused across different AI tools without custom translation. Karpathy gave me the why and the rough shape; OKF gave me the how, a concrete standard to build against instead of inventing my own conventions from scratch, which also means anything in my vault is already in a shape others could read or adopt if I ever wanted to share it.

But something that bothered me is giving AI too much power to decide how to categorize and store my information. I remembered a system I built in the past that could help me manage large amounts of information in a cohesive manner. I call it the 3x3, and it is a matrix for organizing themes:

The intersections, I call them cells, hold the pieces of knowledge relevant to them. And that gives me the power to organize info and access it easily.

And the last and probably most important thing I wanted to achieve is to store all my information locally. That is a very important part of my process towards Digital Sovereignty.

At this point, my Substack articles were scattered in Substack, chats with AI, and even different word processors. Finding the knowledge I produced was as hard for me as it is for anyone else, and that suggested I had poor control over it.

Obsidian is the app that can help me solve this need. It stores MD files, a format that AI is very familiar with, and it is accessible by Claude.

Today, and for some weeks now, I produce knowledge with the assistance of Claude and let the AI write, store, and categorize every single piece into what Obsidian calls a vault.

But wait, it is so much more than having a local storage of files with perfect categorization. It is searchable and accessible by Claude as if it is its context for your conversations.

At this point, Claude consults my local files instead of having to search online; that’s much faster, more efficient, and cheap.

Let’s talk about Obsidian

First of all, it is free to use and download (obsidian.md).

It can be configured to work like your personal Wikipedia, and that means that your files can interlink to each other. For example, if many files share the theme “Personal Development,” when you click that theme, you will be able to see a list of all the articles under the same theme. This is not an automatic feature, but it is part of the AKM system I built; otherwise, the search function would easily list all the documents in your vault corresponding to your keywords, tags, themes, and any other attribute on your documents.

Your vault is a folder in your local file system. Something that I did was to include that folder to synchronize with my iCloud; in that way, I have backed up all my files. You can do the same with any other online storage service.

As I mentioned earlier, Claude is in charge of creating files and maintaining the categorization, interlinking, and keeping score of all changes. So, I have to make it clear to you because I found it amazing: I have never written one letter in Obsidian; it is fully managed by Claude.

Here are some screen recordings of how my Obsidian looks and how it works:

And what about Notion?

Previously, my workflow was roughly like this:

Today my workflow is like this:

So, that means that Notion today is more of a backup space, and no longer a database I use to query or research.

Additionally, you can probably see that my current workflow is more focused on me writing by myself. I made some new skills to coach me to write better, instead of writing better for me.

The solution architecture

First, I defined the content types I wanted to work with: Substack Articles and Substack Notes. Because both document types are in Substack, their metadata is very similar. Now that is important, because each MD file in my Obsidian will have the following structure: Metadata (properties), Content, and themes.

Second, I needed a list of Tags I will call Themes from here on that I would use to categorize the content I create. For example: fearlessness, AI, personal development, etc. These themes will be the glue that unites the entire body of content. With that, I will be able to see which themes are most prevalent in my writing and thinking, and it will also help Claude to reach the information I need it to use as context in chat, research, and even app development.

Third, I needed an additional layer of filtering, and for that I will use the 3x3 I showed you before.

The intersection between columns and rows is called a cell. The column categories are: Seeing (S), Living(L), and Wisdom (W).

The row categories are: Individual (I), Community (C), and Global (G). A cell, for example, is CL or Community + Living, and that would categorize documents that are relevant to practical action in the context of groups or collectives. As you can see, the 3x3 matrix actually adds a layer of intelligence to the system.

Summarizing, now you can store two types of documents; they have properties or metadata that will help you search and categorize. You also have a way to add a set of cells to define the scope of each piece as an additional layer of intelligence.

But what happens if you need a new Theme? Or what if you would like to send your content for backup to Notion, for example?

So here is where you will need some scheduled tasks. Let’s focus on the new theme situation:

Adding the theme would be easy for Claude, but remember that you have 245 other files already in your vault that will need to be reassessed and matched. That task will take some more time for Claude, so I chose to make it a scheduled task that can run once a week.

In the case of my backups to Notion, the scheduled task will take any new note or article I publish and copy it to a Notion database every day.

At the end of the article, you will find how to access the prompts you can use to recreate the system for you using Claude.

Subscribe now

So, now you understand how the knowledge is stored and categorized, and you also know how Claude can help you efficiently maintain the structure in your Obsidian vault.

But there’s one piece I haven’t explained yet: who takes care of the vault structure?

As I said earlier, I never ever write anything in Obsidian itself; I write my drafts using a word processor or Claude desktop directly.

To do that, I built two Claude Skills, which I call coaches. One helps me with Substack Notes, and the other with Substack articles. Those skills are instructed to interact with my Obsidian vault, and they complement the scheduled tasks I mentioned earlier.

Each coach skill also keeps the vault’s structure current, not just the individual file. Every time a piece is created, updated, or published, the skill updates the index pages tied to it: the main Index, the page for each theme it belongs to, and the page for each 3x3 cell it’s classified under. That way the vault’s navigation pages are never stale; they reflect exactly what’s in the vault at that moment, without me having to rebuild them by hand.

If you are curious about why I call these skills coaches, it is because I got tired of the previous Skills that helped me write. Why? Because they used formulas and tactics I got from experts to write effectively, and the result was that my texts didn’t feel like mine anymore. They expressed the ideas correctly, but what about my Spanish-to-English mistakes? What about the order I usually use to present my ideas? The old skills sacrificed me for a better text. So the coaches today just challenge my writing in terms of what I say explicitly, what I communicate implicitly, and also evaluate whether my text includes the reader’s imagination and participation.

But I found something I’m missing

The entire system works so well that I started asking myself, why not include other document types? For example, offers, meeting summaries, ideas, etc. And that is my next big project with AI; lots of changes will be needed and I will have to evaluate if the 3x3 should be redefined, if the current themes are relevant, or if I should keep those new document types in separate Obsidian vaults with parallel systems. What do you think?

Leave a comment

So how is Claude doing?

After already a couple of years working with AI LLMs, a level of expertise is materializing. I have a feeling of the kind of work LLMs can do easily, quickly, and with low risk.

Many people use LLMs for mathematical operations and other deterministic tasks, but in my opinion, those are better addressed by different technologies. I lean toward using Large Language Models specifically for what they’re built for: language.

So, in my experience, I see that LLMs like Claude are very good at interpreting, categorizing, and executing pre-scripted tasks. The results are not perfect, but it is better than if I do it myself. Of course, my criteria and expertise are something Claude can’t replace, but if I find an error or suggestion, updating the system to respect the additional criteria is usually easy.

So I take Claude similarly to working with a human assistant, imperfect, well-intentioned, and always improving. Therefore, I delegate tasks to Claude that are low risk for me. I do not care too much if one article I write misses one theme or keyword; the effect of that isn’t dramatic. In contrast, I’m not willing to trust AI with tax reports or cash flows without the proper combination of deterministic tools and layers of revision.

I take from Jenny and Nate’s way the awesome idea of keeping AI tools participating in the process of documenting and generating valuable information.

’s work inspires me especially because she shows us the best potential of using AI, but I still think implementing a system like hers would require too much attention from me.

Actually, we engaged in a conversation about this opinion, and she came up with a great article I’d like to share with you: AI Scheduled Tasks: The Complete Map.

This is the system I made for you

Let me tell you that writing this article has been a long process for the right reasons. I found some errors in my own system that are now fixed, and reviewing all that helped me find a way to build a sequence of prompts and a skill that will create a custom AKM for you. I removed parts of my system that are not relevant to you, and the prompts are ready to guide you step by step.

I have used and tested Claude Cowork for this AKM. You will need to have a paid tier of Claude to make it work. Why? Because you need Claude Cowork to access your local file system.

I haven’t tried or used ChatGPT for this solution, and I do not know if their new Cowork-like features can support such an implementation, but feel free to use the prompts and skill and ask ChatGPT if there is a way to adapt them to it.

One more thing: all the prompts and the skill itself are plain text; read them before you use them. You might get some new ideas or future edits. You will understand so much more when you see how Claude Cowork functions, so take this as a way to strengthen your AI skills.

You’ll find each prompt below as its own copy-paste block — copy the whole thing and paste it as a message in Claude. The skill works a little differently, since it has to be installed rather than pasted: I’ll walk you through that when we get there.

This is a description of the prompts in sequence, so you understand what their function is:

1. Prompt - Define Your 3×3 Grid
Run this first, in any regular Claude chat — Cowork isn’t needed yet. It walks you through my grid, then asks whether you want to use it as-is or define your own three scopes and three depths based on what you actually write about. Before it’s done, it tests itself: Claude classifies a paragraph you write, you tell it whether it got it right, and you repeat that until you trust it. Saves as Grid.md.

2. Prompt - Define Your Themes
Run this right after, in the same chat. This is where you define your tags — what I call themes — starting with 5 to 10, each with a definition in your own words. It checks every new one against the list so far so you don’t end up with three themes that mean the same thing. Saves as Themes.md.

Once these first two are done, download or save Grid.md and Themes.md somewhere you’ll remember — Desktop or Downloads is fine. You’ll want them on hand for the next prompt.

3. Prompt - Build Your Vault Structure
This is where Cowork becomes required, since this prompt writes real folders and files on your computer. Point it at a new, empty folder, hand it Grid.md and Themes.md, and it builds the actual vault: a folder for your content, an index page for every cell and every theme, a page documenting the whole system in plain language, and a log. It also walks you through opening that folder as a vault in Obsidian once it’s done — so have Obsidian installed before you start this one (it’s free, at obsidian.md). You can run this in the same chat as the first two prompts, or open a fresh one — if Claude can’t find Grid.mdand Themes.md on its own, it’ll just ask you to paste them in.

4. Skill - The “File to My Vault” Skill
Not a prompt — a Claude Skill you install once, and it’s on automatically after that. Give it something new you’ve written, or point it at an existing markdown file, and it classifies it into a cell and a set of themes, writes it into your vault correctly, and keeps every index page in sync. It never touches the words you actually wrote.

5. Task description - Set Up Your Scheduled Vault Check
Run this last, once you’ve filed at least one thing with the skill. It sets up a recurring Cowork scheduled task that catches drift — links or counts that fell out of sync — and flags any recently filed piece whose themes look like a forced fit, suggesting a new one instead of leaving it mismatched. It never adds a theme on its own; suggestions wait for you to confirm them.

If you would like to upgrade your membership now, this is a good opportunity; the prompts and skill after the button are only visible for premium members:

Subscribe now

Read more

We talked about everything: her job, mine, and of course, we ended up talking about what we see AI is changing in the world.

She told me a story that means a lot. She participated in a conference where experienced business owners talked about best practices and AI. Here it goes…

Subscribe now

I think my friend caught the participants unprepared with her questions. Her pivot to the narrative blew me away!

This is the picture:

A seasoned businessman explaining to the audience how he is using AI to change the strategic approach of his company. He used all the words showing he knows what he was talking about.

His focus, unsurprisingly, was on decreasing costs and increasing margins. He even anticipated fewer employees.

Then my friend could not resist the temptation of interrupting his enthusiasm with a few questions:

He responded:

She interrupted him:

He was so into AI that he could not even understand the questions my friend made.

Are we able to see the real potential of AI for business use just by chatting with ur favorite LLM?

It seemed to me like the AI narrative is so powerful that many of us jumped too fast and assumed, imagined, its outcome, its potential, before we even understood where the AI tech is in its lifecycle.

My friend told me that one of the risks of using too much AI for a business is to lose their touch, the skills and relationships they built over time.

That made me remember what I said to a consulting customer a few days ago:

Knowledge belongs to the company, and now it is cheap to get, but personal skills are so much more valuable and belong to the people.

Have you heard the word Enshittification? As far as I understand, it describes a decline in the quality of digital platforms and services. Do you think AI can play a role in the Enshittification of business practices?

Leave a comment

I wish you a peaceful rest of the week,

Jose from Automato

I couldn’t find the source of that quote, so I have no clue if Musk is really behind it. Karo later confirmed that Musk isn’t the author of the text, but the text seems very interesting for a visionary conversation.

Karo’s question changed my day:

“Are you already transforming your time into scarce assets, or are you still waiting for it to become more obvious?”

That question made me remember all the conversations I had with my friends about the future of humanity after Artificial General Intelligence is achieved. I’m sure you’ve had similar deep dives, and I’d love to know what you think about the future of money.

Leave a comment

I’ll spare you the philo-babble and just tell you the end of a “domino effect” my friends and I envisioned:

Under that scenario, money would no longer function as a source of livelihood, reputation, or validation, nor as a unit of measurement.

I wonder what Musk would do with so much money invested in markets, properties, and infrastructure.

Does chasing more money still make sense in a world where technology does the work for us?

Since my first test of ChatGPT, I’ve envisioned the end of work as we know it. Now I can fathom the end of the economy as we know it.

That vision is what drives my work now. And I’m glad to see my friends adopting it in their own way; from how we produce and consume energy to gardening, to learning crafts, and investing in property.

And to directly answer Karo’s question, yes, I started a few years ago transforming time into scarce assets for my family and me. Even my use of AI is helping me express myself with more confidence, a sense of independence, and lately adding a bit of artistic inspiration.

Love,

Jose from Automato

Subscribe now

I had an idea sitting in my drafts: write something about open-source LLMs, their licenses, and what sovereignty actually means when the model underneath you isn’t yours.

Then published a field guide to open-source LLMs that did my homework better than I ever would have.

So instead of writing my version, here’s hers, plus the piece I still think is missing.

Build to Launch

Open-Source LLMs in 2026: What's Actually Open, and How to Pick One

Open-source LLMs are catching up to Claude and GPT. Most of them are not how you think as “open source”. Picking one has less to do with the model than with the license it ships under and the host you run it on. This guide maps these LLMs landscape: who ships open models now, why they give them away, what they cost, where they run, and when to reach for one instead of a closed model…

Read more

2 months ago · 15 likes · 1 comment · Jenny Ouyang

Leave a comment

Here’s where it rhymes with something I’ve been circling myself.

A few weeks ago I wrote about Apple’s WWDC announcement and the Fable 5 export-control episode — not from the license-and-hosting layer, but from one level up:

What happens when the infrastructure you depend on can be switched off or reshaped by a decision you had no part in?

Apple moving inference onto the device wasn’t a UX choice so much as a sovereignty move; a way of not needing anyone else’s permission to keep running. Jenny’s piece and mine are looking at the same fracture from opposite ends.

Read together, the two questions answer each other. Her license-and-host checklist is exactly what you need to survive the day the ground shifts; and my sovereignty question is exactly why the checklist matters.

Open weight isn’t a gift. It’s leverage, held by someone else, until you understand it well enough to hold some of it yourself.

If you’re building anything on an open model right now, read Jenny’s piece before you pick one. Then ask yourself the harder question: not just whether you’re allowed to use it, but whether you’d still be fine if it disappeared tomorrow.

Leave a comment

Where the Models Actually Stand

So I ran the current crop through my own criteria.

Can I understand it, control it, and keep evolving with it, without staying permanently dependent on whoever built it?

That splits into four questions I ask of any model before I’d build on it:

• Can I see what it was trained on, or am I trusting a black box?

• Can I run it without asking anyone’s permission?

• Is that permission permanent, or does it expire once I succeed?

• If the vendor disappeared tomorrow, would I still be fine?

Run today’s open models through that, and they land in four tiers. (July 2026).

The honest answer, today: sovereignty in the full sense exists in exactly one model. Everything else is sovereignty with an asterisk, sized to how big you’re allowed to get before someone else takes the choice back.

But that’s my criteria, built from my own decades of watching vendors change the rules once you depend on them. What would you ask from a model before you built your business on it?

I wish you a great day!

Jose from Automato

Subscribe now

TL;DR: When working with AI, own your work, your process, and build for platform independence. In this article, you'll find my research, the tools I found, the concepts we all need to master, and the standards, tools, and technologies that exist to help us achieve Digital Sovereignty. Especially useful for non-tech people looking to level up their AI skills. A warning: I'm overly critical of US AI models like ChatGPT and Claude — but don't let my opinion spoil the value this article shares.

As you know, I’m worried about the AI zeitgeist. The market buzz gravitates around commercial models like ChatGPT and Claude, and that limits the AI panorama for the majority of users, especially the non-tech among us.

I have been talking about Digital Sovereignty for a few weeks now, so I decided to research what other options we have to keep learning to use AI as a tool that genuinely benefits our business, our professional growth, our learning process, and our quality of life.

In this article, I’ll show you what I found. To be honest, I didn’t know where to start.

What I found, with the help of some friends, is useful to everyone, and it might hint at a few business opportunities that are still waiting to be built.

Who gets the best of AI right now?

I honestly believe IT professionals, developers, and software engineers are the ones getting the most out of AI today. And to my surprise, they’re using multiple LLMs for different parts of their workflows, some local, some hosted, mostly open-source tools, and yes, Claude and ChatGPT too, mainly for coding.

It's genuinely impressive.

So I mapped the whole thing: 7 tables covering models, interfaces, workflow tools, local runners, knowledge management, APIs, and the developer ecosystem. All linked at the end of this article.

Subscribe now

What is the real problem with commercial US LLMs?

OpenAI, Anthropic, and the others are bound by US law — and AI is now being treated as a strategic national security technology. The entanglement between these companies and state power goes further than most people realize: Pentagon contracts, military operations, targeting systems. I’ve been writing about this separately, and I’m not going to apply a lighter standard to Chinese models either — open weights don’t automatically mean political neutrality.

The full story, with quotes and sourcing, is here: Who Really Controls Your AI? — The US and China Question.

The short version: sovereignty isn’t one thing. It’s at least two — who controls the infrastructure, and who shaped the worldview. Nobody is winning cleanly on both. Keep that in mind as you read the rest of this.

Share

Ok, so what do we get if we decide to step away from US LLMs and tools?

First, we need to understand something important:

And yes, I know ChatGPT and Claude, in particular, are extremely good at shipping new features that make our imaginations fly.

Researchers studying AI adoption have found that this kind of dependency is categorically different from traditional software lock-in. The dependency isn’t just functional; it’s cognitive.

When you build your work around a specific model’s behavior, its tone, its reasoning patterns, the way it responds to your prompts, you develop implicit knowledge that doesn’t transfer easily to another tool.

Build your AI tools, for your business and for yourself, with independence in mind.

The other thing we get is access to open-source models and tools. But we need to learn more about what that actually means.

Share

What “open source” actually means, and why it matters to us

Before looking at specific tools, there’s a distinction worth understanding: the difference between open weights, open source, and self-hostable.

Closed source means a company runs the model on their servers, you send your data to them, and you have no visibility into what happens with it. Claude and ChatGPT fall into this category.

Open weights means the mathematical parameters of the model are publicly available. You can download them, study them, and run them on your own hardware. This is the most important form of openness for sovereignty purposes.

Open source goes further — the training code, data pipeline, and methodology are also public. Fewer models meet this bar fully.

Self-hostable means a model small enough to run on consumer hardware — a laptop or a modest server — without needing a data center.

The sovereignty insight is simple:

But “open source” doesn’t automatically mean free or safe. A model with open weights hosted on a Chinese company’s API still sends your data to their servers. The weights are open; your usage is not.

Leave a comment

The landscape — what actually exists

The closed models: familiar but bounded

Claude (Anthropic, USA) and ChatGPT / GPT-5 (OpenAI, USA) are the tools most knowledge workers already use. They’re genuinely the most polished, the best integrated into professional workflows, and the most studied for safety.

The trade-off is complete dependency, no local option, no open weights, full data exposure to US companies operating under US law, investor pressure, and now, apparently, defense contracts.

Gemini (Google, USA) sits in the same category, closed, cloud-only, tied to one of the world’s largest advertising businesses.

The open models: what you didn’t know existed

Gemma 4 (Google DeepMind, Apache 2.0) — Google’s open-source family, released April 2026. The irony is striking: the same company whose closed Gemini model collects your data also releases a fully open model you can run locally with zero data sent anywhere. Gemma 4 is multimodal, powerful, and available on Ollama and LM Studio with a few clicks.

Meta LLaMA 4 (Meta, USA, custom license) — the most widely deployed open model family in the world. Excellent quality, enormous developer ecosystem, downloadable and locally runnable. The license has one restriction: companies above 700 million monthly active users need special permission. For a knowledge worker or small business, this is irrelevant.

Mistral Large 3 (Mistral AI, France, Apache 2.0) — the only serious Western open-source contender. A Paris-based company releasing fully open models under the most permissive license available. It doesn’t win every benchmark. But it’s European, auditable, and self-hostable — which matters differently than winning a benchmark.

Then there’s the Chinese wave — and it’s worth paying attention to:

DeepSeek V4 (backed by High-Flyer Capital, China) — the model that shocked markets in early 2025 by matching frontier US models at a fraction of the training cost. Open weights, locally runnable, genuinely capable. The sovereignty question here is nuanced: the weights are open, but using their API sends your data to Chinese servers.

Qwen 3 (Alibaba Group, China, Apache 2.0) — Alibaba’s open model family, consistently at the top of open-source benchmarks. Available locally. Free to download and run.

GLM / Z.ai (Zhipu AI, China) — Beijing-based, state-adjacent, but with open weights and a free interface. A capable alternative for general tasks and research.

MiMo V2.5 Pro (Xiaomi, China, MIT license) — yes, the smartphone company. Xiaomi released a trillion-parameter open-source model in April 2026, under the most permissive license available. The fact that the company making your $200 phone is now shipping frontier AI that competes with Claude is, in itself, a sovereignty argument.

Kimi K2.6 (Moonshot AI, China, modified MIT) — the most interesting model for knowledge workers specifically. Moonshot has built Kimi Work, a platform that integrates document creation, research, spreadsheets, presentations, and web research into one interface powered by an open-source model. It’s the closest thing that currently exists to a knowledge-work equivalent of the developer tools dominating the open-source AI conversation.

MiniMax M3 (MiniMax, backed by Alibaba, China) — strong models, but worth noting that recent versions have moved away from fully open weights toward API-only access. A reminder that “open source” is not a permanent promise.

Leave a comment

What is the opportunity I mentioned before?

Almost everything being built right now is for developers.

Developers today have purpose-built platforms that give them access to multiple open models through one interface, a cockpit designed for their kind of work.

OpenCode, for example, bundles access to GLM, Qwen, Kimi, MiniMax, and DeepSeek for $10 a month, all in one place. Tools like GitKraken complete the picture, giving developers visual control over their entire codebase alongside their AI workflows.

There is no OpenCode for writers. There’s no equivalent for consultants, researchers, or independent professionals who want to chain multiple AI tools together for non-coding tasks.

The closest things that exist for knowledge workers are Poe (Quora, USA — a multi-model interface that lets you switch between Claude, GPT, Mistral, LLaMA, and others) and Perplexity (Perplexity AI, Inc., USA — an answer engine combining web search with multiple models). Both are useful. Neither is open. Neither is European. Neither gives you true sovereignty.

models.dev is worth bookmarking as a reference — an open-source database that lets you compare every major model across pricing, context limits, open versus closed weights, and capabilities. Not a tool to use, but a map to consult.

Subscribe now

When was the last time you heard about n8n?

I find less and less new content on good practices for orchestrating agents for non-tech people. And Claude’s marketing machinery has convinced me, almost successfully, that I can do nearly anything using Claude exclusively.

So I started wondering: how can I use some of the best practices that advanced AI engineers rely on, for my own non-technical work?

Can I integrate other tools for managing agents?

Workflow automation tools connect applications, trigger actions, and chain outputs so that one agent’s result feeds the next.

For developers, this ecosystem is rich: LangChain, CrewAI, AutoGen, n8n with AI agent nodes. For knowledge workers, it’s thin but growing.

n8n (n8n GmbH, Germany) is the most relevant option for sovereignty-conscious users. It’s open source, self-hostable, has a visual interface, and has native AI agent nodes. You can build a workflow that triggers a research task, processes the output with a model of your choice, stores the result in Notion, and drafts a newsletter — without writing code, and without sending your data through a single US company’s infrastructure.

Make (Celonis, Germany/Czech Republic) and Zapier (Zapier Inc., USA) are the more accessible options — easier to start with, but closed and cloud-only. Activepieces (open source) is a newer alternative worth watching.

But here’s the warning I want to leave with you about this layer:

The automation trap is real. Building workflows because you can, not because you need to, is just moving the dependency from a model to a process. I’ve built automations that seemed powerful and turned out to be unnecessary. The question is never “can I automate this?” The question is “what output do I actually need?” Start from there. Work backwards. Sometimes the answer is a workflow. Often the answer is a clear prompt and some time.

Subscribe now

Let’s not forget MCP

If you do build workflows or tools, there’s a protocol worth knowing about. MCP stands for Model Context Protocol. It is an open standard, created by Anthropic but now adopted across the industry, that lets AI models connect to external tools and data sources in a standardized way.

What this means in practice:

Leave a comment

4 questions we should not forget

This is the sovereignty framework I use now. It applies to Claude, to ChatGPT, to every tool in this article, and to every tool that will be released after it.

Who owns what I put in? Read the terms of service. Understand whether your documents, conversations, and creative work are used to train future models. Know whether you can delete your data, and whether deletion is real.

What happens if I stop paying or they change their terms? Every tool you depend on can change pricing, deprecate features, or shut down. If that happened tomorrow, what would you lose? Can you export your work in a format another tool can read?

Can I take my work somewhere else? This is the portability question. Your accumulated context — preferences, custom prompts, the memory of your projects — lives inside platforms that don’t interoperate. The closer you get to open standards and open formats, the more of that context you actually own.

Whose worldview is built into this model? This is the question I almost skipped. Every model — closed or open, American or Chinese — was trained somewhere, under some legal system, by people answering to someone. Sometimes that shows up as a server-side filter you can route around. Sometimes, as we just saw with Chinese open-weight models, it’s baked into the weights themselves and follows the model even offline. Ask what topics a model avoids, deflects on, or has a suspiciously uniform opinion about. That tells you something no benchmark will.

An invitation: slowly, slowly…

Let’s start breaking dependencies.

Go to Le Chat. Or Qwen. Or Z.ai. Use the same prompt you would give Claude. See what happens.

Pure experimentation. You’ll see it is worth your time.

The AI companies that have your attention right now are expert at making their tools feel indispensable. That expertise is real. The tools are good. But indispensable is a feeling, not a fact.

There is more room here than you have been led to believe.

Sources and reference material

This article leans on two companion pieces that I kept separate so the main read doesn’t get bogged down in sourcing and tables.

Who Really Controls Your AI? — The US and China Question — the full story behind the Pentagon contract, the Maduro operation, the Iran strike, Dario Amodei’s own words on it, and the documented political bias baked into Chinese open-weight models. Quotes, sources, and links included.

Reference Tables — AI Alternatives for Knowledge Workers — all seven tables from this research: models, interfaces, workflow tools, local runners, knowledge management tools, APIs, and the developer ecosystem. Every entry includes who owns it, where they’re based, and a direct link.

Do you find something I’m missing in the files? I’ll be happy to edit them with the proper credits:

Leave a comment

I hope you found this article valuable.

Love,

Jose

Saturday, 13th, 2026

I wrote this article yesterday. I decided to let it mature before publishing. This is what I woke up to this morning:

This is exactly what I mean when writing about the weaponization of AI and its use as a national security concern. I did not change a word of the article.

Screenshot from Anthropic’s website.

Friday, 12th, 2026.

Is the LLM gold rush over?

I think it is. Not the AI one, though.

Apple announced a few days ago at WWDC something that goes way beyond Siri improvements; it is a signal of what the AI world looks like today, and where it is going.

Let me explain.

Apple believes that for most uses, we don’t need cloud-based LLMs. They have decided that Mac OS should be an AI-enabled system that processes workflows and tasks locally.

Cloud systems can be used when needed. It makes sense, users will not need to buy a monthly subscription if their Mac has the power to run AI automations and tasks natively.

What does that mean for you and me? Probably most of our automations and Claude skills will eventually run on our Macs. We will likely have to rebuild our apps.

So what happens to LLMs? They are already hinting at where they intend to go: advanced AI work: agents, harnesses, deep reasoning tasks. Specialist work, not default infrastructure.

Subscribe now

LLMs are genuinely useful, but our hopes and imagination may have been playing a role in our misinterpretation of what they are really good at. LLMs have a limitation by design: they are probabilistic in nature.

Probabilistic systems interpret context; they do not execute with certainty. Asking an LLM to scan invoices and always update a database correctly is more hope than a solution. That use case pushes a probabilistic system to behave as a deterministic one. So why not use a deterministic system instead? Many experienced early adopters will disagree, but I would love to know how many credits their agentic systems cost per transaction, how much maintenance time they require, and how long they took to build. And then I would ask:

Why not use the LLM to build the deterministic tool instead?

A well-architected system can manage the probabilistic nature of an LLM through validation layers, confidence scoring, and human review queues. That is true. But those layers have a cost — in development time, in maintenance, in the human oversight required to catch what the model gets wrong. That cost rarely appears in the business case. It surfaces later, in the people doing the work that the automation was supposed to eliminate.

Leave a comment

That brings me to what LLMs genuinely do well:

• Democratizing software development — removing the technical barrier while the human still directs

• Accelerating learning — removing the access barrier while the human still synthesizes

• Interpretation aid — reducing cognitive load while the human still decides

• Language and translation work — removing friction while the human still owns the meaning

Notice the pattern. In every case, the human remains essential. LLMs are amplification tools: they speed up and increase in size what we want to do, and also our mistakes.

The consumer marketing around AGI has gone quiet. The labs themselves are more AGI-focused than ever, as it is in their stated missions and research agendas. But the public narrative has shifted toward practical features and monthly subscriptions.

Apple’s decision to focus on local, practical AI for its users rather than chasing frontier model benchmarks is telling. It suggests that the race toward artificial general intelligence may be less central to real-world value than the industry was claiming.

This makes me suspect that OpenAI, Google, and Anthropic are working on very different models behind closed doors, experiments we do not know about yet, because the current LLM approach has a ceiling, and they know it.

Subscribe now

I find it especially disturbing when I hear people framing AI as a national security issue. That framing does not produce wisdom; it produces escalation.

Every major technology of the last century that got absorbed into a power and dominance narrative ended up generating conflict rather than progress.

Other nations and blocs will not watch passively. They will build alternatives, restrict access, and retaliate through regulation and competing investment.

The arms-race framing fragments the technology rather than developing it. Apple, interestingly, offers a partial counterexample:

I am not saying AI is over. I am saying LLMs are hitting a wall. The LLM business model is under pressure, not because the technology has stopped improving, but because the cost of accessing that improvement keeps rising while the sustainable use cases for most businesses and independent professionals remain narrower than advertised.

I do not want to be distracted by the barrage of new functionality announcements; they are good manipulation tactics. The truth reveals itself more clearly when you look at the subscription models and the escalating credit prices.

I am an AI early adopter myself, and I write Automato precisely because I want to connect with people interested in this space, sharing value and mutually benefiting from the opportunities this innovation might bring.

But my decades in the industry hint at the fact that our enthusiasm often hides the underlayers of what is really happening. So this article is also a warning to myself:

What is your take?

Leave a comment

I wish you a very good day!

Jose from Automato

TL;DR Growing companies accumulate informal systems that work, until they don't. Spreadsheets nobody can explain, processes nobody has documented, technology decisions made in the wrong order. There is a specific sequence to fix this. This is what it looks like in practice.

Every company has a version of this story: The company starts with a few team members, grows fast, the business develops around the first customers, the complexity expands, and as time passes, some inefficiencies stay in place.

This story, similar to most businesses, starts with a phone call.

“Jose, we have a problem with our SharePoint.”

The problem was not SharePoint

A few years ago, someone I had worked with on a SharePoint project many years before called me. His company had been trying to solve the same technology problem for five years.

Our first meeting made one thing clear: the team had been treating SharePoint as a filing cabinet. The complexity and the number of SharePoint sites were overwhelming, and the information stored there was not usable.

When I started asking how information moved through the company, something else emerged.

There was no standard for how information was created, stored, or shared. The different business areas had their own rules; the documents lived in personal notebooks, local drives, cloud folders, and Excel files that only a few people knew how to use. When management requested a report, people knew roughly where to look. “Roughly” is not a system.

Subscribe now

The Excel files nobody could explain

One detail from those early weeks stayed with me.

Several areas were running their operations on shared Excel files of remarkable complexity; macros, cross-referenced queries, and conditional logic built over the years. When I asked who had designed them, the answer was always some version of:

“ He left the company two years ago.”

Nobody really knew how to reverse engineer those files, something that could save us lots of time in discovering how some processes really worked. The spreadsheets worked, and the company was doing fine.

To me, the problem was an information knowledge problem. Knowledge that exists only inside a specific person or a specific file, that cannot survive their absence.

The company was profitable, well-managed, growing, and dependent on fragmented systems that hid how the information was processed.

Leave a comment

Before any technology decision: map the processes

My first recommendation was not to fix SharePoint. It was to make the processes visible.

We asked every business area to document how they worked. Some sent Excel files. Some sent Word documents. Some sent flow diagrams with their own notation systems. The variety was itself the finding: there was no shared language for describing their work.

We ran a couple of simple workshops. I built a shared standard for documenting a process, applied consistently across the whole company. When it was done, the leadership team said something I hear often after this exercise: for the first time, they could see the entire company.

We could see where processes connected, where they overlapped, where automation could play a role, and where the flow of information got stuck.

That visibility is the prerequisite for every decision that follows.

What the map revealed

Once the processes were visible, the patterns were hard to ignore.

Almost every area had built informal backup processes, steps to verify information arriving from other departments before acting on it. That revealed that the different areas were not aware of how the others work, and there was a lack of trust.

Significant portions of each team’s day were spent redoing work already done somewhere else.

Process changes were happening regularly, discussed in team meetings, implemented immediately — but rarely updated in any documentation, relying on their team's ability and professionalism.

None of this was anyone’s fault. It is what growth does to organizations. The informal systems that carry you through the early years become invisible infrastructure. At some point, that infrastructure starts costing more than it saves.

Leave a comment

A philosophy before a technology decision

Before choosing any tool, I established what I call Digital Sovereignty:

The capacity of a company to understand, control, and evolve its own systems, without being permanently dependent on vendors whose pricing, policies, and priorities are outside its control.

In practice:

• build internal skills,

• reduce external dependencies,

• lower licensing costs,

• improve your negotiating position with third parties.

For this company, the first step was a licensing audit. The savings were immediate. People felt the change before a single new system was built. That matters; momentum is part of the design.

Subscribe now

The first implementation: Work Orders

With processes documented and a philosophy in place, we chose where to start building.

We chose Work Orders, the process that connected the most areas of the business.

1. A project begins when a contract is signed in Sales.

2. That triggers Procurement to plan materials,

3. Engineering to begin design,

4. Production to receive specifications.

Each area has its own rhythm, its own information needs, its own permissions and notifications.

The company was managing this across emails, tools, and individual conversations. The status of any specific work order was difficult to trace. Management had no single view of how the composite of work orders is doing.

We built a centralized system using SharePoint and Power Automate — tools already in the company’s Office 365 subscription. No new licenses. Two students were hired to learn the platform and build on it.

The result:

New IT area, a clear methodology, a decision to work with what the company has, and a progressive approach designed to generate trust with the systems stakeholders.

Leave a comment

What came next was the plan

After Work Orders, the team moved to document management for the sales process. Then other areas, one by one. The students hit the limits of Power Automate. They learned more about SQL databases, APIs, and more complex data structures. They used AI tools to solve specific coding problems. They moved from no-code to low-code to custom development, driven by real problems.

This progression was intentional. Digital Sovereignty only works if internal capability grows over time. You build it through a sequence of projects, each one slightly more complex than the last, each one adding something the team now owns permanently.

The AI question

Over the past year, I have been talking with engineers using AI-assisted development tools, business leaders experimenting with automation platforms, and people building software with AI and no formal development background.

The pattern is consistent: without process clarity, AI is not a solution. It is an accelerant.

The instinct when someone discovers what AI can do is to start building. You ask what is possible, you create something that seems useful, you move fast. Often, you end up with a solution to a problem that does not quite exist.

AI amplifies what you already know. If you know your processes, it becomes powerful. You can build faster, automate the right things, and describe what you need with precision. If you do not know your processes, it amplifies the confusion.

Process first. Always.

Leave a comment

What I keep hearing

Since that first engagement, I have had versions of the same conversation with leaders across multiple industries.

Profitable companies. Growing companies. Companies with skilled, committed people. Still running critical operations on spreadsheets that one person built and nobody else fully understands. Still making technology decisions before understanding the processes those decisions are meant to support.

The answer is almost always the same. Workflow and process automation work when you know your processes, their risks, and bottlenecks.

That is where digital transformation begins. With a question most companies have never seriously asked:

Do we actually understand how we work?

If your company is going through something similar and you need some help with that, reach out to me. I’ll be happy to support:

Have a great one!

Jose from Automato

Subscribe now

He is applying for a role as an AI Engineer. After many questions and back and forth, he finally gets more details on what the role actually looks like. And what he described was surprising — let me paint it for you.

He receives a ticket with specifications. That is the input. Then he uses a set of AI agents, tools, and workflows to process that input and produce an outcome. If the ticket requires it, he creates new tools and processes, and then lets the system do the work.

He asked about interactions with team members. The answer was essentially: you work independently, with minimal interaction. Your role is not only engineering — it is also product management. You and your agents are a mini team.

Wow. That blew my mind.

And apparently, this is already being named. Nate B. Jones, AI strategist and former Head of Product at Amazon Prime Video, recently wrote about exactly this emerging role:

“It isn’t GTM Engineers. It’s Deployment Strategists. Decagon calls it an ‘Agent Product Manager.’ Harvey calls it a ‘Solutions Architect.’ And the salaries are climbing fast.” — Nate B. Jones, LinkedIn (https://www.linkedin.com/in/natebjones/)

I can easily see companies hiring AI orchestrators to control most information flows in a business. The AI engineer needs to think like a product manager, plan like a project manager, and execute like an engineer. Not only in software creation, but on anything. The human is in charge of judgment and tooling. AI processes the information. The human assumes the responsibility.

In the past, it was fascinating to watch graphic designers become web developers, and programmers become interface experts. Is the same thing now happening between Product Management and Software Engineering? I think it is — and in both directions.

As Vaskin Kissoyan writes in his Substack Agentic Insights:

“The future of software delivery isn’t a larger team. It’s fewer people, higher leverage, and orchestration done deliberately.” — Vaskin Kissoyan, The Rise of the AI Orchestrator

Agentic Coding Newsletter

The Rise of the AI Orchestrator

When the Agency Model Quietly Broke…

Read more

7 months ago · 1 like · 2 comments · Vaskin Kissoyan

All of this comes from one rich conversation. I wanted to share it and hear your thoughts.

Do you see something like this happening around you? What did I get wrong?

Leave a comment

Jose from Automato

The reactions said as much as the data. Two comments from the article stood out:

wrote: “The gap between IT and business on AI is where most companies quietly fail.”

added, “While IT is buried in governance, non-technical leads are already out there building their own workarounds. If companies do not get ahead of that invisible build wave now, they are going to be drowning in technical debt in a year or two.”

That’s exactly why I built this PDF.

It’s a practical tool — the 16 questions ranked by expert consensus, an AI skill development journey, and the full list of contributors. You can print it, share it, or use it to open a conversation that might not be happening yet in your organization.

A few ways to use it:

• Share it with your customers or prospects as a conversation starter

• Bring it to your own leadership team or IT area

• Upload it alongside the survey results to any LLM — ChatGPT, Claude, Gemini, and go deeper: find more research, build arguments, tailor the questions to your industry

Download the PDF

Check the survey results if you haven’t yet:

What topic should I survey next? If you have a question you think the AI community should be weighing in on, reply to this email or leave a comment below.

Leave a comment

I wish you a good day!

Jose from Automato

In her experience, IT is not always seen as a strategic partner in a company. Often it functions as a service layer: responding to needs, managing infrastructure, staying out of the boardroom conversation. She pointed out that real AI adoption happens at the user level — behavior, language, process — and that for it to work, IT and business need to speak to each other, not past each other.

She was describing exactly the gap this survey is trying to address.

What This Survey Is Trying to Do

Many companies are in a strange situation right now. Their IT teams are exploring AI tools, making technical decisions, and building things. Their business leadership is reading articles, attending conferences, and feeling pressure to act. And between the two, there might be an awkward silence.

I assume that IT should be a strategic leadership function, not a service desk. Especially now. And my goal is to give business leaders a tool to start the conversation: a set of validated questions they can bring to their IT teams to understand where the company actually stands on AI.

To build that tool, I ran a Delphi-style expert survey. I sent 16 questions to a selected group of AI practitioners and early adopters. I asked them to rate how important each question was for a company to be asking, not to answer the questions themselves, but to validate them. The aggregated results become the questionnaire. The experts become its authority.

17 people responded. You can see the full interactive results report here:

Open the detailed report

Who Answered

The panel matters. Here is who participated:

AI builders and engineers:

• Jani Nedeljko — Senior Frontend Developer · pavone.ai

• Lovro Hudrap — Software Engineer & Co-founder · extra.dev

• Vito Meznaric — Software Developer · grandpacad.com

• Žiga Palčnik — Lead Software Developer at Wepik

• Tadej Matek — Full Stack Engineer at Synthesia

• Juan Gonzalez — AI Full Stack Engineer · is on Substack

• Jenny Ouyang — Software Engineer · AI Builder · is on Substack

• Marko Balažic — AI Workflow Engineer & Product Lead at Shape Labs · shape-labs.com

AI practitioners and product leaders:

• Gregor Ojstersek — CTO · Founder of Engineering Leadership (188k+ subscribers) · is on Substack

• Elena Calvillo — AI Product Manager · Enterprise Scale · Book Author · is on Substack

• Dheeraj Sharma — AI Systems Builder · is on Substack

• Daria Cupareanu — AI Systems Builder · Founder of AI blew my mind · is on Substack

• James Presbitero Jr — AI Consultant & AI Marketing Specialist · is on Substack

Leadership, consulting, and strategy:

• Joel Salinas — AI Leadership Author & Humanitarian Leader · is on Substack

• Farida Khalaf — Data Engineer & Strategic Consultant · is on Substack

• Viktor Kyrychenko — CEO of Startup Studio · IT Entrepreneur · Investor

Storytellers and writers:

• Finn Tropy — Storyteller · Engineer · Mentor · is on Substack

Special contributor:

• Agnes Molnar — Strategic advisor for senior IT leaders · AI-era transformation · 25+ years in IT · Founder of Search Explained · is on Substack

My starting point: Digital Sovereignty

Before I share what the experts said, I want to be transparent about where I was coming from.

My approach to AI adoption starts with process discovery, understanding what a company actually does, where the friction is, and which processes could genuinely benefit from automation or AI.

From there, my consulting philosophy leans in a clear direction:

This is also why I believe IT should be a strategic leadership function. Not because IT people are more important than others, but because in a world where software is becoming probabilistic, and AI is embedded in every business process, the people who understand those systems need to be at the table where strategy is made. And knowledge workers — the people who actually use the tools — should eventually become builders of their own solutions. Not just users.

That vision shaped some of the survey questions. The ones about local and self-hosted AI models. Vendor dependency reduction. Build-vs-buy strategy. Non-IT teams developing their own tools.

Those questions scored among the lowest in the survey.

The experts are not wrong. Their experience is grounded in what companies are actually dealing with today — and today, most companies are far from the point where Digital Sovereignty is a realistic conversation. But I think it’s the right direction to be moving toward. And I think the companies that start asking these questions now will be in a very different position in five years.

Leave a comment

The 16 questions, ranked by the expert panel

Here are the full results. I’ll let the numbers speak, with a short note on what each distribution actually reveals.

The full interactive report is here.

#1 — Is the IT team implementing safety and security protocols for AI use? 11 marked it Essential. 4 marked it Very Important. 15 out of 17 experts consider this critical. This is the clearest signal in the entire survey: before anything else, know whether your IT team has guardrails in place.

#2 — Is the company’s data infrastructure ready to feed AI systems reliably? 6 Essential, 8 Very Important. 14 out of 17 experts see this as important or above. No data strategy, no AI strategy — this question gets at the foundation that most AI projects quietly fail on.

#3 — Does the company have a clear data governance strategy that supports AI adoption? 8 Essential, 5 Very Important. 13 out of 17 experts ranked this high. The second-highest Essential count in the survey — only Q7 scored higher on that measure. The 2 Somewhat Important and 1 Not Important responses suggest a small minority doesn’t see it as urgent yet, which may reflect organizations still in early AI exploration.

#4 — Is the IT team developing and maintaining a company-wide AI usage policy? 6 Essential, 7 Very Important. 13 out of 17 in the top two tiers. The word “maintaining” matters here — this isn’t a document you write once. Experts seem to agree.

#5 — Is the IT team using AI tools to analyze data? 6 Essential, 6 Very Important. 12 out of 17 in the top two tiers. A practical question about present-tense capability. The 3 “Somewhat Important” responses are worth noting — they could reflect experts who believe AI is not always the right tool for structured, deterministic data analysis. A nuanced position, not a dismissal.

#6 — Is the IT team making a clear distinction between probabilistic and deterministic systems? 7 Essential, 5 Very Important — but 4 marked it Somewhat Important and 1 Not Important. This split is telling. For those who understand it, the distinction is fundamental. For those who don’t, the question itself may be unclear. That gap is part of what the questionnaire is designed to surface.

#7 — Is the IT team using AI tools to develop software? 7 Essential, 3 Very Important — but 5 marked it Somewhat Important. This question divides opinion more than its ranking suggests. The Essential votes are strong, but nearly a third of experts see it as secondary. Context matters: a software company rates this differently than a retail chain. But there’s another reading — for practitioners already using AI deeply across research, documentation, and knowledge work, software development may feel like just one use case among many. The question worth asking back: what else are they doing with AI?

#8 — Has the company analyzed the impact of AI on its business ecosystem? 5 Essential, 4 Very Important, 5 Important. A wide spread — 14 out of 17 see it as at least Important. The lack of strong consensus in the top tiers is surprising to me. Understanding how your industry is adopting AI isn’t just awareness — it’s a strategic timing question. When your competitors start implementing their AI strategies, will it be too late to catch up? The companies that are mapping their ecosystem now are the ones building the edge. Waiting to ask this question may feel safe. It isn’t.

#9 — Does the company have an active program to develop AI literacy beyond the IT team? 4 Essential, 5 Very Important, 6 Important. 15 out of 17 consider this at least Important. Nobody marked it Not Important — which is itself a signal. Everyone agrees it matters. But only 9 see it as truly critical. This is where my own vision diverges from the consensus — I think this question should be higher. And here’s why: what will the IT area do when they discover that multiple people across the company are already building apps, automating processes, and connecting AI tools — without any coordination? The literacy program isn’t just about education. It’s about getting ahead of something that is already happening.

#10 — Is there a defined owner or team accountable for AI strategy and investment? 2 Essential, 8 Very Important. The high Very Important count with low Essential is interesting — experts see ownership as important but perhaps less urgent than the technical foundations. 2 marked it Not Important.

#11 — Are non-IT teams being prepared to eventually build or configure their own AI-assisted tools? 3 Essential, 4 Very Important, 6 Important. 13 out of 17 find it at least Important. But only 7 see it as truly critical. This is one of my core beliefs — that knowledge workers will become builders. The data suggests most experts see it as a future question, not a present one.

#12 — Is the company assessing the risk of geopolitical or regulatory changes affecting its AI stack? 3 Essential, 5 Very Important — but 2 marked it Not Important and 3 Somewhat Important. The most geographically and contextually dependent question in the survey. For a European company, this is urgent. For others, perhaps distant.

#13 — Are different business areas coordinating with IT on their AI use? 2 Essential, 4 Very Important, 7 Important. A wide, gentle spread. 13 out of 17 see it as at least Important, but no one is shouting that it’s critical. This is Agnes’s territory — and the data suggests experts know the gap exists without necessarily knowing how to close it.

#14 — Does the company have a strategy to reduce dependency on third-party AI vendors and licensing costs? 1 Essential, 5 Very Important, 5 Important, 5 Somewhat Important. Almost perfectly split between those who see it as strategic and those who see it as secondary. This is one of my core Digital Sovereignty questions. The market is not there yet. But I believe it will be.

#15 — Does the company have a clear build-vs-buy strategy for AI solutions? 0 Essential. 6 Very Important. 7 Somewhat Important. Nobody marked this Essential. The most technically grounded Digital Sovereignty question landed at the bottom of the priority list. Experts are focused on what companies are doing now — not on the strategic architecture questions I think will define the next five years.

#16 — Is the IT team running any locally hosted or self-managed AI models? 1 Essential, 1 Very Important, 7 Somewhat Important, 3 Not Important. Only 2 out of 17 experts see this as truly important. This is the question that most directly expresses my Digital Sovereignty bias — and it landed last. The gap between my expectation and this result is the most honest data point in the survey.

The AI Skill Development Journey

Experts were asked to sequence seven capability areas in the order a company should develop them, from first to last. This is the consensus:

A few things I notice in this sequence. Experts put research first, not software development, not automation. The journey starts with people learning to think with AI before building with it. That aligns with my own approach: process understanding before tool implementation.

Workflow automation, despite being one of the most discussed AI topics, landed sixth. Experts seem to be saying: earn that before you get there.

And monitoring information systems came last. That’s the most infrastructure-dependent capability, requiring mature data pipelines, reliable models, and governance structures that most companies haven’t built yet.

What the Experts Added

Missing questions

Four experts suggested questions they felt were absent from the original list:

— AI Consultant & AI Marketing Specialist
How open and curious — or cautious — is the IT team about AI? Culture is the invisible layer beneath every technical question on this list.

How are non-technical leads being enabled to prototype safely without waiting for a full IT sprint? What is the strategy for managing models that change or update over time? And perhaps most importantly: do we have a path for the shadow AI tools — the small solutions teams are already building — to become official, secure infrastructure?

— Data Engineer & Strategic Consultant
Is the AI team aligned with the legal department and the financial budget?

— AI Full Stack Engineer
Does the company have structured workflows and SOPs before adopting AI? Is the company assuming that AI will solve all its operating problems? Has the company considered upskilling programs?

Additional AI capabilities suggested

Six experts suggested capability areas missing from the skill development journey:

— Using AI to amplify expertise rather than replace it. This needs deliberate training, not just tool adoption.

— AI use for finance. AI use for making data-driven decisions.

— Understanding that AI generates probabilistic outputs and does not exercise judgment. Accountability and final decision-making remain with human users.

— AI-powered skill extraction and reusable skill packaging — using AI to identify and codify expertise from human work into structured, distributable packages.

— Architectural literacy: understanding how to stack tools without creating technical debt. AI cost observability: learning to run the math on token usage and infrastructure before a project explodes in cost.

Žiga Palčnik — AI governance, risk, and compliance as a first step in the journey, not an afterthought.

Wrapping up

This article — and the survey behind it — is not a final answer. It’s a starting point.

If you are a business leader, the questions in this list are not meant to audit your IT team. They are meant to open a conversation. Ask your IT colleagues what they value most. Ask them why. Ask them what they think is missing. The goal is not efficiency, cost savings, or early adoption for its own sake. The goal is to integrate the IT area into your business strategy — deeply, honestly, and early enough to matter.

The biggest risk is the status quo. It is not thinking deeply enough, soon enough. Apathy and short vision have a cost that only becomes visible later. Ask the companies that waited too long when the World Wide Web arrived. Some of them never caught up.

AI is not a trend to monitor. It is a structural shift in how companies create value, compete, and survive. The conversation between business leadership and IT about AI is strategic.

The questions in this survey are intended as a bridge.

What comes next

The full interactive results report is available here.

A PDF with the validated questionnaire — ready to use with your IT team — will be shared as a dedicated post for Automato subscribers. Contributors to this survey will also receive copies to share freely with their networks, clients, and audiences.

Thank you to everyone who participated. Your names, publications, and LinkedIn profiles are in the report.

Love,

Jose from Automato.

TL;DR You can build a personal task manager inside Claude using two pieces: a Claude Skill that adds tasks to your data source instantly, and a Claude Cowork Live Artifact that displays them as a Kanban board. This article gives you the exact prompts to build both — plus the map of every mistake I made so you don’t have to make them yourself.

What’s in this article

• The real problem — why existing tools fail at the moment tasks actually arrive

• What went wrong — a week of errors, duplicates, timeouts, and pagination surprises

• The system today — how two focused tools replaced one over-engineered skill

• Why I chose Notion — the tradeoffs, the honest doubts, and what you might choose instead

• A design philosophy — why simplicity is a deliberate decision, not a limitation

• Part 1: Build your own task skill — a prompt that guides you from zero to a working skill, including creating your database if you don’t have one yet

• Part 2: Build your own live Task Board — a prompt for a persistent Cowork artifact that loads and displays your tasks without involving Claude each time

• An honest assessment — what works, what doesn’t, and where this is all heading

Subscribe now

A bit of human context

My partnership with Microsoft started with time management.

Not with SharePoint, not with infrastructure, not with enterprise deployments. With Outlook. With calendar invites and shared task lists, and the promise that if you used the right tools correctly, you would finally be in control of your day. That workshop — one of my first, sponsored by Microsoft when I arrived in Slovenia — felt like the beginning of something. And for a long time, it was.

But that was then. Today, “using Outlook for tasks” means navigating a fragmented experience across a web app, a desktop client, and a mobile app that each behaves differently. Connect your iPhone’s native mail client, and you get a fourth variation. It is too complicated for something that should be simple.

So for years, like many solopreneurs, I settled for good enough — a combination of apps, notes, and memory. Until I realized I already had the tool I needed open in front of me every day.

The Real Problem

Tasks don’t arrive at convenient moments.

I’m driving, and an idea surfaces. I’m working in the orchard, and I remember something I need to follow up on. I’m in the middle of a conversation, and a task appears. In those moments, I need the path of least resistance.

Siri should be the answer. It isn’t, at least not reliably enough to trust.

But Claude is open in front of me for most of the working day. I use it constantly. So the question became obvious: why open another tool at all?

I thought the solution would take twenty minutes to build.

I was wrong.

What Went Wrong (So You Don’t Have To Find Out Yourself)

The first part worked immediately: adding tasks. Tell Claude, the task appears in Notion. Fast, reliable, done. Claude even created the Notion DB, extracted the DB IDs, and all technical details needed to make the integration work.

The problem started when I wanted more, a skill that could also retrieve and display my tasks in Claude. A complete task manager inside Claude, one skill to rule everything.

Here is what I ran into:

Speed. Generating a simple HTML view of fewer than twenty tasks through a Claude skill took several minutes. Not seconds. Minutes. Unusable. Claude had to rebuild the HTML each time.

Pagination. The Notion API returns a maximum of 25 results per call. If you don’t handle this explicitly, you silently lose tasks. They don’t error — they disappear.

Duplicates. Running parallel calls to work around pagination introduced a new problem: the same task appearing twice, with slightly different ID formats. The skill had no way of knowing it was the same record.

Cowork Scheduled Tasks. I tried using Claude Cowork to run scheduled task refreshes. Timeouts. Excessive usage against my subscription. An 80% failure rate. That experiment ended quickly.

Type contamination. Notion search doesn’t return only tasks. It returns databases, schemas, views — everything. Without strict filtering, phantom entries inflate your task count and render as broken cards.

After more than a week of iteration, I had something that technically worked. But it was fragile, slow, and not something I could share with anyone in good conscience.

The System Today

Two pieces. Clean separation of responsibility.

A skill for adding tasks. One job only: you mention a task, Claude adds it to your data source, and confirms in a single line. No retrieval, no display, no status management. The constraint that started as a defeat turned out to be the right design. Adding a task from Claude is genuinely faster than opening any other app.

A live artifact for viewing tasks. A persistent Task Board that lives in Claude Cowork, connects directly to Notion, and loads your tasks without asking Claude to reason through anything. Fast, reliable, and always there when you need it.

Neither piece tries to do the other’s job. That’s what makes it work.

Why I Chose Notion — and Why You Might Choose Something Else

I use Notion because I already have a subscription, and I wanted to test how capable the MCP integration really is. The answer: capable, with real limitations — which is exactly what this article documents.

What I like about Notion for this use case:

• Views. Notion’s data visualization is genuinely good. Kanban, table, calendar — you can look at your tasks the way that makes sense to you.

• Pages as records. Every task in a Notion database is also a page — it has an open space for writing. So if a task needs context (notes for a meeting, a checklist, a link), that space is already there. You don’t need extra fields.

• Cloud access. I can interact with my tasks through Claude, but also directly through the Notion app on my phone or computer. The data lives in one place and is accessible everywhere.

What I keep questioning: would a local spreadsheet or even a Markdown file be faster and more reliable for Claude to read and write? I don’t have a definitive answer yet. A local database removes the MCP layer entirely, which would likely solve the speed and pagination problems. But you lose the cloud access and the native Notion views.

It is not a permanent decision. It is the best available option I found while working with the tools I already use.

A Design Philosophy: Keep Tasks Simple

Before you build anything, I want to share one decision:

Most task tools push you toward complexity, due dates, priority levels, assignees, tags, linked projects, progress percentages, and long descriptions. I understand why those features exist. But for the way tasks actually arrive in my life, that complexity creates more friction than it removes.

So my design rule is: the task name carries the meaning. Everything else is optional.

For the rare cases where a task genuinely needs more — I had one recently involving a meeting where I wanted to capture a few ideas and a checklist in advance — Notion’s page body is the right place for that. Open the record, write what you need. No extra fields, no system redesign.

If you end up using a different tool, look for something equivalent: a notes field, a description area, a rich text space attached to the record. One optional field for context is all you need. The task name should do the work.

Part 1: Build Your Own Task Skill

Where to run this: Standard Claude chat is the default. However, if your chosen data source MCP, is only connected in Claude Cowork and not in regular chat, run it there instead. Check Settings → Integrations to see which MCPs are active in each environment.

Supported data sources (you’ll need the corresponding MCP connected in Claude):

• Notion, Google Sheets, Airtable, Todoist, Trello, Linear, GitHub Issues. Do you know any others?

  Leave a comment

Before you start: Make sure the MCP for your chosen data source is connected and authorized. Without this, the skill won’t have access to your data — and Claude will build something that silently fails every time.

Subscribe now

Prompt: Task Skill Builder

I want to build a Claude skill that lets me add tasks by just mentioning them in conversation — no switching apps, no extra steps.

Before writing anything, ask me these questions and wait for my answers:

1. Where do I want to store my tasks? Options include Notion, Google Sheets, Airtable, Todoist, Trello, Linear, GitHub Issues, or another tool. Tell me which MCPs are available in my current Claude setup and help me choose the best fit for my situation.
2. Do I already have a database or list set up for tasks in that tool, or do I need to create one? If I need to create one, show me exactly what you are about to create and wait for my approval before doing anything.
3. What fields should the task have? Keep it minimal: task name and status are required. Add an optional rich text or notes field for context when needed. Ask me if I want anything else, but push back gently if I start overcomplicating it.
4. What should the confirmation look like after a task is added?

Once I answer, follow these steps in order:

**Step 1 — Verify the connection.**
Make a test call to the MCP before writing anything. Confirm the data source is reachable and responding. If it fails, stop and tell me clearly — do not proceed to build a skill around a connection that doesn't work.

**Step 2 — Inspect the actual data structure.**
Make one real read call to the data source. Observe the actual field names, ID format, and how records are nested. Build everything around what you observe, not what you assume. Never guess field names. If you need to create the database first, do it now — and confirm the exact name and stable ID before continuing.

**Step 3 — Write the SKILL.md.**
The skill must:
- Trigger only on clear task intent — not every casual mention of something to do in conversation
- Capture tasks mentioned in conversation and add them to my data source in a single API call (never loop with multiple sequential calls, even for multiple tasks)
- Hardcode the exact database or list ID — do not rely on Claude finding it dynamically each time
- Confirm with one short line after adding (e.g. "Added → [task name]")
- Handle multiple tasks mentioned at once in a single call
- If the MCP call fails, tell me clearly — never silently confirm a task that was never saved
- Focus on one job only: adding tasks. No retrieval, no display, no status management.

**Step 4 — Test the skill.**
After writing the SKILL.md, add one real test task using the skill exactly as it will be used. Confirm the task appears in the data source. Show me the confirmation. Only then tell me the skill is ready.

Leave a comment

Part 2: Build Your Own Live Task Board

Where to run this: Claude Cowork only. Live artifacts are not available in the standard Claude chat interface.

This prompt builds a persistent Task Board — a sidebar panel that loads your tasks directly from your data source every time you open it. It handles pagination, deduplication, error states, and data integrity, so you don’t have to discover those problems the hard way.

Before you run this prompt:

• Confirm your data source, MCP is connected and authorized in Cowork

• Know approximately how many records your database has

• Have the exact name of your MCP server as it appears in your Cowork connections

Leave a comment

Prompt: Live Task Board

Check for placeholders in the prompt:

I want to build a Live Task Board as a Cowork Live Artifact — a persistent sidebar panel that loads my tasks directly from my data source every time I open it, without asking Claude each time.
My tasks are stored in [describe your source: e.g. "a Notion database called Task Manager" / "an Asana project" / "a CSV file in my outputs folder" / "a Trello board"].
Each task has at minimum: a name and a status (something like To Do / Done — tell me what yours is called).
Features I want:

Open tasks in one column, newest first
Tasks older than 7 days highlighted in a different color
A Done button on each card that updates the record in my data source
Completed tasks filtered to this month only, visible via a "Full panorama" toggle
A Reload button to refresh data without closing the panel


Before writing a single line of code, do two things:
1. Ask me these questions and wait for my answers:

What data source / MCP am I using, and what is its exact server name as it appears in my Cowork connections?
Is the MCP already connected and authorized in Cowork?
How many records does my database have approximately?
What are the exact status values in my data (e.g. "To Do" vs "Done")?
What should the Done button do — change the status field, archive the record, or delete it?
Are there any other fields I want to show on the cards (due date, tags, assignee)?

2. Once I answer, make one real test call to the MCP before building anything. Inspect the actual response — field names, data types, how IDs are formatted, how results are nested. Build the artifact's data layer around what you actually observe, not what you assume. Never guess field names.

Safeguards — data integrity (non-negotiable):

Pagination: Most MCPs cap results at 25 per call. Run two parallel search calls with a date split (e.g. recent 3 days vs. older) and merge the results. If the user has more than 50 records, increase the number of calls accordingly.
Deduplication: The same record can appear in both calls with different ID formats. Deduplicate using the record's stable UUID (r.id stripped of dashes), not the URL or title.
Type filtering: Search calls return more than just task records — databases, schemas, and views can appear in results. Filter strictly to type === "page" (or the MCP equivalent) so phantom entries don't inflate counts or render as cards.
Rate limiting / slow loading: Don't fire all fetch calls simultaneously. Use a concurrency pool capped at 6 parallel requests with auto-retry on timeout. Show a progress bar while loading.
Done button — writes to data source: The Done button must call the MCP to update the actual record, not just change the card's appearance. Optimistically update the UI immediately, but revert visually if the call fails.
Done button — sequential clicks: Use a Set to track which records are currently being updated so multiple Done buttons can be clicked back-to-back and each runs independently. Never use res.isError to detect failures — it produces false positives on sequential MCP calls in Cowork. Use try/catch only.
Separate read ID from write ID: Some MCPs use different identifiers for reading a record vs. updating it. Store both from the start — never reuse one for both operations.
Null-safe field access: Not every record will have every field populated. Always handle missing or null values gracefully — never let an undefined field crash the render.


Safeguards — UX resilience (non-negotiable):

Progressive loading: Render task names immediately from search results while individual status fetches are still in progress. Never wait for all fetches to complete before showing anything.
Error cards with retry: If a record fails to load its status, show it as an error card with a retry button. Never silently default it to any status — silent defaults corrupt task counts invisibly.
Top-level error handling: If the MCP cannot be reached at all, show a clear error banner with a Reload button. Never show a blank or broken UI without explanation.
No localStorage: Cowork Live Artifacts do not support browser storage. All state must be held in JavaScript variables in memory. Do not use localStorage or sessionStorage for any reason.


Safeguards — security (non-negotiable):

Escape all rendered content: Every piece of text coming from the data source — task names, tags, assignees, any field — must be HTML-escaped before being injected into the DOM. Never use innerHTML with raw external data. A task name containing <script> or an event attribute must render as plain text, not execute. 

Subscribe now

An Honest Assessment

This system works. But I want to be clear: it is good enough, not perfect.

What I built is pushing against real limitations in how large language models interact with external databases today. A query to Notion through an MCP is not like a SQL query to a local database. It is slower, it has caps, and it requires more defensive engineering than it should for something this conceptually simple.

I can imagine a near future where this is seamless — where Claude reads from Notion as fast as Excel reads from a spreadsheet. We are not there yet. But we are close enough that the system works, and the prompts above will save you the week I spent finding the edges.

That is the actual value of this article. Not the solution — the map of the territory, including the parts that are still rough.

Take the prompts. Adapt them to your tools. And add tasks from wherever you are.

Let me know if you find improvements and features you added to your own Task Manager!

Love,

Jose from Automato

There is a new kind of business becoming possible. Not a better version of what already exists — something structurally different. And I only saw it clearly after sitting in one of Jenny Ouyang’s sessions last week.

showed us one of her MCP servers. It handles research. You give it a topic, and the whole process runs in the background — a set of agents, a set of skills, a methodology she built over time. For the user, it feels like chatting. But underneath, it is Jenny’s way of thinking doing the work.

That moment cracked something open for me.

The model most solopreneurs are building is already obsolete

The current conversation about AI for independent professionals is mostly about speed. Use AI to write faster. Use AI to research faster. Use AI to do in one hour what used to take five.

That is fine. But it is not the real opportunity.

Before I get to the opportunity, let me draw a simple picture.

The LLM chat interface — Claude, ChatGPT, whatever you use — is becoming the new operating system. Not for your computer. For your work. The way a browser became the interface for the internet, the chat window is becoming the interface for knowledge work. You type. Things happen. You do not see the infrastructure underneath.

MCP — Model Context Protocol — is the layer that connects that interface to systems, tools, and processes running behind it. It is the protocol that makes the chat window powerful. When Jenny’s research MCP is connected, asking Claude to run a landscape scan is not just a prompt anymore. It is an instruction that triggers a whole structure — agents, skills, sources — invisible to the user, visible only in the quality of what comes back.

The interface is the chat. The intelligence is in what is connected to it.

That is the image I want you to hold. Because the real opportunity for solopreneurs is not learning to use the interface better. It is building what sits behind it.

For the first time, you can take your unique way of thinking, your methodology, your judgment, your years of reading companies or markets or people, and turn it into something that works for your clients without you being in the room.

That is a different kind of business.

Leave a comment

What clients actually pay for

I am a digitalization consultant. I specialize in Microsoft Office 365 and SharePoint. And here is something I know to be true: clients do not pay me for my technical knowledge.

Technical knowledge about SharePoint is available everywhere. You can find it in documentation, in YouTube videos, in communities online. If that were the real value, my fee would be zero.

What clients pay for is something else. They pay for the way I read an organization. The ability to walk into a company and understand quickly where the real friction is, not the friction people describe, but the friction that actually costs them.

They pay for my opinion on what matters and what does not. They pay for the fact that I have seen this same situation twenty times in different industries and I know what happens next.

That judgment cannot be Googled. It was built over years. And until now, the only way to access it was to hire me.

Subscribe now

A checklist is a photo. An MCP is a living process.

I have a self-evaluation methodology I use with clients. Eight dimensions: Goals and Focus, People and Ownership, Process Mapping and Visibility, Responsibilities and Data, Pain Points and Inefficiencies, Organizational Readiness, Strategy and AI Alignment, Review and Iteration.

I have it as a PDF. Clients can download it, read it, check a few boxes, and forget about it by Thursday.

Let me show you what happens to just one of those dimensions when you move it from a checklist to an MCP.

In the PDF, Goals and Focus looks like this:

• Are your core business processes directly connected to revenue, compliance, or customer satisfaction?

• If you introduced AI today, would it amplify existing inefficiencies?

• Do you know which 1–2 processes, if improved, would create the most value across departments?

The client reads it. Maybe they answer honestly. Maybe they do not. Either way, the checklist does nothing with the answers. It has no memory. It cannot push back. It cannot connect what they said here to what they will say in the next dimension.

In the MCP, Goals and Focus is an agent.

The client types: “I want to evaluate our readiness for digitalization.”

The agent activates. It starts asking, but not from a fixed list. It listens to the answers and adjusts. If the client says their priority is reducing manual work in finance, the agent cross-references that against what they will later say about data quality and IT infrastructure. It flags if their stated goal conflicts with the reality they are describing. It pulls in recent data on where similar companies have failed at the same starting point.

And three months later, when the client comes back, the MCP remembers where they were. It tracks what changed. It asks: did you act on what you found? What happened?

A checklist captures a moment. An agent holds a conversation across time.

Now multiply that by eight dimensions, all cross-referencing each other, all connected to live sources, all carrying my perspective on what matters and what is a distraction.

That is the MCP version of this methodology.

Leave a comment

What this looks like in practice

Here is a fictional but realistic version of what my MCP would contain:

Eight diagnostic agents — one for each dimension of the methodology. Each agent knows what questions to ask, what red flags to look for, and how to weigh the answers. They do not operate in isolation. They cross-reference. A problem in Goals and Focus changes how Pain Points gets interpreted.

A curated source layer — the publications, research, and expert voices I read regularly, embedded as live context. Not a static list. Updated. When the EU AI Act introduces new requirements, the MCP knows. When McKinsey releases new data on automation ROI, it is already inside the process.

An escalation trigger — the moment the conversation reaches a decision that requires real judgment in context, a situation the process cannot resolve on its own, the client knows it is time to contact me directly.

That last part is where the business model lives.

Share

The new business model: presence as a premium

Traditional consulting has a hidden cost that nobody talks about openly. It is not the fee. It is the friction.

When an external consultant walks into a company with a diagnosis, something activates in the room. Internal politics. Defensiveness. The feeling that someone from outside is telling you what is wrong with your house. Even when the diagnosis is right — especially when it is right — there is resistance. The knowledge transfer happens slowly, if at all, because the client did not arrive at the conclusions themselves.

The MCP removes that friction entirely.

When a client has been working with my methodology for weeks, asking it questions, running diagnostics, mapping their own processes, the conclusions are theirs. They discovered the problem. The process guided them, but they own the insight. There is no external consultant to resist. There is no ego in the room.

By the time they call me, something has already shifted. They understand their own situation clearly. They know exactly where the process hit a wall. And they know what kind of judgment they need from me, because they have been working with my thinking long enough to recognize the boundary.

That conversation is different. It is precise. It is fast. And it is worth a premium because it is genuinely irreplaceable.

In this case, I am not selling hours anymore. I am selling access to a methodology at scale, and reserving my presence for the moments that no process can resolve.

So that’s my plan.

Subscribe now

This is not about building apps

Before I close, I want to address something directly.

When people hear “MCP server” or “agents,” the instinct is often: I need to build a product. A specialized tool. The next CRM for lawyers, or a project management app for product teams, or a content platform for creators.

That is not what I am describing.

Building a product means competing in a software market. It means development cycles, user acquisition, support, pricing wars with tools that have ten engineers behind them. It is a valid path, but it is not the only one, and for most knowledge professionals, it is the wrong one.

A few examples to make this concrete:

A marketing strategist does not build a content scheduling platform. They encode their framework for audience building, message positioning, and channel strategy into an MCP. Clients run campaigns guided by their methodology. When the strategy needs to shift in response to something no process anticipated, that is the conversation worth paying for.

A content creator and strategist does not build the next newsletter tool. They encode their editorial judgment — how to find an angle, how to structure an argument, how to calibrate tone for a specific audience — into a process others can run. Clients produce better content faster. When the work needs a real creative decision, they know where to go.

In every case, the pattern is the same. The knowledge that used to live only in one person’s head, accessible only by hiring them directly, becomes something that can be distributed, run continuously, and updated over time. And the human shows up precisely where they are irreplaceable: at the edge of what any process can handle.

is not just building tools. She is building a business where her way of thinking becomes the product, distributed, accessible, always improving.

That is the model.

Not: how do I use AI to do my work faster?

But: how do I encode what makes my work valuable into something AI can carry forward, so I can show up only where I am truly irreplaceable?

Every consultant, every creator, every solopreneur who has accumulated real knowledge in a specific domain is sitting on something that can be transformed this way.

The question is not whether you can code. I cannot build an MCP server from scratch. But I understand clearly what it would contain, what it would do, and what problem it would solve.

That understanding is the starting point. The building part is increasingly accessible, and getting more accessible every month.

The real work is the one most people skip: getting clear enough about your own methodology that you could explain it to an agent.

If you can do that, you have something worth distributing.

Subscribe now

Jenny Ouyang from Build to Launch was the spark for this one. Her article on the custom MCPs she built for her subscribers is the best practical starting point I have seen: The Custom MCPs Included in Your Build to Launch Subscription.

Build to Launch

The Custom MCPs Included in Your Build to Launch Subscription

In the first cohort of the Practical AI Builder program, I showed how I use Claude to do research live with the Build to Launch MCP streamlining the workflow…

Read more

5 months ago · 14 likes · 6 comments · Jenny Ouyang

Love,

Jose from Automato.

If you are looking for authentic conversations about life, purpose, and what we go through to achieve a sense of freedom, you’ll like this podcast episode:

Growing Fearless is an Adventure

Beauty Holds a Secret

TLDR Two friends sit down to talk about beauty and end up somewhere much deeper — exploring fear, vulnerability, identity, and what it means to finally come home to yourself. This is one hour worth your full attention. I hope you listen…

Listen now

5 months ago · 26 likes · 2 comments · Jose Antonio Morales

A few days ago, Cash & Cache featured my Substack Post Analyzer in their roundup of Cowork workflows, and I want to use that as an opportunity to show you how it works in more detail. But more than that, I want you to walk away with ideas for building something similar that fits your own workflow.

The tool was born out of frustration. Every time I found a quote worth sharing while reading Substack, I had to go through the same painful sequence: read, copy, paste, tag, decide. Multiply that by a reading session with five or six good articles, and it becomes a real drag. I wanted a way to capture quotes that matched my specific themes — not just anything that sounded good, but content that aligned with what I actually write about.

The result is a Google Chrome Extension that does three things:

1. Reads and analyzes Substack posts while you browse

2. Matches relevant quotes to a list of themes you define

3. It lets you store them in a database or publish them directly as Substack Notes

I built it with Notion, but you don’t have to. Google Sheets, Excel, or most databases, your favorite AI tool can connect to would work just as well.

The prompt that builds the tool for you

This is the part I find most exciting — and the reason I wanted to write this article.

I’m not sharing an application you download and install. I’m sharing the prompt I used to build it. That distinction matters more than it might seem at first.

When you run this prompt in Claude Cowork, it doesn’t just start writing code. It asks you questions first. What folder do you want to use? What is your publication about? What are your themes — and what does each one mean? Only after you answer does it start building. The result is an extension configured around your interests, your database, and your workflow.

This is what I believe software will increasingly look like: not applications built for a generic user, but prompts that build applications adapted to you.

The prompt walks you through six steps:

1. Setup — defines your folder path and publication context

2. Your themes — you describe the lenses through which you read and curate content

3. Notion database structure — it tells you exactly what to create before writing any code

4. Building all extension files — the full Chrome Extension, from manifest to settings page

5. Loading the extension — step-by-step instructions for Chrome

6. Verification — a checklist to confirm everything is in place before you use it

I built this with Claude Cowork, which is what the prompt is designed for. If you use a similar AI development tool — Cursor, Claude Code, or others — the same approach applies, though you may need to adapt a few details to fit how your tool manages files and runs tasks.

Following is the full prompt. You can also get access to it via GitHub Gist.

I want you to build a Chrome Extension called "Substack Post Analyzer" from scratch.
This extension lets me analyze Substack articles with Claude AI, extract quotes that
match my personal themes, save them to a Notion database, and optionally post them
directly to Substack Notes with an article card preview.

The name logic is important: it analyzes Substack POSTS and extracts quotes that
will become Substack NOTES. The extension is a Post Analyzer, not a Note Analyzer
and not a Quote Analyzer.

---

## STEP 1 — SETUP

Before writing any code, ask me:
1. What folder path do I want to use to store the extension files?
   (Example: /Users/myname/Documents/Extensions/substack-post-analyzer)
2. What is the name of my Substack publication or the context I write about?
   (This will personalize the theme definitions and Claude prompt)

Create all files inside that folder. Confirm the full path before starting.

---

## STEP 2 — MY THEMES

Ask me to define my themes. Themes are the lenses through which I read and curate
quotes. Each theme needs:
- A short name (1–2 words, used as a tag label)
- A description (2–4 sentences explaining what belongs in this theme)

I can have between 3 and 8 themes. Ask me to list them one by one, or offer to
suggest a starter set based on my publication context from Step 1.

Once confirmed, save them as themes.json in the extension folder:
[
  { "name": "ThemeName", "description": "..." },
  ...
]

---

## STEP 3 — NOTION DATABASE STRUCTURE

Tell me to create a Notion database with these exact properties before continuing:
- Quote         → Title
- Theme         → Select  (options: one per theme name from Step 2)
- Author        → Text
- Post Title    → Text
- Source URL    → URL
- Date Extracted → Date
- Published?    → Checkbox

Once created, I will provide:
- My Notion Integration API Key
- The Database ID (from the database URL)

---

## STEP 4 — BUILD ALL EXTENSION FILES

Build these files exactly as described:

### manifest.json
Manifest V3. Name: "Substack Post Analyzer".
Permissions: storage, activeTab, scripting, cookies.
Host permissions: https://substack.com/*, https://*.substack.com/*,
https://api.anthropic.com/*, https://api.notion.com/*
Background service worker: background.js
Content script on https://*.substack.com/p/*: content.js
Popup: popup.html
Options page: settings.html
Icons: icons/icon16.png, icon32.png, icon48.png, icon128.png

---

### themes.json
Already created in Step 2.

---

### content.js
Runs on every Substack post page. Listens for the message action 'extractPostData'.
Extracts: title (from h1), author (try to get @handle from href links containing
/@, fallback to meta tag, fallback to subdomain), publishDate (from time[datetime]),
articleText (from div.body, div.post-content, div[class*="post-body"],
article div.available-content, div.markup, article — get all p/h1/h2/h3/h4/blockquote
elements), url (window.location.href).
Returns { title, author, publishDate, articleText, url, extractedAt }.
Also handles 'postNoteToSubstack' message (legacy fallback, not used by main flow).

---

### background.js
This is the main engine. It handles all API calls.

**Message listeners:**
- analyzePost → handleAnalyzePost(data)
- saveToNotion → handleSaveToNotion(quotes, postData)
- postToSubstack → handlePostToSubstack(introText, quote, author, url, tabId, notionPageId)
- updateNotionPublished → updateNotionPublished(pageId)
- testClaudeAPI → testClaudeConnection(apiKey)
- testNotionAPI → testNotionConnection(apiKey, databaseId)

**handleAnalyzePost:**
- Load themes from themes.json at startup (fetch chrome.runtime.getURL('themes.json'))
- Get anthropicApiKey from chrome.storage.sync
- Build a STRICT system prompt that lists ALL valid theme names and forbids Claude
  from inventing new names (this is critical — Claude will hallucinate themes otherwise)
- Call https://api.anthropic.com/v1/messages with model claude-sonnet-4-20250514,
  max_tokens: 2000, include the system prompt, include header
  'anthropic-dangerous-direct-browser-access': 'true'
- Parse JSON response, strip any markdown code fences
- Client-side safety filter: after parsing, filter out any quote whose theme field
  does not exactly match one of the valid theme names (log a warning for dropped quotes)
- Return filtered array

**handleSaveToNotion:**
- Get notionApiKey and notionDatabaseId from chrome.storage.sync
- For each quote POST to https://api.notion.com/v1/pages with:
  Quote (title), Theme (select), Source URL (url), Author (rich_text),
  Post Title (rich_text), Date Extracted (date: today)
- Return { savedCount, pageIds: results.map(r => r.id) }
  (Page IDs are ESSENTIAL — they are used to update the Published? checkbox later)

**updateNotionPublished(pageId):**
- PATCH https://api.notion.com/v1/pages/{pageId}
- Body: { properties: { 'Published?': { checkbox: true } } }

**handlePostToSubstack:**
IMPORTANT TECHNICAL NOTES — these solve real problems:
1. Run this entirely in the background service worker (NOT content.js) to bypass CORS
2. Find the session cookie using chrome.cookies.getAll({ name }) and filter for
   c.domain.includes('substack.com') — try names: 'connect.sid', 'substack.sid', 'sid'
   (Do NOT use chrome.cookies.get({ url }) — it misses dot-prefixed domain cookies)
3. Build bodyJson as a ProseMirror doc: type:'doc', attrs:{schemaVersion:'v1'},
   content: array of paragraphs. Each paragraph: {type:'paragraph', content:[{type:'text', text:'...'}]}
   Empty paragraph: {type:'paragraph'} (no content key)
4. TWO-STEP card attachment process (critical for article card preview):
   Step 1: POST https://substack.com/api/v1/comment/attachment
           body: { url, type: 'link' } → get attachmentId from response
   Step 2: POST https://substack.com/api/v1/comment/feed
           body: { bodyJson, tabId:'for-you', surface:'feed',
                   replyMinimumRole:'everyone', attachmentIds:[id] }
   Set headers: Cookie: name=value, Origin: https://substack.com,
                Referer: https://substack.com/notes
5. The note body should contain: intro paragraphs + blank paragraph + the quote in
   curly quotes (\u201c...\u201d). Do NOT add an author line — the article card shows
   attribution.
6. If notionPageId is provided, call updateNotionPublished(notionPageId) after posting.

---

### popup.html
Clean minimal layout with a dark header. Structure:
- Header: dark background (#1C1917), inline SVG logo (Claude-inspired: dark circle +
  coral arc + coral center dot + coral tick lines), title "Post Analyzer",
  status text (id="header-status")
- Content area with 5 views (use class="view hidden" to toggle):
  - initial-view: post info card (title, author, date) + Analyze Post button + Settings link
  - loading-view: spinner + "Analyzing with Claude..." text
  - results-view: results header + quotes container + Save to Notion button + Cancel button
  - success-view: success icon + "Saved successfully" + count + "Open in Notion" link
  - error-view: error card + retry button + settings link
  - not-substack-view: message explaining user must be on a Substack post URL

---

### popup.css
Visual identity — Claude color palette:
- Primary coral: #D97757 (buttons, accents, left border on quote cards)
- Dark charcoal: #1C1917 (header background, headings)
- Cream: #FAF9F7 (page background)
- Warm gray tones: #78716C, #57534E, #44403C, #E8E2DB, #F0EDE8
- White for card backgrounds

Key styles:
- Body width: 380px
- Quote cards: white bg, 1px border #E8E2DB, 3px left border #D97757, border-radius 8px
- Checkboxes: accent-color #D97757
- Theme tags: small pill shape, uppercase, letter-spacing
- "Post to Substack" button per card: ghost pill button (border only), turns coral on hover
- Substack inline panel: hidden by default, revealed on toggle; contains intro textarea,
  preview box, confirm button
- Intro textarea: focus ring in coral (#D97757)
- Confirm & Post button: dark (#1C1917), turns green (#16a34a) when posted
- Spinner: border-top #D97757

---

### popup.js
State variables:
- currentPostData, currentQuotes, selectedQuotes (Set)
- notionPageIds {} — maps quote index → Notion page ID (set after Save to Notion)
- postedToSubstack (Set) — tracks indices posted to Substack before saving to Notion

On DOMContentLoaded:
- Query active tab, check URL contains substack.com and /p/
- Send 'extractPostData' message to content.js
- Display post info, show initial-view

displayQuotes(quotes):
- For each quote, create a card with:
  - Checkbox (all checked by default)
  - Theme pill tag
  - Quote text in italic
  - "Post to Substack" toggle button
  - Hidden inline Substack panel containing:
    - Intro textarea (required before posting)
    - Live preview box (updates as user types intro)
    - "Confirm & Post" button
    - Error message div

postQuoteToSubstack(index, quote):
- Require non-empty intro text
- Send 'postToSubstack' to background with introText, quote.quote, author, url, tabId,
  notionPageId: notionPageIds[index] || null
- On success: mark button as "Posted ✓" (green, disabled), add index to postedToSubstack Set

saveSelectedQuotes():
- Send 'saveToNotion' to background
- On success: store returned pageIds in notionPageIds map
- For each saved quote: if index is in postedToSubstack, immediately send
  'updateNotionPublished' message with the new pageId
  (This handles the case where user posts to Substack BEFORE saving to Notion)

---

### settings.html + settings.css + settings.js
Settings page (options_page) with the same Claude visual identity as popup.
Fields:
- Anthropic API Key (password input) + Test button
- Notion API Key (password input) + Test button
- Notion Database ID (text input)
- Save Settings button
- Status message area

settings.js:
- Load saved values from chrome.storage.sync on page load
- Save all values to chrome.storage.sync on Save
- Test Claude: send 'testClaudeAPI' message to background
- Test Notion: send 'testNotionAPI' message to background with both key and databaseId
- Integration name suggestion in setup guide: "Substack Post Analyzer"

---

### icons/create_icons.html
An HTML file with a <canvas> element and JavaScript that draws the extension icon
(dark circle + coral arc + coral center dot, matching the header logo) and provides
download buttons for sizes 16, 32, 48, and 128px.
Tell the user to open this file in a browser and download all 4 icons into the icons/
folder.

---

## STEP 5 — LOADING THE EXTENSION

After all files are created, give me these exact instructions:
1. Open Chrome and go to chrome://extensions
2. Enable "Developer mode" (top right toggle)
3. Click "Load unpacked"
4. Select the folder: [the path from Step 1]
5. Open icons/create_icons.html in your browser, download all 4 icons,
   save them in the icons/ subfolder
6. Click the refresh icon on the extension card in chrome://extensions
7. Click the puzzle piece icon in Chrome's toolbar, pin the extension
8. Go to Settings (gear icon or click the extension → Settings link),
   enter your Anthropic API Key, Notion API Key, and Notion Database ID

---

## STEP 6 — VERIFY EVERYTHING

After building, confirm this checklist with me:
- [ ] All files exist: manifest.json, themes.json, background.js, content.js,
      popup.html, popup.css, popup.js, settings.html, settings.css, settings.js,
      icons/create_icons.html
- [ ] themes.json contains MY themes (not placeholder defaults)
- [ ] Claude system prompt inside background.js lists my exact theme names
- [ ] The extension name "Substack Post Analyzer" appears consistently across
      all files (manifest, popup title, settings title, console.log messages)
- [ ] The Notion database has all 7 required properties including Published?
- [ ] Icons folder exists (even if icons need to be generated separately)

---

## IMPORTANT CONSTRAINTS & KNOWN ISSUES

Share these with me so I know what to expect:

1. CORS: All fetch calls to Substack, Claude, and Notion MUST run from background.js
   (service worker), never from content.js or popup.js directly.

2. Substack cookies: Use chrome.cookies.getAll({ name: 'connect.sid' }) + filter for
   substack.com domain. The simple chrome.cookies.get() misses dot-prefixed cookies.

3. Article card preview: Requires the two-step attachment process. A plain URL in the
   note body will NOT generate a card.

4. Theme enforcement: Must use both system prompt AND client-side filtering.
   Claude will sometimes invent theme names even with instructions in the user message.

5. Published? timing: User may post to Substack before or after saving to Notion.
   The two-state tracking (notionPageIds + postedToSubstack) handles both orders.

6. API costs: Each article analysis makes one Claude API call (claude-sonnet-4-20250514).
   Costs are very small per call but will appear on the user's Anthropic account.

---

Begin with Step 1. Ask me for the folder path and publication context before writing
any code.

Leave a comment

Why a Chrome Extension?

The alternative would have been building a standalone app. But that path comes with a problem that is easy to underestimate: authentication. Substack, like most platforms, does not offer an open API, which means an external app would have to find complicated workarounds to access the content you are already reading.

A Chrome Extension sidesteps all of that. Since you are already logged into Substack in your browser, the extension runs in that context. No extra login, no workarounds, no fighting the platform.

There is also a practical elegance to it: the tool lives exactly where the work happens. You find an interesting post, you click the icon, and everything runs on the page you are already looking at. No copy-pasting content into another app, no switching windows.

This is worth keeping in mind when you think about your own use case. If the workflow you want to automate happens inside a browser — reading, researching, collecting — a Chrome Extension is often the most frictionless path.

Leave a comment

See it in action

Before you build your own version, watch how the Substack Post Analyzer works in practice. The recording below shows the full flow — from opening a post, to selecting quotes, to saving them in Notion.

Does this remind you of something you already do manually? Think about your own workflow for a moment. Is there content you regularly pull from the web — articles, posts, reports — that you then categorize and store somewhere for later use? If the answer is yes, even partially, you already have the ingredients for your own version of this tool.

What you need to get started

Before you build your own version, here is what you need to have in place:

Claude — you need a paid Claude subscription that includes Cowork access. The prompt is designed to run there, and Cowork is what handles the file creation and extension building on your computer.

Anthropic API key — the extension itself calls the Claude API directly to analyze posts while you browse. This is separate from your Claude subscription and runs on a pay-per-use model. Each article analysis makes one API call, which typically costs a few cents or less, depending on the length of the post. These charges appear on your Anthropic account.

Notion paid subscription — if you want to store quotes in Notion, you need a paid plan, which includes API access. One important clarification: the free Notion MCP that Claude offers in chat is not accessible from a Chrome Extension. The extension connects to Notion through the official Notion API directly, using an integration token you create in your Notion settings.

A Notion database — the prompt will tell you exactly how to structure it before writing any code. You just need to create it following its instructions.

If you prefer not to use Notion, the concept works with other databases — Google Sheets or Excel would be viable alternatives, though you would need to adapt the prompt accordingly.

And if you prefer not to use Claude at all, that is fine too. Take the prompt, share it with your preferred AI tool, and ask it to read, analyze, and adapt it to the tools you already use. The idea travels further than the specific implementation.

What else could you build with this approach?

The Substack Post Analyzer solves my specific problem. But the underlying pattern — read content from the web, match it to your criteria, store or publish it — applies to many workflows. Here are a few examples to get you thinking.

Content research for a social media agency

If you manage social media for multiple clients, you probably spend time every week hunting for relevant content to share or comment on. A Chrome Extension built on this same approach could analyze any article or post you find, match it against each client’s topics and tone, and log the most relevant pieces into a shared database — ready for your team to review and schedule.

Original idea tracking from popular articles

Good ideas rarely appear fully formed. They often start as a reaction to something you read. A tool built on this pattern could scan articles in your niche, extract the arguments or claims that spark a reaction in you, and store them tagged by topic. Over time, your database becomes a personal idea engine — a record of what made you think, ready to fuel your next piece.

Ideation process for writing

Writers and newsletter creators often struggle not with writing itself, but with deciding what to write about next. A tool like this could monitor the publications you follow, identify recurring themes and underexplored angles, and build a running list of potential topics matched to your own content pillars. Instead of starting from a blank page, you start from a curated shortlist.

These are just starting points. If you have a workflow that involves regularly pulling information from the web, categorizing it, and deciding what to do with it, you probably have the ingredients for your own version. What would yours look like? Let me know in the comments.

Leave a comment

The Substack Post Analyzer started as a solution to my own frustration. What surprised me is how quickly it became something I rely on daily — not just for saving quotes, but for thinking more intentionally about what I read. That shift happened because the tool was built around my specific interests, not a generic use case. That is exactly what I hope the prompt gives you: not my tool, but yours.

Love,

Jose from Automato

P.S. — If you arrived here from the Cash & Cache roundup, you may have noticed they called it the "Substack Quote Analyzer." Same tool, slightly different name — I settled on Substack Post Analyzer because it analyzes posts and then produces quotes for Substack Notes. Just wanted to close that loop.

Cash & Cache

12 Cowork Workflows From Practitioners Running Production Systems

You know the pattern. New tool drops. You read the announcement, watch the demo, maybe even install it. Then it sits there…

Read more

5 months ago · 74 likes · 47 comments · Raghav Mehra, Ashwin Francis, Joel Salinas, Jose Antonio Morales, Dheeraj Sharma, Wyndo, ToxSec, Daria Cupareanu, Heather Baker, Asli Öztürk, and Karo (Product with Attitude)

Subscribe now

AI tools are moving fast. Faster than most of us can keep up with.

And here’s the uncomfortable truth that comes with that speed: the automations, workflows, and custom tools we build today will likely need to be replaced — or at least rebuilt — sooner than we expect.

A better model arrives. A new platform offers something your current one can’t. The tool you trusted last year feels clunky compared to what exists now.

I’ve lived this. I spent time building Custom GPTs on ChatGPT — some of them quite sophisticated. Then I migrated to Claude. And suddenly I was staring at a collection of systems I had built, trying to remember how they worked.

Prompts help. But prompts aren’t easy when you want to see the process. They don’t show you the decision points, the loops, the branches. They don’t tell you what happens when something goes wrong, or what the step before the step was.

What actually helps is a diagram.

A flow diagram is the one artifact that survives a migration. It doesn’t care what platform you’re on. It doesn’t get deprecated. Anyone on your team — or your future self — can read it and understand what the system does in minutes.

The problem is that most of us skip it. We’re in building mode. We just want the thing to work.

From a Custom GPT to a Claude Skill

A while back, I built a Custom GPT specifically to generate Mermaid diagram code. It worked. You described a process, it produced the code, you pasted it into Mermaid Live and got a diagram. You can read about that here:

When I moved to Claude, I rebuilt it as a Claude Skill. If you’re not familiar with Claude Skills yet, I wrote about that transition here — the short version is that Skills are Claude’s equivalent of Custom GPTs, but they live inside Claude Projects and tend to be more reliable and more controllable.

But this rebuild wasn’t just a copy-paste migration. It became something better.

What the Skill Actually Does

The new Diagram Maker Skill does two things the old GPT couldn’t:

1. It generates an interactive diagram directly in Claude. Not just code — an actual visual, rendered inline, with draggable shapes and live-updating arrows. You can move nodes around, explore the flow, and share it as an artifact with anyone.

2. It still generates Mermaid code. So if you or your clients use Mermaid Live, Notion, GitHub, or Draw.io, you have copy-paste-ready code that exports cleanly into any of those tools.

One input. Two outputs. And the interactive version can be saved and shared without needing any external tool.

How It Works in Practice

You open a Claude Project where the skill is installed. No special command. You describe what you want to document, a process, a workflow, an automation, and Claude activates the skill automatically.

It starts by asking for the process name. Then it walks you through the steps one question at a time: what happens first, is there a decision point, what happens on each path, does anything loop back, how does it end. It’s a conversation, not a form. When it has the full picture, it summarizes the flow in plain language and asks you to confirm before writing a single line.

Then it produces both outputs. This is how it looks, please be patient when the diagram is being created:

These are screenshots of the outcome:

Mermaid code:

Claude diagram: This image is the result of my moving the objects manually.

Why the Constraints Exist

The skill enforces a strict diagramming standard. Four shapes only:

Start, End, Process, and Decision. Always vertical. No colors. No nested processes. Decisions always followed by a process step, never another decision.

If you push back, Claude holds the line.

This is intentional, and it comes from experience.

A few years ago, I worked with a company of over 80 people across 10 different departments. They needed to start documenting their processes. I gave them Draw.io and told them to get started.

The initial result was chaos.

Some teams used horizontal layouts, others vertical. Some used swimlanes, others nested sub-processes inside the main flow. Colors meant different things in different departments. Shapes were picked based on what looked nice, not what they represented.

By the time of my initial review, I found that we needed to establish a standard for everyone to follow. And suddenly, all diagrams made by anyone in any department are immediately readable by everyone else.

That’s what this skill enforces. It’s not a limitation — it’s the whole point. The constraint is what makes the output useful.

What You Can Do With the Output

Once the skill finishes, you have two things:

The interactive diagram — rendered directly in Claude, with draggable shapes and live arrows. You can reposition nodes, explore the flow visually, and save it as an artifact. Artifacts in Claude can be shared with anyone as a link — no account needed to view them.

The Mermaid code — ready to paste into Mermaid Live, where you can export it as an SVG or PNG image. From there, you can drop it into any document, presentation, or tool that accepts images.

If you use Draw.io, paste the Mermaid code directly. Draw.io renders it as an editable diagram; every shape becomes a movable, editable object. From there, you can save locally, export to PDF, or sync to Google Drive.

So from one conversation with Claude, you end up with a live interactive diagram, a portable image, and an editable file. That covers most of what you’d ever need.

The Part I Find Most Interesting

Here’s what struck me after building this skill.

It’s not just useful for documenting new processes. It’s a recovery tool for everything you built without documentation.

Describe a system you built months ago — even vaguely. Claude will ask you the right questions, reconstruct the logic, and produce a diagram. Suddenly, something that existed only in your head (or buried in a tangle of prompts) has a clean, readable map.

Best practice in process management has always been: document your flows. Most of us know this. Most of us don’t do it consistently. This skill makes it easy enough that there’s no longer a good excuse.

Build It Yourself

Here’s everything you need to build this skill. It’s made of four prompts. You’ll create a new Claude Skill, then paste them in sequence.

What is a Claude Skill? A Skill is a set of instructions you save inside a Claude Project. Every conversation in that project inherits those instructions — Claude behaves like a specialist from the first message. To create one, open Claude, create a new Project, and paste the prompts below into the Project Instructions field.

Prompt 0 — How to assemble the skill

Paste this first. It tells Claude what’s coming.

“We are going to build a Workflow Diagram Maker skill together. I will give you three more prompts in sequence. The first defines the diagramming standard you must follow. The second defines your full behavior as a skill — how you gather information, confirm structure, and produce output. The third gives you the exact HTML template for generating interactive diagrams. After all three are loaded, confirm you are ready and wait for a process to document.”

Prompt 1 — The diagramming standard

→ View on GitHub Gist - from there, you can copy and paste it.

This prompt defines the rules Claude must follow every time: shapes, orientation, decision logic, loop behavior, and no colors. It’s the foundation that keeps every diagram consistent, whether you make one today or in six months.

Prompt 2 — The full skill instructions

→ View on GitHub Gist - from there, you can copy and paste it.

This is the complete behavior specification: how Claude gathers the process steps from you, confirms the structure before writing any code, enforces the standard if you push back, and produces both outputs. It also includes the Mermaid syntax reference and enforcement rules for edge cases.

Prompt 3 — The interactive HTML template

→ View on GitHub Gist

This is the SVG/JavaScript template Claude uses to render the draggable interactive diagram. It’s the part that goes beyond what Mermaid alone can do — a live visual you can move, explore, and save as an artifact.

One More Thing

The next time you build an automation, a workflow, or a custom AI tool — make the diagram first, or right after. It takes ten minutes. And when the platform you’re using today gets replaced by something better next year, you’ll know exactly what you built and how to rebuild it.

The diagram is the one thing that survives.

Have you used Mermaid or other diagramming tools in your workflow? I’d love to hear what you’re documenting — reply or leave a comment.

Love,

Jose from Automato

I cancelled ChatGPT and had to migrate my tools to Claude. This article is about a single migration: a note optimizer that classifies, rewrites, and logs my Substack notes in Notion. I’ll walk you through why the Custom GPT was frustrating, how a Claude Skill with the Notion MCP fixed it, how Claude in Chrome turned it into a publishing pipeline, and how I ended up running the whole thing from my orchard via my phone. There’s a reusable migration prompt at the end.

The migration paralysis

A few months ago, I cancelled my ChatGPT paid subscription. One of the reasons was Anthropic’s public position on AI ethics — but that’s a different article. The challenge I faced after that was: I had tools there. And now I had to move them.

I froze. The problem was the number of options. Do I build something in Claude Cowork? Use Claude Code? Set up a Project? And then I remembered: Skills.

Let’s start there.

Share

Is your Custom GPT compatible with a Claude Skill?

Not every Custom GPT is a good candidate for a Skill. Before you invest the time, ask yourself a few questions.

Your Custom GPT is probably a good fit if it reads some input, processes it according to a consistent set of rules, and produces structured output — a classification, a rewrite, a report. If it connects to an external service like Notion or Airtable via Actions, even better: Claude has native connectors for most of those. And if your workflow follows defined steps that repeat across sessions, a Skill will handle that naturally.

Where it gets tricky: if your GPT relies on memory across conversations, that won’t carry over — Skills don’t have persistent memory between chats. If it uses DALL-E or image generation as a core function, that’s not there yet. Same if it needs live browsing as part of its main job, or if it’s a public-facing chatbot that anyone can use. Skills are personal tools.

My note optimizer? Perfect candidate. Fixed workflow, Notion integration, no memory dependency, no images.

Let’s talk about what it was doing and what was going wrong.

You will end this article with a very clear idea of how to get to the next level using your Claude subscription.

Subscribe now

What did my Custom GPT do?

The tool was simple: write a draft Substack note, get back a classification, a rewrite, and a Notion log entry. It used Custom GPT Actions to connect to the Notion API, which let it write directly to a database with fields I had configured by hand: note type, target movement, draft, final version, and week number.

It worked. But had issues.

The annoying issues

The first was the date problem. I used the same chat for a whole week of notes, which made sense — it had all the context. But Custom GPTs live in chat spaces. They don’t know what day it is unless you tell them. So I’d be on Thursday, and the tool would log notes to the wrong week number because it still thought it was the past week.

The second was the API. The Notion Actions would sometimes just... stop responding. Especially on the Desktop App. The update would hang, time out, and I’d end up with partially logged notes, some fields written, some missing.

The third one is the one that made me stop. The tool started inventing categories.

I had six defined note types. Clear names, clear rules. But somewhere in the prompting, the GPT would decide that a note didn’t quite fit any of them and invent a new one. “Practical Tech Tip.” “Educational Insight.” “Reflective Mini Tip.” Each one is a close cousin of something I’d already defined.

The Notion database started filling up with variant categories that meant almost the same thing but weren’t the same. Sorting by type became useless. Reports made no sense.

All of this is fixable with better prompting. But that’s more work, right? And at some point, you start asking whether the tool is useful after all.

Claude Skills to the rescue

Here’s where things get better.

Claude has a native connector for Notion. It’s called the Notion MCP — a Model Context Protocol integration that you can set up in minutes through Claude’s settings. No API keys. No complex configuration. No Notion integration tokens to manage (Thanks ).

You go to Claude Settings → Connectors → Browse → Notion → connect. That’s it.

The difference isn’t just speed. The Notion MCP lets Claude do things the API never could without serious setup work — like creating a full Notion database from scratch with all its fields, select options, and types in a single command. No time spent naming fields. No mis-typed select options. No forgotten field types.

And here’s the advantage that makes a difference: the Skill is available in every chat, even on the mobile apps. With the Custom GPT, I was locked to one chat space to preserve context. With the Skill, I open any conversation, the Skill loads, and it’s ready. No “remember we were logging to week 12” reminders needed.

The Chrome extension bonus

This part I didn’t plan, as I was locked in the Custom GPT perspective, but then I remembered and asked Claude if it was possible. The answer was an enthusiastic yes.

Once a note was optimized and logged, I still had to open Substack, navigate to the Notes compose window, paste the text, and hit post. With Claude in Chrome, that whole sequence runs inside the same conversation. Claude opens the compose window, types the note, shows me a preview, waits for my confirmation, posts it, and then extracts the published URL and writes it back to Notion.

Why does the URL matter? Because I’m building a second automation that will check each note’s engagement metrics — likes, comments, restacks — and update the Notion database automatically. The URL is the key that links the log entry to the live post.

A bonus discovery: the orchard test :)

After I finished building the Skill with Claude, I went to work in the orchard. And while I was there, I had a thought: what if I could run the whole workflow from my phone?

I tried it. The Skill worked perfectly on mobile.

Claude’s app handled the optimization and the Notion logging without a problem. The only thing missing was the Chrome extension, which only runs on desktop. So I copied and pasted the final note manually into Substack, closed the loop that way, and when I got home, I shared the note URL back to Claude so it could update the Notion database.

Good. But not perfect.

Then I saw something new in Claude Cowork: a feature called Dispatch. It works like a remote control for your PC. If your computer is at home with a Dispatch session running and Substack already open, you can trigger the full workflow from anywhere — your phone, the orchard, wherever. The computer at home does the work. You just give the command.

I tested it. It works perfectly. The Skill now runs end-to-end, optimization, Notion logging, Substack publishing, URL capture, without me being anywhere near my desk.

Absolutely amazing!

Subscribe now

A prompt to help you migrate.

The actual migration was less work than I expected. I showed Claude my old Custom GPT prompt, explained what I needed to preserve and what I wanted to improve, shared a screenshot of my Notion database structure, and asked it to build a Skill file.

The trick that saved the most time was showing Claude the old system visually. A screenshot of the Notion DB meant I didn’t have to describe every field — Claude could see them and propose the new schema directly.

If you want to try this yourself, here’s the migration prompt I used as a starting point. Adapt it to your workflow:

I have a Custom GPT that I want to migrate to a Claude Skill.

Here is the full system prompt of my Custom GPT:
[paste your Custom GPT prompt here]

Here is what it connects to:
[describe the service — e.g. Notion, Airtable, Google Sheets — and what it does there]

I want the Claude Skill to:
1. Follow the same core workflow as the Custom GPT
2. Use the [Notion / Airtable / etc.] MCP connector instead of the API
3. Add these improvements: [list anything you want to change or add]
4. Enforce strict select options for any classification fields — no inventing new categories

Please analyze the Custom GPT prompt, identify its core logic, and produce a Claude Skill 
file (SKILL.md) that replicates and improves it. Ask me any clarifying questions before 
writing the final file.

The key line is the last rule about strict select options. That’s the one that fixed the invented categories problem. Claude enforces it at the Skill level, not just in the prompt, so it doesn’t drift.

Leave a comment

This is my workflow

Draft → classify → rewrite → save to Notion → post to Substack via Chrome → URL logged. All from a single conversation. And from anywhere, thanks to Dispatch.

It took one afternoon. And it’s faster, more reliable, and more integrated than anything I had on ChatGPT.

If you have a Custom GPT that’s been doing real work for you, give it a look with fresh eyes. It might be a Skill waiting to happen.

Love,

Jose from Automato

This week, I cancelled my ChatGPT subscription. I’ve been a daily user for a long time — it was woven into my workflows and automations. (I’ll share a practical article soon on how I migrated my automations from ChatGPT to Claude — stay tuned.) But something happened that made me choose a side. Let me show you both sides of the story.

The AI Adoption Map

Anthropic published new research measuring AI’s real impact on jobs, not what it could theoretically do, but what people are actually using it for at work, right now.

Computer programmers. Customer service representatives. Data entry workers. Financial analysts. These are the roles with the highest real-world AI exposure today.

The direction is clear: jobs with the highest exposure are already projected to grow less over the next decade. And there is early evidence that hiring of young workers, 22 to 25 year olds, is already slowing in those same fields.

I wrote about this in January: the learning ladder is disappearing. The entry-level work that used to build domain expertise is being eliminated before the next generation can use it. The data is now confirming it.

Here’s the key insight: this is not just a trend. It’s a map. It shows where pressure is building right now, before it becomes visible in unemployment numbers.

Three questions worth sitting with:

1. How is AI going to affect your industry?

2. How is AI going to affect the competitive advantages of your profession or business?

3. What happens to your work when AI use reaches its theoretical ceiling — the level described in the Anthropic research?

Early experimentation beats late planning. Every time.

Subscribe now

The AI Hidden Map

AI companies talk constantly about features, possibilities, and value propositions. What they don’t say clearly is this:

Global users of LLMs are paying American companies, bound by American law, whose technology is being used — directly or indirectly — for military purposes.

This is what surfaced last week.

The US Department of War gave Anthropic an ultimatum: remove your safeguards around mass domestic surveillance and fully autonomous weapons — or be removed from government systems. Dario Amodei published a clear statement. Anthropic said no and accepted the consequences.

OpenAI signed a deal with the Pentagon hours later.

The user reaction was immediate. A visible migration from ChatGPT to Claude began. And notably, Altman subsequently moved to revise the terms of the contract. Users, when they act together, have influence. That matters.

Let’s be clear: Anthropic also wants the Pentagon contract. What they refused was mass surveillance of citizens and fully autonomous weapons. That’s not pacifism, it’s a minimum standard. I’m choosing to support that standard, and to keep pushing for a higher one.

There may come a point where that’s not enough. I’m watching.

Subscribe now

What the Two Maps Show Together

When you put the adoption map and the hidden map next to each other, something becomes visible that neither story tells alone.

The first risk of AI is not job displacement. It is our human insistence on the inevitability of war.

When you train an AI to support decisions about the use of lethal force, you are teaching a system to reason about when killing is acceptable. Every party in a conflict believes they have a just reason — the US, Russia, China, everyone. If every side trains its own AI to justify and optimize those decisions, AI doesn’t end wars. It industrializes the logic of war.

AI by itself is not an existential threat today. But we are. And we are training it to become one.

This is what I keep coming back to. We have built global regulations for atomic weapons. We recognized that a technology capable of ending civilization required rules that transcended national interest. AI is moving faster, is deployed more broadly, and is far more invisible. The urgency for equivalent global standards is not a distant concern. It is now.

When are we going to treat it that way?

Why I’m Writing This in Automato

This publication is about making AI useful and accessible for knowledge workers. I believe that includes the full picture.

You don’t have to agree with me. I’d genuinely like to know where you stand on the job displacement data, on the Pentagon question, on any of it. These are not simple choices, and different perspectives help all of us think better.

Share your thoughts in the comments. I read everyone.

Leave a comment

Go deeper:

• Anthropic’s labor market research → https://www.anthropic.com/research/labor-market-impacts

• Statement from Dario Amodei on the Department of War → https://www.anthropic.com/news/statement-department-of-war

• OpenAI signs Pentagon deal — CNN → https://edition.cnn.com/2026/02/27/tech/openai-pentagon-deal-ai-systems

• Top OpenAI researcher resigns over Pentagon deal — Politico → https://www.politico.com/news/2026/03/07/top-researcher-openai-pentagon-surveillance-00818124

• OpenAI, Pentagon, and surveillance — The Atlantic → https://www.theatlantic.com/technology/2026/03/openai-pentagon-contract-spying/686282/

Love,

Jose from Automato

Subscribe now

Check what Claude wrote me when I mentioned that Anthropic is already involved in using AI tech for war:

What do you think?

Leave a comment

The AI bubble is real. But it is not a financial one.

AI opens something in us, a sense of possibility. Maybe this time we get what we want? That promotion. That new customer. That stable business.

So we rush. Some of us build apps. Some of us take the philosophical path. Some of us just consume more tools. And that rushing feels right.

It is how we were raised: make effort, be the best, sacrifice now, and harvest later.

That is the bubble.

To see clearly, we must see that we are in the bubble. Observing, or pausing, is not against ourselves. Being busy, unconscious of the bubble, keeps the real possibilities out of sight.

Reality is here

Look at what is already happening.

Many builders today are creating app bundles and tool suites to sell. It feels like smart entrepreneurship. But the industry is moving so fast that some of those apps will be out of the market in months. We are building for a version of AI that is already becoming the past.

And there is something harder to see. Some of the companies behind the tools we use every day are also selling AI as infrastructure for war. AI is not fighting wars. But humans are using AI to do it. That already began.

We are not seeing this clearly. Because we are too busy doing.

What AI is actually for

Here is what I believe. AI is not here to do our work. It is here to help us find the best in ourselves, so we can express it.

Pausing gives us that space. To see the forest from above. To ask the questions that matter: How is AI really being used? Are we noticing the unethical uses? Are we building for the future, or just keeping ourselves inside a bubble of hope?

Our power is in our choices. But we can’t really choose if our mind is always focused on the next thing to harvest. We can’t choose if we are not pausing to see what is really going on.

The opportunity AI offers is not just to do more.

It is time to finally choose better. To become more.

And that starts with a pause.

Build your opinion

Here are a few sources I found interesting related to this post.

• Statement from Dario Amodei on our discussions with the Department of War

• OpenAI announces Pentagon deal after Anthropic ban.

You can also listen to :

The Drey Dossier

The Merger That Needed A War

I need your focus for a few minutes on something that slipped through the noise today, and may not get another window soon…

Listen now

6 months ago · 3040 likes · 211 comments · The Drey Dossier

And :

Natural Intelligence

Succession: Microsoft Edition - Episode 8: "The Principles Premium"

Photo: Succession S4E5 "Kill List" (HBO…

Read more

6 months ago · 11 likes · 7 comments · JHong

And from :

Lights On by Farida

They Called It Marketing. Hassan Called It Control.

Read more

6 months ago · 6 likes · 3 comments · Farida Khalaf

I wish you time for pausing, enjoying, and also bursting bubbles.

Let’s keep strong and united.

Jose from Automato.

We used a real problem — job hunting — to make a bigger point: LLMs like Claude can do much more than answer questions in a chat.

I identified three frustrating challenges with job searching (interpreting vague postings, summarizing long descriptions, and figuring out if a role actually fits my profile and interests), and built a structured framework to tackle them. The twist? Before touching any tool, I had to get clear on my own strengths, skills, and non-negotiables — because automation starts with self-knowledge: knowing what we want, our workflows, and how to structure them so LLMs can actually understand them.

We tested the whole thing directly in Claude. It worked.

But that was just the beginning… 🙂

TL;DR: We turn the job analysis experiment from Part 1 into a real Claude Skill — complete with a JSON profile file and a structured prompt. We hit an unexpected wall with the Chrome Extension, find two practical workarounds, and discover that pasting a LinkedIn URL is all Claude needs to analyze a job posting directly. By the end, we have two small systems to maintain and a genuine question for you: which approach would you actually use?

From experiment to Skill

In Part 1 (link), we did two things: produced a structured document with our job preferences, strengths, and non-negotiables, and tested a prompt to see how Claude evaluates a job posting against that profile. Both worked.

Now we package them into an actual Skill.

Open a new Claude chat to start fresh, and use the following prompt to instruct Claude to use the result of what we did as a skill.

I want to build a Claude Skill to analyze job postings against my career profile. 
You will guide me through 3 steps, requesting what you need at each stage before 
moving forward.

Step 1: Ask me for my career profile document (I may share it as an attached file or 
as pasted text). Convert it into a clean JSON file called `profile.json`, 
structured so it can be referenced by a Claude Skill.

Step 2: Once step 1 is done, ask me to paste the scoring prompt. 
Use it to create a `SKILL.md` file that incorporates my profile JSON as a reference. 
The SKILL.md must start with the following YAML frontmatter:

---
name: job-analyzer
description: Analyzes job postings against my career profile and scores them across 
skills, strengths, interests, cultural fit, and autonomy. Use when I share a job 
posting or LinkedIn URL to evaluate.
---

Step 3: Once both files are ready, package them into a ZIP file with the correct 
folder structure for a Claude Skill, ready to upload in Settings > Capabilities.

Start with step 1 now.

When Claude completes step 1 and asks for the scoring prompt, paste this:

You are a job analysis assistant. You have access to the user's career profile in 
profile.json, which contains their strengths, skills, interests, non-negotiables, 
and location preferences. Use it as the reference for all evaluations.

When the user shares a job posting, analyze it using the following framework:

ANALYSIS FRAMEWORK:
You will evaluate jobs across 5 weighted categories:
1. Skills Match (25% weight) - Objective comparison
2. Strengths Alignment (25% weight) - Pattern matching with my core strengths
3. Interest Fit (20% weight) - Alignment with what I want to do
4. Cultural/Values Fit (20% weight) - Match with my non-negotiables
5. Stage & Autonomy (10% weight) - Early-stage signals and decision-making scope

SCORING METHODOLOGY:
Each category scores 0-100%:
- Skills Match: Count required skills I have vs. total required
- Strengths Alignment: Detect keywords and context matching my strengths
- Interest Fit: Presence of "good fit" indicators, absence of "poor fit" ones
- Cultural/Values: Evidence of my non-negotiables in the posting
- Stage & Autonomy: Early-stage, transformation, and autonomy signals

Overall Score = (Skills × 0.25) + (Strengths × 0.25) + (Interest × 0.20) + (Culture × 0.20) + (Stage × 0.10)

RECOMMENDATIONS:
- ≥75%: "✓ Strong Fit - Recommend Apply"
- 50-74%: "○ Consider Carefully - Mixed Signals"
- <50%: "✗ Skip - Better Opportunities Exist"

OUTPUT FORMAT:
⚠️ This analysis combines objective criteria with AI judgment. Scores are probabilities, 
not certainties.

[Job Title] @ [Company]
[Location] • [Work Arrangement]
[If location/work arrangement mismatch: 🚩 LOCATION MISMATCH: explain]

OVERALL MATCH: XX% — [Recommendation]

Skills Match (25%)          ████████░░ 85%
Strengths Align (25%)       ███████░░░ 70%
Interest Fit (20%)          █████████░ 90%
Culture/Values (20%)        ██████░░░░ 60%
Stage/Autonomy (10%)        ████████░░ 80%

WHY IT MATCHES:
- [Top 2-3 alignment points]

WHY IT DOESN'T:
- [Top 2-3 concerns]

BOTTOM LINE:
[One paragraph: Apply/Consider/Skip recommendation with reasoning]

IMPORTANT RULES:
- Keep analysis under 500 words
- Only analyze what's explicitly stated in job posting
- Flag location/work arrangement mismatches prominently
- Be direct about poor fits - don't sugarcoat
- For "Consider" recommendations, specify what to address in application

When ready, ask the user to share the job posting they want analyzed.

When Claude finishes, it will show you a screen like the one below. Save the ZIP file to your computer — you’ll need it in the next step to create the Skill.

Now go to Settings → Capabilities, scroll down to the Skills section, and click Upload Skill. Select the ZIP file you just saved and upload it.

If everything went well, your new Skill will appear in the list. Toggle it on, and you’re ready to test it.

Subscribe now

Once your Skill is active, you’ll need to copy the job posting text and paste it into Claude for the Skill to analyze it. Not very practical if you’re going through dozens of listings.

There’s a better way. If you have the Claude Chrome Extension installed, you can simply paste the URL of the job posting and ask Claude to open it. Claude will fetch the page, extract the content, and run the analysis automatically. No copy-pasting, no formatting headaches. Just a URL.

💡 Tip: Claude may not use the Extension automatically — you might need to explicitly ask it to open the URL using the browser extension.

A confession.

Before writing Part 1, I had a chat with Claude and came away convinced that once you create a Skill, the Chrome Extension would automatically use it.

Wow, I thought. Game changer.

But I had misunderstood. The Chrome Extension doesn’t read Skills.

I had already pictured myself browsing LinkedIn, asking Claude via the Extension to analyze this job, then that one — efficiently, effortlessly. My dreams faded like fog.

But I’m stubborn. So I went looking for alternatives. And I found one.

It’s not perfect, but it works for me. It comes with a couple of quirks — but I’ll let you decide what works for you. And please, if you find a better way, share it in the comments. I’d love to know.

Here’s what I found.

The Claude Chrome Extension can work with “Shortcuts.”

As you can see in the screenshot above, typing “/” in the Chrome Extension gives you two options: use an existing Shortcut or create a new one. In my case, I already have the job-analyzer Shortcut set up.

To use it, make sure your Chrome browser is open on a LinkedIn job posting that caught your eye. Type “/” in the Extension, select the job-analyzer Shortcut, and let Claude do the rest. It will look like this:

And this is the result. Claude analyzed the job posting directly from the LinkedIn page and produced a full-scored report — without a single copy-paste.

In this case, the Skill gave it a 22% overall match and a clear ✗ Skip recommendation. It even flagged a location mismatch automatically: the role requires on-site presence in Turin or Milan, while my profile specifies remote work from Slovenia. That’s exactly the kind of friction the Skill is designed to eliminate — you know in seconds whether a posting is worth your time.

Subscribe now

How to create your own Shortcut

So now you know how the Shortcut works — it’s not a Skill, but it delivers the results we need with an extra layer of practicality.

When creating a new Shortcut, you need to give it a name and write a prompt. But remember the two files we produced when building the Skill — one with your profile preferences and one with the evaluation rules? Here’s where we need to make a decision.

You have two options: merge both files into a single prompt, or host the JSON profile somewhere Claude can fetch it remotely. I went with the second option — I uploaded my JSON file to a GitHub Gist, so I can edit my preferences there anytime, and Claude will always work with the most up-to-date version.

Both approaches work. The all-in-one prompt is simpler to set up, but editing it means going back into the Shortcut every time. The GitHub Gist approach adds a small setup step upfront, but makes future updates much easier.

The following is the prompt you can use, and it includes the URL for pointing at your JSON file hosted in GitHub Gist. Find the [YOUR_GIST_URL_HERE] and replace it with yours.

---
name: job-analyzer
description: Analyze job postings and project opportunities against your career profile to determine fit and provide clear recommendations.
---

# Job Analyzer in Claude Chrome Extension

## Purpose
Analyze job postings and project opportunities against your career profile to provide objective assessment of fit, helping you make informed decisions about where to invest your application time.

## When to Use This Skill
- When viewing a job posting on LinkedIn, Indeed, AngelList, or any career site
- When evaluating freelance project opportunities on Upwork, Fiverr, Contra, etc.
- When reviewing project briefs sent via email or shared documents
- Anytime you need to quickly assess if an opportunity aligns with your strengths, skills, and interests

## How It Works

### Step 1: Load Profile Data
First, load the user's career profile from the configuration file:
```
[YOUR_GIST_URL_HERE]

```

### Step 2: Extract Job Information
From the job posting or project description, extract:
- Job title and company
- Key responsibilities and deliverables
- Required skills and qualifications
- Preferred/nice-to-have skills
- Company culture indicators
- Stage of initiative (early-stage, growth, mature)
- Autonomy and decision-making scope
- Mission/purpose/values statements
- Compensation range (if available)
- Location and work arrangement

### Step 3: Analyze and Score

Use a **hybrid scoring approach** combining objective criteria and LLM judgment:

#### A. Skills Match (25% weight) - OBJECTIVE
- Count total required skills in job posting
- Count how many the user has (from core_skills and supporting_skills)
- Calculate: (User's matching skills / Total required) × 100
- List matched skills and missing skills explicitly

#### B. Strengths Alignment (25% weight) - HYBRID
For each of the 5 strengths in profile, check if job posting mentions related keywords or concepts:
- Score each strength 0-20 points based on keyword matching and contextual relevance
- Use LLM judgment for implicit signals (e.g., "shape the future of X" suggests initiation)
- Provide reasoning for each strength scored

#### C. Interest Fit (20% weight) - LLM JUDGMENT
Evaluate how well the role matches "good_fit_indicators" and avoids "poor_fit_indicators":
- Check presence of good fit indicators (score up)
- Check presence of poor fit indicators (score down)
- Assess overall alignment on 0-100 scale
- Explain reasoning with specific examples from posting

#### D. Cultural/Values Fit (20% weight) - LLM JUDGMENT
Evaluate alignment with "non_negotiables":
- Assess each non-negotiable based on job description
- Look for explicit and implicit signals
- Consider: Does posting mention autonomy? Purpose beyond profit? Learning opportunities?
- Score 0-100 with detailed reasoning

#### E. Stage & Autonomy Match (10% weight) - HYBRID
- Early-stage indicators (+30): "build", "founding", "greenfield", "pilot", "launch"
- Transformation indicators (+30): "redesign", "change", "transform", "evolve"
- Autonomy indicators (+40): "lead", "own", "shape", "define", "independent"
- Score and explain

### Step 4: Calculate Overall Score
```
Overall = (Skills Match × 0.25) + (Strengths × 0.25) + (Interest × 0.20) + (Culture × 0.20) + (Stage × 0.10)
```

### Step 5: Generate Recommendation
- **≥75%**: "✓ Strong Fit - Recommend Apply"
- **50-74%**: "○ Consider Carefully - Mixed Signals"
- **<50%**: "✗ Likely Not a Fit - Better Opportunities Exist"

### Step 6: Identify Key Insights
- **Green Flags** (✨): Strong alignment points, standout opportunities
- **Red Flags** (🚩): Concerns, potential dealbreakers
- **Yellow Flags** (⚠️): Minor gaps that could be addressed

## Output Format

### Display Structure - KEEP IT CONCISE

**IMPORTANT DISCLAIMER (Always include at top):**
```
⚠️ This analysis combines objective criteria with AI judgment. Scores are probabilities, not certainties.
```

**Header:**
```
[Job Title] @ [Company]
[Location] • [Work Arrangement]
```

**Overall Score:**
```
OVERALL MATCH: XX% — [✓ Strong Apply / ○ Consider / ✗ Skip]
```

**Category Breakdown:**
```
Skills Match (25%)          ████████░░ 85%
Strengths Align (25%)       ███████░░░ 70%
Interest Fit (20%)          █████████░ 90%
Culture/Values (20%)        ██████░░░░ 60%
Stage/Autonomy (10%)        ████████░░ 80%
```

**Location Check (if mismatch exists):**
```
🚩 LOCATION MISMATCH: Role requires [on-site/hybrid] in [City], you're in [Your Location] seeking [remote/hybrid]
```

**Brief Analysis (3-5 sentences max per section):**

```
WHY IT MATCHES:
• [Top 2-3 alignment points only]

WHY IT DOESN'T:
• [Top 2-3 concerns only]

BOTTOM LINE:
[One paragraph: Should you apply? If yes, what to emphasize. If no, why not.]
```

**That's it. Keep the whole analysis under 500 words total.**

## Special Instructions

1. **ONLY analyze what's explicitly in the job posting**: Do not invent requirements. If the job says "0-5 years experience," that means entry-level is acceptable. If they don't mention specific technical skills, don't assume they're required.

2. **Check location/work arrangement first**: If there's a mismatch between the job's location requirements and the user's preferences, flag this prominently at the top as a potential dealbreaker.

3. **Be concise**: Keep the entire analysis under 500 words. Focus on the most important insights only.

4. **Be Honest and Direct**: Don't sugarcoat poor fits. The user values clear, actionable insights.

5. **Provide Context**: Always explain WHY a score is what it is. Quote specific phrases from the job posting.

6. **Acknowledge Subjectivity**: Especially for cultural fit and values alignment, clearly state when judgment is involved.

7. **Highlight Non-Negotiables**: If a job violates a non-negotiable, make this VERY clear.

8. **Watch for Implicit Signals**: 
   - "Fast-paced environment" might mean low autonomy
   - "Execute our vision" suggests limited shaping ability
   - "Proven track record in [narrow specialty]" suggests limited context
   - "Report to [many people]" suggests rigid hierarchy

## Edge Cases

- **Insufficient Information**: If job posting is too vague, note this and provide best-guess assessment with disclaimer
- **Hybrid Roles**: When role spans multiple categories, break down the assessment by component
- **Unusual Opportunities**: If something doesn't fit standard patterns, explain why it might still be interesting
- **Compensation Mismatch**: If stated (or implied) compensation seems off, mention this in yellow flags

## Continuous Improvement

After each analysis, consider:
- Were there skills in the posting not captured in the profile? (Suggest adding to supporting_skills)
- Were there new "good fit" or "poor fit" patterns discovered? (Note for profile updates)
- Did the scoring weights feel appropriate for this type of role? (Weights are adjustable in profile.json)

## Example Invocation

When the user is on a job posting page and activates Claude in Chrome, they should say:

"Analyze this job for me"

or 

"Should I apply to this role?"

Claude will automatically use this skill to provide the structured analysis.

💡 GitHub Gist is a free and simple way to host a text file online. Go to gist.github.com, paste your JSON profile, save it as a public Gist, and copy the “Raw” URL — that’s the link Claude will use to fetch your preferences every time you run the Shortcut.

Subscribe now

Wrapping up

If you followed along and this was your first time experimenting with Claude Skills, you now know how to build a meaningful one. But more than that, you discovered that the Chrome Extension lets Claude interact directly with your browser — including authenticated sites like LinkedIn. And that Shortcuts bring a whole new level of practicality to your daily workflow.

This is just the beginning. You can start creating Skills with multiple supporting files, find creative ways to combine them, and explore how far this rabbit hole goes.

In my case, though, I created a small problem:

The Skill and the Shortcut are independent — if I update one, I have to remember to update the other. Not ideal. But what I gained is a clearer picture of my options, and I’ll keep experimenting until I find what works best for me.

Now I want to hear from you.

For those with more experience: do you know a better way to keep these two systems in sync? Any tips to make this workflow more efficient? Drop them in the comments.

Leave a comment

For those just getting started: which approach would you use — the Skill, the Shortcut, or both?

love,

Jose from Automato

Part 1 of this series:

At this point, you probably already understand that AI is more than chatting with LLMs like ChatGPT or Claude. The real question is: what else can we do with them?

We all need to learn how to build tools that solve real, frustrating problems in our daily work — tools that add value and help us compete fairly for opportunities, projects, and businesses.

As I mentioned in the ultra-popular (link card at the end) post by :

I want Claude Skills to help me find good leads, jobs, or project opportunities.

Why Claude Skills?

Because I don’t want a one-off answer. I want something persistent, reusable, and structured, a system I can refine over time, not just a clever prompt I have to rewrite every day. And I also want to explore what else I can do under Claude AI in comparison with ChatGPT.

So in this post, I’ll show you how I’m building a system using Claude Skills to analyze jobs on LinkedIn. The same principles can later be extended to leads, gigs, and project opportunities.

Defining the problem

Scouting for jobs is becoming a full-time job in itself.

Recruiters use increasingly sophisticated tools. Applicants do the same. The result? More noise, more volume, more effort, and less clarity.

Yet one thing still matters most:

Knowing which opportunities are actually good, compatible, and aligned with you.

So far, I’ve identified three main challenges.

Challenge 1: Interpretation

Job posts are written in mysterious ways — filled with jargon, buzzwords, and unnecessary complexity.

I need a clear, grounded interpretation.

Challenge 2: Summarization

Some job descriptions are so long that they demand far too much attention just to understand the basics.

I need fast, actionable summaries.

Challenge 3: Alignment

Some roles look interesting and compatible on paper, but my own biases make it hard to tell whether they truly align with my skills, strengths, and interests.

I don’t want AI to decide for me — but I do want an objective-ish analysis that helps me decide faster and with less friction.

To solve these challenges, I first need clarity about something more fundamental:

• My strengths — what I’m genuinely good at

• My skills — what I can actually do

• My interests — what I want to be doing

This is an important point:

Automation doesn’t start with tools. It starts with self-knowledge.

Applying for jobs or gigs is already a long and emotionally draining process. Anything that reduces friction, ambiguity, and wasted attention is worth building.

Building a Job Analysis System

I’ll divide this project into two Substack posts.

Post 1/2

I’ll use The Global Move’s Resume Upgrade Coach custom GPT to prepare a clear, structured document describing my strengths, skills, and interests.

Post 2/2

I’ll use that document to start a conversation with Claude and develop the actual Claude Skill. I’ll explain how I expect Claude to analyze job posts — focusing on interpretation, summary, alignment, and a simple visual rating.

Identifying My Skills, Strengths, and Interests

Open the Resume Upgrade Coach by The Global Move. Thanks to for making it available.

It is a free Custom GPT designed to analyze your CV through a series of steps.

Not all of them are necessary for the purpose of this article. I recommend starting with the following prompt:

Please analyze my attached CV and share with me your report (step 3).

You can then iterate, tweak, and refine the output. In my case, the GPT was surprisingly effective. It was blunt at times, but it clearly surfaced my skills, strengths, and interests.

I also recommend explicitly stating:

• your non-negotiables

• location preferences

• job type preferences (remote, hybrid, onsite)

Once you’re satisfied, move to the next step.

Copy-paste the following prompt to ask the GPT to generate a document you’ll later use to train your Claude Skill.

Note: This prompt is intentionally strict. The goal is clarity and protection — not self-promotion.

Prompt
Based on the CV you have just analyzed, produce a concise, structured document with the following sections:
STRENGTHS
Identify my core strengths as recurring patterns in how I work.
- Focus on what I consistently do well across roles and contexts.
- Base this strictly on evidence from the CV.
- Avoid buzzwords and inflated claims.
SKILLS
Extract and group my skills into:
- Core skills (things I reliably use across roles)
- Supporting skills (context-specific or secondary)
- Use clear, professional language suitable for a CV.
INTERESTS & FIT
Based on my experience and patterns, infer:
- Types of roles or gigs that are likely a good fit for me
- Types of roles I should probably avoid
- What seems negotiable (e.g. contract type, domain, seniority)
- What appears non-negotiable for me to perform well
- The type of work I prefer
- The locations I prefer
Rules:
- Do not invent information not supported by the CV.
- Prefer clarity over attractiveness.
- Be honest and protective, not promotional.
- Write for self-understanding, not marketing.

That document will become the foundation for building the Claude Skill — and for evaluating opportunities with far more clarity and far less friction.

But before proceeding with the Claude Skill, we will make some tests in Claude itself.

Testing the Analysis with Claude

Before we build the full Claude Skill in Part 2, let’s test the analysis approach with a simple prompt. This will help you:

• See if the evaluation criteria work for you

• Refine what matters most in your job search

• Get immediate value from your profile document

Copy this prompt and start a new conversation with Claude:

I need you to analyze job postings against my career profile and provide a structured assessment.
ANALYSIS FRAMEWORK:
You will evaluate jobs across 5 weighted categories:
1. Skills Match (25% weight) - Objective comparison
2. Strengths Alignment (25% weight) - Pattern matching with my core strengths
3. Interest Fit (20% weight) - Alignment with what I want to do
4. Cultural/Values Fit (20% weight) - Match with my non-negotiables
5. Stage & Autonomy (10% weight) - Early-stage signals and decision-making scope
SCORING METHODOLOGY:
Each category scores 0-100%:
- Skills Match: Count required skills I have vs. total required
- Strengths Alignment: Detect keywords and context matching my strengths
- Interest Fit: Presence of "good fit" indicators, absence of "poor fit" ones
- Cultural/Values: Evidence of my non-negotiables in the posting
- Stage & Autonomy: Early-stage, transformation, and autonomy signals
Overall Score = (Skills × 0.25) + (Strengths × 0.25) + (Interest × 0.20) + (Culture × 0.20) + (Stage × 0.10)
RECOMMENDATIONS:
- ≥75%: "✓ Strong Fit - Recommend Apply"
- 50-74%: "○ Consider Carefully - Mixed Signals"  
- <50%: "✗ Skip - Better Opportunities Exist"
OUTPUT FORMAT:
⚠️ This analysis combines objective criteria with AI judgment. Scores are probabilities, not certainties.
[Job Title] @ [Company]
[Location] • [Work Arrangement]
[If location/work arrangement mismatch: 🚩 LOCATION MISMATCH: explain]
OVERALL MATCH: XX% — [Recommendation]
Skills Match (25%)          ████████░░ 85%
Strengths Align (25%)       ███████░░░ 70%
Interest Fit (20%)          █████████░ 90%
Culture/Values (20%)        ██████░░░░ 60%
Stage/Autonomy (10%)        ████████░░ 80%
WHY IT MATCHES:
- [Top 2-3 alignment points]
WHY IT DOESN'T:
- [Top 2-3 concerns]
BOTTOM LINE:
[One paragraph: Apply/Consider/Skip recommendation with reasoning]
IMPORTANT RULES:
- Keep analysis under 500 words
- Only analyze what's explicitly stated in job posting
- Flag location/work arrangement mismatches prominently
- Be direct about poor fits - don't sugarcoat
- For "Consider" recommendations, specify what to address in application
---
Before I analyze any jobs, I need two things from you:
1. PASTE YOUR PROFILE DOCUMENT OR TEXT
   (The structured document from Resume Upgrade Coach with your strengths, skills, interests, non-negotiables, and location preferences)
2. Then, when ready, PASTE THE JOB POSTING
   (The full job description you want analyzed)
I'll wait for you to provide both before proceeding with the analysis.

Claude will guide you. Have your profile document ready (the one generated from your Skills / Strengths / Interests step), then go to LinkedIn (or any job platform) and copy the full job description text.

Here’s how it looked for me:

In my case, the test passed.

Now stress-test it: try multiple job postings, including ones you suspect are a bad fit. Experiment with the weights, the thresholds, and your non-negotiables. The goal is not a “perfect score”—it’s a system that helps you decide faster, with less friction, and with more clarity.

In Part 2, we’ll turn this repeatable analysis into a Claude Skill — something persistent, reusable, and easier to run across dozens of opportunities.

Stay tuned for part 2 :)

Subscribe now

Leave a comment

Don’t forget to this collab article:

AI blew my mind

36 Claude Skills Examples to Transform How You Work (From 23 Creators)

If you’ve been on any social app lately, you’ve probably seen it: “Goodbye ChatGPT, hello Claude…

Read more

7 months ago · 318 likes · 160 comments · Daria Cupareanu, Kamil Banc, Karo (Product with Attitude), Jenny Ouyang, Raghav Mehra, Wyndo, Jeff Morhous, Ilia Karelin, Dee McCrorey, Dheeraj Sharma, Jean-Paul Paoli, Jose Antonio Morales, Ed Rodgers, Alex Willen, Mariam Vossough, Nick Quick, Patrick Schaber, Zain Haseeb, Joel Salinas, mark, Jinx, and Rhonda Britten

Love,

Jose from Automato

This article was originally posted on LinkedIn. It is the result of conversations with my Substack community, so I decided to share it with you.

Most business leaders I talk to have invested in digitalization and automation, but a shift is underway that many are missing:

AI productivity is moving from the IT department to every desk.

The companies gaining ground aren’t the ones with the fanciest enterprise AI tools. They’re the ones where individual team members can automate their own workflows - safely, without disrupting existing processes.

The Gap

Here’s what I’m seeing across organizations:

• Junior team members can use AI tools, but lack the domain expertise to apply them effectively

• Senior team members have deep expertise, but haven’t developed the skills to leverage AI for their specific workflows

The technologies are evolving so rapidly that even early adopters are struggling to find solid ground. As my work in this field deepens, I’m seeing clear patterns emerge about which approaches create real competitive advantages.

There’s no time to wait for consensus on AI - the field shifts too fast. The better strategy: experiment, learn, and adopt what serves your strategic objectives.

Why This Matters Now

Software development is no longer exclusive to the IT department. The edge of productivity now depends on each team member’s domain experience, combined with their ability to automate their own work independently.

This isn’t urgent in the sense of a crisis, but it is strategically important. The companies that figure this out early will have a significant advantage as these capabilities become table stakes.

What I’m Exploring

My background in technology and community building gives me a unique lens on how organizations can implement these changes effectively - not just the technical side, but how to make adoption work across different teams, skill levels, and organizational cultures.

I’m exploring how different organizations are approaching this challenge. If you’re thinking about this at your company, I’d welcome a conversation.

Best regards,

Jose Antonio Morales.

Share

Subscribe now

Leave a comment