But if you’ve ever felt the luxurious satisfaction of a tool that understands you at the core and helps you get your work done without you even having to explicitly ask, then you too might root for the survival of great applications.

To survive, 2 questions to ask:

1. Does it own durable and defensible moat against frontier models’s progress? (moat iceberg below)

2. Does it have clear positioning on what it’s really selling and does the market understand it?

Let’s explore through a few company examples: Notion, Linear, Harvey, Ramp, Sierra, Airtable, Canva, Figma, Atlassian, Shopify… (scroll to the bottom to see full company map).

Subscribe now

Part 1. Durable moats

If the product is a nice interface built on frontier inference with traditional per seat pricing - it’s just a wrapper. A moat is durable and hard to replicate as frontier models continue to get better.

Vertical moats

1. Permission and accountability

For vertical applications, the liability you’re willing to take on is the first defence against pure intelligence horsepower. This moat is central in Sarah Guo’s “the untrainable“ memo. These are often regulatory fields where the correctness of the work is private and walled off - hence untrainable.

The moat is in taking on accountability (often legally) for the outcome. Getting past the compliance gate and then building the trust with every customer takes incredible depth and time. It’s hard enough for one domain, frontier labs certainly can’t do that for every vertical.

• For Epic (Electronic medical record giant): the gate is federal regulation. Hospitals must run ONC-certified EHR technology to qualify for Medicare reimbursement. Epic has the ONC certification, signed HIPAA business associate agreements with every hospital, carries legal liability for every patient’s record, and is accountable for multi-year implementations across labs, pharmacies, and billing.

• For Harvey: the gate is liability/privilege. It has to pass every law firm’s security audit (SOC 2 grade) to fulfill the duty of client confidentiality, commit to zero retention and no training on client data, and maintain the ethical walls between matters. Harvey II launch strengthens the promise with per firm/per matter walls.

• For Sierra: the gate is contractual for each client. It needs SOC 2 at minimum, PCI DSS when payment is involved, HIPAA agreement when the client is in healthcare, telemarketing and bot-disclosure rules if agent talks on the phone, etc. On top of that, it takes on the liability for the outcome of its customer conversations (literally its business model) - a risk that a general model can’t easily take on.

2. Definition of good

Data alone is not a moat. Every company already has state data, which is often transferable through connectors and MCPs (unless you have moat 1). The defensible moat is the outcomes - what was done and the result, and the evals - the definition of good for those outcomes. You can only uniquely own this if the end to end work only happens in your product.

• Sierra‘s agents can log which save offer kept which subscriber.

• Ramp can see what thousands of companies pay for the same software to tell a CFO if they’re overpaying.

• Harvey open-sourced the Legal Agent Bench (LAB) benchmark (1,200 client-matter tasks graded against 75,000 expert-written rubrics) that it uses to post-train open models on.

  • Owning the definition of good allows them to distill specialized models. It just announced Tenet, its first model post-trained for legal reasoning, which seems to perform at frontier level at open-source cost.

  • They also acknowledge that there is no neutral test. (ie. LAB benchmark is not the definitive test.) Domain expertise adds defensibility. Different firms have different preferences for what “good” outcome looks like. So the benchmark is only one instrument, it also grades models with expert side-by-side review, and get extensive customer feedback (since they can’t train on customer data) in production. That is hard to copy.

• Notion just started building the horizontal version of this moat. Its new Knowledge Board evaluates 15 models on a random slice of live agent traffic. It’s judged by an ensemble built with the frontier labs on the tasks users actually delegate: meeting follow-ups, ticket triage, etc.

The horizontal moats

3. Agent consumability

A product is agent-consumable when its architecture allows agent to read and write and directly do things for the user without manipulating the GUI. The labs can absorb the UX but the object model is harder to absorb.

Notion‘s original block architecture set the perfect foundation for this. Users work inside the object model. By nature it’s built to be legible to machines. An agent can read the user’s (and the orgs’) own organizing decisions, which is the highest signal context in the workspace.

Notion didn’t have to build an agent abstraction layer on its existing product. The block system is the interface to both humans and models. This is why Notion is one of the few incumbents that managed a real architectural transition to agent-native vs traditional SaaS product with a AI chat bolted on. This architecture is defensible against frontier models because the structure a company builds inside Notion can’t be synthesized by a model from the outside-in, even if it can replace the product UX.

Canva is an example of when the shallow layer gets absorbed. It cut revenue growth forecast to 20% because customers are going to ChatGPT for the shallow use cases (since the image models got really good at design). The casual users didn’t depend on Canva’s structured design objects, they just needed high quality generation - which the models replaced.

The opposite bet is Atlassian: just had a great quarter while exposing Jira and Confluence to external agents, because the org pays for the system of record and switching cost is too high.

4. Owned unit economics

AI brought back COGS to SaaS business that had always benefitted from having zero marginal costs. Free tiers (PLG motion) worked because free users are free to serve. But AI features have real cost on usage. Canva had to throttle its popular AI features due to increasing cost. It’s hard to compete with the labs’ subsidies for token usage.

To build moat in unit economics, one way is to distill specialized models on your own task distribution so you own your inference costs. This requires having evals and outcome data from moat 2.

• Cursor built Composer 1 so its supply isn’t always rented from the labs, even though it is in the direct path of frontier model capabilities (coding).

• Harvey recognizes that the labs have way more resources and they can’t compete head on on compute, so they play the Moneyball game, leveraging the frontier ecosystem (neolabs, post training providers, harness tools) to post train open weight models efficiently without having to train frontier models from scratch. Recently announced Tenet, their first legal model.

The other way is to price for outcome, ie. charge for the work you can prove.

• Sierra and Intercom Fin are examples of this, they charging by resolution (outcome). The pricing model is GTM and reflects the moat you have. Pricing for outcomes bet on you owning the definition of good.

The fundamentals

These classic moats built durable software businesses long before AI, and they are even more rare and defensible now.

5. Distribution.

This is the hardest moat to build for new AI-native startups because it takes time.

1. Can you still acquire customers if your product becomes less differentiated?

• Slack was the better product. But Microsoft Teams came with the Microsoft 365 bundle that enterprise procurement already approved.

• Gemini is gaining consumer share by riding the distribution of Android and Search, not necessarily being a better model.

• Mark’s latest memo for Meta is betting its AI strategy on the distribution moat: delivering its assistant to the 3.5 billion people already on its apps every day.

This is the biggest (and maybe last) moat for large incumbent SaaS.

2. Do you own the channel, does it compound without sales effort? eg. a bundle, a marketplace, an installed base are assets

• Shopify‘s defensible layer evolved from being the software to being the marketplace for millions of merchants and buyers, with an app ecosystem and payment rails.

• Harvey is starting a network effect by selling to both law firms and in-house counsel, where the mutual benefit for legal operations could compound distribution.

3. Do you have strong lock-in’s & high switching cost?
   Large incumbents have an edge here again by being the system of record entrenched deeply within companies.

• Notion‘s org-wide knowledge base (context) is the lock-in,

• Harvey II launch announced per-firm specialization, where Harvey would partner with each firm to build models for their specific legal work. This becomes model lock-in.

6 Specialized craft and taste

Good old product craft: the taste that comes from deeply understanding your specific user and getting the nuance of their days right. This is obviously not new because of AI but it stands out more and can be felt more viscerally when products start looking the same.

• Linear consistently encodes a strong opinion about how software teams should run. Users (developers) can feel the difference in the speed.

• Ramp displays that taste for finance and ops users. Every default is biased toward the admin’s time. A controller can feel that the product was built for them.

• Sierra builds taste for each industry it serves. What a good conversation sounds like for a telco is different from a healthcare organization.

Taste shows up as an accumulation of every small decision, in the UX, in the model, in the entire interaction. Even if each piece is “copy-able”, the aggregation of the whole is not.

Share

Part 2. Positioning

Moats are for the business, not for the customer. Nobody buys a moat. Customers buy the feeling of getting to their desired outcome. Positioning is the translation of moats into that feeling. First is getting to clarity for what you’re really selling. Then it’s making sure the market understands it.

Shreyas has a great piece on this, do you know what you’re really selling? For example, Apple sells taste, Stripe sells deep care, X sells access, etc. When you get to this level of clarity (it’s extremely hard), it’s reflected in every corner of the product and the company, consistently, with conviction. Clarity makes it natural for you to teach the market of the criterion you want it to judge you by.

Notion‘s biggest moat (agent consumability) has become a burden for its positioning. The pre-AI Notion grew successfully with the “note-taking” use case. It used to sell “agency“. Its 2024 campaign is “Think it. Make it.” But the business is about enterprise collaboration for knowledge work.

The new campaign is “AI can’t replace teamwork. Think together.” It’s about enabling an organization to run on one shared brain where people and agents share the same context and think together. And yet, it’s hard to shed that “note taking app” branding from the public. People complain it’s doing too much, it’s too bloated, isn’t this a note taking app, etc.

I think Notion is now selling coherence. But it’s still in the processes of fulling committing to this new positioning. The homepage headline has moved to sell the collective, while the subtitle and feature list still sells the individual. Straddling between the two is what dilutes the main positioning with intensity and clarity.

Zuck’s reframing of Meta‘s positioning was a masterclass for updating the market’s criteria to judge winning by: from “who builds the best model” to “who has the largest distribution”.

The wrapper is important

i don’t mean the interface wrapper, but the branding/positioning wrapper. The discourse in the tech community is often about the moats and capabilities, but positioning sets the price because it sets how the market views you. What are you really selling? Does the market understand it the same way you do?

Share

More related reads

• the amount of value that can be created between the AI model and the ultimate end-user workflow is far larger than many people assumed or realized. Aaron Levie

• the untrainable - Sarah Guo

• no saas is safe - signull

• Harvey’s Sequoia talk about their moneyball playbook

Thanks for reading! If you liked this, hit the 🖤, it helps others see it :) Find me on X.

- Christine

Why is systems thinking now table-stake for PMs?

Software is designed, but models are grown, not designed. As such, AI products are organic. PM’ing AI products has an organic nature, akin to growing a garden. It’s more about creating the system for your plants to grow in rather than specifying exactly how an individual plant will look like. (See Alex Komoroske’s Gardening Platforms.)

To apply systems thinking in building AI/agents in existing products, these truths from Donella Meadows’ Thinking in Systems help explain the principles. (I read the book a few years ago, but only understood it after living through building agent systems in a platform PM role.)

1. Behavior arise from the interactions between the parts, not the individual parts.

A system’s output is the result of how the components connect and interact. You can perfect every individual part and still get a broken whole. The classic PM role (let’s call it the domain PM), especially at a large company, owns a specific scope/product area. By org design (and Conway’s law), every domain PM optimizes for their part. The product is not the part itself, but how the parts are organically assembled to a useful whole. Systems thinking means optimizing the connections between the parts. This matters more with agents because agents don’t know your org chart. Agents need to act holistically across the product on the user’s behalf.

2. Structure, rules, information flow have greater leverage than parameters.

The places to intervene in a system has a spectrum of leverage. Tweaking parameters has the weakest leverage. Next is changing the structure, the rules, and the information flows. Changing the paradigm has the highest leverage on the system. When experimenting with versions of a feature, it’s only parameter level intervention. Designing the contract that many teams can build on is a rule. Capturing the correction signal between what the agent did and what the user corrected is an information flow. Traditional product management is often filled with planning parameter tweaks. The higher-leverage product work is building the contracts and new information flows.

3. When there’s a new actor in the system, value goes to the interface.

Systems don’t exist on their own. Every system nests inside larger systems and they exchange with each other through interfaces. When agents are added to a product system as a new actor, new value goes to the interface for how the agent communicates with the product surface. For example, the shipping container was the interface for trade. API was the interface for software. MCP is the interface for agents and the tools they use. So what’s the interface for agents to use your product?

Building for the agent is systems work. Here are a few specific ways to apply systems thinking in PM’ing AI products - that I learned the hard way. Whether you work on a product domain or a platform, these might be helpful for PMs (especially in incumbent SaaS) trying to add agents in their products beyond just a bolt-on. It’s not a framework or a prescription, but examples of how the above principles were applied.

Subscribe now

Decompose the product

Traditional workflows and UI has always been built for human consumption (so called "UX"). But now the agent enters as a new actor in the product’s system. Most existing products aren’t composed in a way that is easily digestible by agents. New value comes from creating the interface that lets agents understand and act in the product. Let’s call this AX: “agent experience”. It requires thinking from the agent’s perspective, not just the user’s.

To make products more usable by agents, a product can be decomposed into a few ingredients: structured context (so agent can “see” what the user sees), typed actions (so agent can know what actions are available to take in the workflow), preconditions and rules for human intervention, and the coordination to know what action to take given the context.

Decomposing the product into these common primitives and providing the interface for agents to access them give you a chance to differentiate from the AI labs. Otherwise, you’ll end up with bolt-on “AI features” that don’t perform any better than ChatGPT/Claude etc., and your customers will balk at AI’s disruption to their usual workflows without delivering significant new value. New value can only be realized when agents are deeply wired inside the product and can understand and act in workflows that are already familiar to the customers.

Define the contracts

Once the product is decomposed, the parts need a contract (agreement) to specify how they should behave, so different product areas can plug and play to get to their agents to perform the desired behaviour.

Contrast this from the traditional feature-level PM’ing: building a feature on your product area alone only works on your product area. But a contract defines how a capability can work at a system level, what each component means, the rules by which they can interact, etc. This allows any team or agent to add that capability in their own surface or use case. It makes the capability much more scalable, consistent, and cohesive across the entire product. It’s like the electrical outlet - individual appliance maker don’t have to negotiate with the power company because the outlet gives them the contract. Contract thinking was not always needed for domain product PMs but is required when building agents, since agents need defined preconditions to operate in your product.

Design the loop

New features come with a lifespan. But a loop can improve the agent system over time. An example of a valuable loop is capturing the correction signal: the delta between what the agent did and whether the user corrected it. For example, when the user edits the field the agent updated, or accepts the draft but rewrites the last paragraph. Corrections are labeled examples from real usage.

The correction signal gives two benefits. One is eval data: signal for where the agent fails in production, which no offline benchmark can. The second is improvement data: corrections improve the agent by preference pairs or past corrections retrieved into context so the agent doesn’t repeat the same mistake.

Great AI products are built on these loops. Coding agents track what jobs are accepted, rejected, or edited after accepting. The bonus: this valuable signal is simply a byproduct of normal use of the product that doesn’t require any additional action from the user (unlike a feedback form). Improving the agent with this signal makes it perform uniquely better inside your product on subsequent use, compared to starting cold in generic AI chat products. Features depreciate, systems (with loops) compound.

An example:

Sierra’s newly launched Horizon platform is a great illustration of the systems thinking that’s required to conceive new and real value with agents. Horizon’s innovation is in the seams, the interface layer that can feed signals, goals, guardrails, and cross-interaction context to agents. It answers the differentiation question that most companies are struggling with: “How can you build a durable moat in your business as frontier AI progresses so rapidly?”?

Systems thinking’s velocity advantage

Stewart Brand’s “pace layers” explains that buildings work because the things that change fast (furniture, room layouts) aren’t welded to the things that are harder to change (the frame, foundation). The fast layer can change quickly because of the affordances built in the foundation layer that don’t have to change.

In software products, you might think decomposing and building contracts and loops sound slower than shipping that new AI feature in the next cycle. It’s the age-old tension between speed and scale. But in the agent era, the features are so much cheaper to build, so read speed actually comes more from investment in the system, the platform. The system level capabilities and contracts are the product thinking that can bring real differentiation in your existing (especially non-AI native) product. One contract allows you to test many variants of an experience in one sprint, and new variants can easily be added as a new surface on the same contract. But if that capability was built on individual surfaces, then each variant needs to re-implement the same plumbing. Platform is now the speed play.

What now

Elizabeth Stone on Lenny’s podcast shared that Netflix is hiring people who can look across the business domains and abstract the building blocks. It previously let local teams build off the paved path to move fast, but agents operating across multiple systems make common paved paths and solving problems once more important.

To practice systems thinking, pick a high traffic JTBD area and ask a few Qs:

• What would the agent need to see, know, and do at each step to complete this job accurately for the user? What are the structured ingredients it would need?

• How can your agent access those ingredients across multiple surfaces? Is your capability attached to your UI surface only or is it interchangeable with other areas of the product or of the customer’s workflow holistically?

• How are you capturing unique context from your customers and improving your agent?

Software used to be designed. But now, AI products are grown. A great PM/builder acts more like a gardener who creates the conditions, sows the seeds, and tends to what comes up. Don’t stay stuck in the traditional PDLC boundaries of defining the perfect feature and perfect roadmap. Think about the core primitives that make up the system, define contracts and loops, and watch the new organic value that can grow from your system and maybe surprise you.

More fun reads

• Lenny’s podcast with Elizabeth Stone: “we need more systems thinkers in a world with AI.”

• Sierra’s Horizon

• Why product sense is the only product skill that will matter in the AI age

• Seeing people clearly

• Why Elon’s companies consistently achieve wild successes

Share

Thanks for reading! If you liked this, hit the 🖤, it helps others see it :) Find me on X.

- Christine

These automations freed up a lot of my time. With that new found time (and Fable), I started pushing it for higher leverage work, hard tasks that I didn’t trust Claude with before. To my surprise, this made the biggest leap in my productivity, capacity, and sense of flow.

Using Claude for all 3 levels of product work

Shreyas has a lovely concept of three levels of product work: impact, execution, and optics. It’s natural to use Claude to automate busywork like optics & execution and leave the impact work to ourselves, because we tell ourselves the story that judgment is the defensible human thing. But when given the right context, Claude can do the hard work surprisingly well. Lowering the activation energy of starting is a major unlock in itself.

Notion surveyed 6,000 professionals across 10 markets on how orgs are using AI at work. At an individual level, if your org provides Claude/ChatGPT and connectors/MCP access, you can push your personal system further to the right and use Claude for all 3 levels of product work.

As a platform PM, I need to understand and work with the entire product org’s roadmap (thousands of PMs and developers). Here’s how I use Claude at each level of product work that’s allowed me to scale up my scope and scale down my stress.

Subscribe now

For Optics work: use Claude as autopilot to replace production.

Optics work is about making your team’s progress visible to the right people in the right forums. I’ve automated the majority of the production work and simply review/edit the output to use for different forums.

I have a scheduled task on Cowork that runs twice a week, it pulls context from key sources (Slack channels, calendar, meeting notes, git) and writes the latest updates for each work stream into a Google Sheet template. This G-sheet status board is visible for the teams to get their questions answered without asking me. I also have a skill that packages these updates at different levels for the right audience and forums.

Optics work is the layer to automate ruthlessly because it takes care of the dopamine backlog so you can stay in flow. I wouldn’t be surprised if these flows are productized soon in Claude itself.

Sample instruction template for scheduled task - iterate your own

You are updating the [your status sheet] for [name]. Your job is to gather the latest updates from Slack channels, a Google Sheet tracker, and the PM journal, then write a JSON update file to a Google Drive inbox folder (which an Apps Script time trigger picks up and applies to the sheet), and send [name] a compact Slack DM.

## STEP 1: Read current sheet state
[use google connector here]

## STEP 2: Gather signals (last 3 days)
### 2a. Slack channels
Read these channels using slack_read_channel (oldest = 3 days ago unix timestamp):
[insert your slack channels]

### 2b. PM Journal
Read the last 3 days of journal entries from the [folder]. These are local files at:
[insert filepath]
Read the 3 most recent .md files (by filename date). Journal entries contain to-dos, decisions, meeting notes, and context that may not appear in Slack. Treat journal entries as a first-party source alongside Slack messages.

## STEP 3: Determine changes

For each track and detail row, identify ONLY items with verifiable changes:
- Status changed? (e.g., “◉ In Dev” → “Complete”)
- New items shipped?
- Next milestone updated?
- New or resolved risks?
- Notes worth updating?

Be factual. Only report changes backed by [your sources]


## STEP 4: Push updates to the sheet via Drive inbox

Create a JSON file in the Drive inbox folder. The Apps Script time trigger (runs hourly) will pick it up and apply the changes to the sheet.

[work with claude on JSON payload]


## STEP 5: DM Christine

Send a Slack DM to [yourslackID] with a compact summary. Format:

```
At-a-Glance sheet updated — [date]

[N] changes pushed:
• [Track]: [one-line summary of what changed]
• [Track]: [one-line summary]

Heads up: [1-2 sentences on anything notable that doesn’t map to a sheet row — new risks, scope changes, staffing, cross-team signals]
```

If nothing changed across all tracks, send: “At-a-Glance sheet checked — [date]. No changes detected across any tracks.”

## GUIDELINES
- Be factual. No inferences or speculation.
- Use the sheet’s own terminology: “◉ In Dev”, “○ Planned”, “Complete”, “On Track”, “At Risk”, “In design”
- If a Slack channel has no relevant messages, skip it silently.
- Keep the DM under 500 characters. Christine wants glanceable, not comprehensive.

For Execution work: use Claude as copilot to accelerate execution.

Every Monday, I run the context-dump skill (click to install) which reads my past week across Slack, calendar, notes, and repo activity, gives me an analysis of my efficiency and patterns to watch, recommends what I should prioritize, and offers specific work to help me get started on.

It reads less like a summary and more like a coach who notices my personal patterns and actually makes helpful observations and recommendations. Here’s one from last week:

“Leverage dipped below the ~70% target this week, not because overhead crept up (it’s at the line), but because alignment/ceremony load (~35–40% of calendar) crowded out heads-down leverage work. The risk is that too much of it happened in meetings rather than in written, shareable artifacts.”

I also run a planning skill that checks Jira status and the roadmap and shows me how the team is tracking, what to line up next. I’d then ask Claude to make those Jira updates or turn PRDs into tickets. I know this is awesomely automated in Linear but for those still using Jira - there’s no reason to manually create tickets now.

For staying close to the customer, I run a scheduled task monthly that analyses our customer support channel to give me the top themes, the trends, and recommends the top requests or problems I can prioritize.

Get more ambitious with the asks. Don’t treat Claude as just for distilling a summary, ask for judgment.

For Impact work: use Claude as sparring partner to get the hardest work started.

This level is the high leverage work that is the hardest to start - eg. thinking through a new product bet, preparing narrative for a high stakes presentation, testing my strategy against business unit strategies, etc.

I no longer wait for the perfect condition (like a full 2 hour focus block) to start. Even with a 10 minute break, I can quickly voice dump in a Claude project (that’s loaded with my team’s and product’s context and a writing style guide) and spar on Fable high or max to get a shitty first draft. This used to take days.

It’s worthwhile to “onboard” Claude and set it up with the right context. I started with existing product decks, roadmap, our code base summary, then voice dump our goals & metrics, my stakeholders (their roles & what they care about), what I want to learn & grow in, etc. Then I’d work with this context folder in either Cowork project or in Claude Code for my tasks so it always gets the right context and also automatically updates that context as I work.

Judgment is mostly a function of context. A PM’s judgment is pattern-matching over the context they have. As a platform PM executing on 5 to 10 work streams at any given time each with a different team with different timelines and scope, I’m supposed to be the context router. Claude can hold way more context at once, so can certainly offer valuable judgement. (Obviously, it’d miss the context that’s not written down, like team dynamics and incentives, and it doesn’t have accountability for its ideas.)

Using HTML was another breakthrough in helping me understand Claude’s output better and faster than a wall of markdown text. It also gives me ideas on how to visualize my POVs during brainstorming. Simply ask Claude to “generate the output in a HTML artifact”. However, I never share html artifacts. It’s not collaborative, it’s another url for people to keep track of, and subconsciously it feels less personal as it’s obviously “AI-generated”. But structuring an idea visually takes a huge mental load off already. It gives me a starting point to then iterate in Claude Code via Figma connector to build the visual for the doc I actually share.

Fable is noticeably better than previous models for impact work. I especially enjoy its calmness, it sounds more concise. It feels more like a conversation with a smart colleague than reading walls of text with unnecessary dispositions. This makes a big difference in keeping a train of thought going.

Personal growth

Growing technical fluency

As someone who doesn’t have a CS degree, Claude Code is incredible in helping me understand technical concepts. I’d ask questions about a code base to understand its structure, generate an overview.md with architecture and feature list and reference it to debate on PRDs and roadmap. The immediate feedback on feasibility and technical approach helps me make decisions and plan way faster. I’ve pushed small PRs for the first time and had a taste for the real developer’s end to end workflow. I’ve learned a ton from just reading CC’s plans, seeing how it thinks through a problem. It gives me a great starting point to have a meaningful discussion with the engineers.

Seeing patterns & blindspots

The bonus benefit of the weekly context-dump is using the cumulative history to get honest self assessment for personal growth. It has the perfect rolling history of what I worked on, the challenges I encountered, accomplishments I made, and even interpersonal dynamics built up from meeting notes. Every few months, I’d ask Claude to review my rolling context-dump file and give me an assessment on new skills I’ve developed, blindspots I should pay attention to, and recommend me specific and actionable growth areas to work on to get to the next level (you’d have defined what your next level is in your goals.md in this project).

It’s like having a personal career coach to guide you. It’s also handy for writing your monthly and year-end reviews.

The setup (the least important part)

Beware of the second trap coming out of this - automation as a dopamine backlog. Don’t let the “perfect setup” become another distraction from just getting started. With that disclaimer, here’s my basic setup.

• Obsidian: a folder of daily markdown notes where I keep daily notes including meeting summaries (from your AI meeting summary of choice)

• Claude Code in VSCode: a workspace with my team’s repos,

  • Q&A to understand the codebases and spar on technical ideas;

  • skills to understand our latest releases and write prds;

  • Figma connector to make visualizations for docs;

  • scheduled tasks for automations through internal MCP connectors

• Claude desktop: Cowork for scheduled tasks; chat projects for brainstorming and writing, personal writing guide md file in the project folder

• Connectors: Slack, Google Drive, Outlook, Jira, GitHub, HeyMarvin

• PM folders: store evergreen markdowns that are important context, eg. context folder (with meeting notes), writing style guide, company annual goal, templates/frameworks etc.

Getting over yourself

Given the progress of models over the past few months and how impressive Fable is, it’s inevitable that Claude is going to be increasingly capable of harder tasks and do them better than us. I’ve come to admit that Claude with enough context can already outperform my thinking in many ways. The best you can do now is to keep pushing the model with higher expectations, ask it for unknown unknowns, and get the most out of it for all 3 levels of product work. Test it yourself: pick the highest leverage task you’ve been putting off, give Claude the context it would need, your current thinking, and ask what you’re missing.

Share

Thanks for reading! If you liked this, hit the 🖤 above, it helps others see it :) Find me on X.

- Christine

But wait, what would “product work” actually look like then? What is the kernel of true product work? How much have you practiced it? Let’s trace this possible evolution step by step.

The inversion of the PDLC

Traditionally, the PDLC is a funnel, wide at the top and narrow at the bottom because there are more ideas than resource to build. That’s why a large part of PM’s work is a “rationing” exercise: discovery, planning, prioritization, etc.

When resource to build is no longer scarce, the PDLC shape gets inverted. The triangle is narrow at the top. The right ideas have to be narrowed in early, because many prototypes and experiments can run in parallel. Here’s what the PDLC could look like for the product triad (PMs, designers, and engineers, though their points of collaboration will be different).

• You have an always-on Context OS and Discovery OS. Context OS intakes signals throughout your day (meeting notes, docs, slack, code base) and makes connections across your work. Discovery OS automates harvesting user insights. It crawls your feedback river (support tickets, social media, user feedback channels), generates and structures insights. Together, they surface the right signals and suggest ideas to invest in. This gives your product sense and creativity a faster starting point.

• You have a folder of dynamic PlanRDs (instead of a PRD) which hosts your collaborative thinking with Claude in plan mode. Here, you also define the eval for the idea. CC has your product’s code base, and can help you explore technical approach and feasibility with user flows and architecture diagrams without relying on asking engineering each time.

• You prototype with the PlanRD. Your AI design tool and CC are encoded with your design system and components so it builds a high quality, functional prototype in one shot, with a few iterations to get it to exactly match your vision.

• You run user testing (eval) by simulation. Instead of booking users, you have agents that simulate the target personas. They interact with the prototype and give feedback. You run hundreds of simulated sessions to get meaningful signal on your prototype. (This is additive to real user testing).

• Your top bets move to development for production and scale. The PlanRD is updated with technical specs and designs, connected to your development tool which plans tickets and assigns them to coding agents. Code review and testing agents follow the plan to complete the build.

• After final reviews, the idea is deployed to production. Many ideas are simultaneously deployed, because you can learn faster.

Subscribe now

Implications of an inverted PDLC

The messy middle. The inverted PDLC has a “barbell problem.” AI is great at the two ends: input of incoming work (customer feedback, requests, etc) and execution (coding agents building to specs). But this is a recipe for slop. That messy middle of correctly deciding what’s the actual problem and what’s worth building are not well solved by AI, because AI optimizes for the mean and every PM has access to the model’s power. Human judgement in the messy middle is what creates the edge.

Human review as the bottleneck. When the messy middle is not properly managed, it worsens the downstream human review bottleneck. AI generates more docs, more ideas, and more code than humans can review. PMs now have way more signals (automated by AI) to ingest and synthesize. PMs generating prototypes and creating PRs with good intentions also transfer hidden cost to engineers, because they might not be aware of the downstream implications like maintenance and tests.

The new hard skills

In this world, PMs need to develop new hard skills to manage the messy middle.

1. Portfolio Management. With multiple experiments running simultaneously, PMs are managing a portfolio of ideas. The work becomes more of an editing motion, thinking in bets akin to a portfolio manager: monitor signals, size bets, allocate bets, decide what to double down on and what to kill.

2. Product Eval Sense. When products are probabilistic, the definition of “what good looks like” is a qualitative judgement that’s constantly shifting. PMs need clarity on what good looks like to define eval from the start and continuously update that definition. Eval formalizes taste. The skill is turning the sense of “this is technically correct but feels wrong” into something systematic that agents can test against. It’s also about actively monitoring for model drift and recognizing when previous constraints are lifted and something new becomes possible mid-development.

3. Simulation Design. Designing a useful simulation for user testing requires deep understanding of the personas for your product and user research methodologies applied to synthetic users. The skill is building specifications for getting meaningful signals that can be reliably tested by simulation, such as task completion, failure modes, etc. It doesn’t replace the emotional signals and non verbal response that would always require real human user testing.

The path to accumulating these skills is like layers of a Russian doll. You ask Claude to do thing 1, when that’s successful, ask Claude to do the natural subsequent thing that you were going to do, and so on and so on. Soon, you put yourself out of as many layers of your job as you can, until you get to the kernel.

The Kernel Skills

With this inverted PDLC, every idea is possible, for everyone. The race comes down to having the best idea.

What leads to the best idea?

We generally underestimate how well AI can perform on having good ideas and even good “taste”. (Taste is a fragile fortress.) To a certain extent, AI can synthesize insights, pattern match, and decide what to build under ambiguity, all with fewer cognitive biases than most PMs. The spaces that will remain hard to automate are a few elusive human qualities, “the kernel”, that can’t be captured or encoded.

1. The upstream creative act of asking the right questions.

It’s framing the unique problem you’re solving for from a lens that most people are not even thinking about. Slack didn’t ask “how do we make email faster”. Airbnb didn’t ask “how do we make hotels cheaper”. People with this innate sense are category creators. The “what” becomes obvious to them before the “why”. Naturally, they are also the greatest storytellers. They’re just working backwards to explain the thing they already know rather than to justify or persuade.

2. The antenna for novel signals through direct human contact.

At the macro level, it’s having a sharp cultural antenna, sensing cultural currents through immersion. Snapchat understood that the generation of teenagers wanted an ephemeral space that’s not archived. At the micro level, it’s noticing a hesitation in the user’s answer, what wasn’t said in observing users. These signals don’t exist in data, are out of reach for AI. This empathy still requires being physically in the room, living in and sharing the world with fellow humans.

3. The conviction across small decisions over time.

It’s the consistent, accumulated directional bias of hundreds of small decisions that’s driven by a strong conviction for a specific point of view. It certainly cannot be built by committee or consensus. AI mostly gives the reasonable answer, and yet great products are often unreasonable/non-obvious by regular standards. Linear’s every product decision biases towards “fast and opinionated”. The keyboard-first design, intentional simplification of customization and options, all accumulate to reflect a product that has a clear sense of what an agent-native tool should feel like. Conviction carries the cost of saying no, and AI doesn’t have skin in the game.

The truth is, most PMs spend most of their time on the operational, analytical parts of judgement which AI will replace. The kernel work is maybe 10-15% of a great PM’s time and 0% of an average PM’s time.

The knee-jerk reaction is to chase the latest AI tools, install the latest skills, set up the perfect agent workflow. These are the minimum basics, the easiest part that makes you feel like you’re “not left behind”. They’re certainly not differentiation. The next level of new hard skills to develop are: portfolio management, product eval sense, and simulation design.

But the true differentiation for PM work in 2027 is the same kernel skills needed for building great products pre-AI: creative problem understanding, a tuned antenna for signals, and conviction in a specific point of view. In fact, these are the essence of the original PM role that’s been under-utilized, now finally unshackled from the layers of operational overhead.

The question to ask is, have you been practicing the kernel skills, or consumed in the operational ones?

Share

Thanks for reading! If you liked this, hit the 🖤 above, it helps others see it :) Find me on X.

- Christine

In the AI era, intelligence is becoming free (inference cost dropped 1,000x in 3 years). The new scarce resource is context, which is currently wildly fragmented for the knowledge worker. The average company uses over 130 SaaS applications. Knowledge workers juggle between 8-12 different tools daily, manually copy/pasting context between their tools to make their AI useful.

The Context Aggregation theory: when intelligence becomes free, context is the scare resource that makes agents effective. Those who aggregate the richest context will commoditize both the models and SaaS/system of records and will control the value.

There’s a growing wave of companies investing in context aggregation. Recently, Notion acquired Sequin, a startup that syncs data across a company’s tech stack into database for developers, and Embra, which built a graph context engine for B2B. Glean repositioned from enterprise search to AI middleware between agents and corporate data. Salesforce acquired Informatica for $8B to invest in context infrastructure.

What remains defensible

SaaS tools that are primarily data stores will get commoditized first. That’s CRMs, project trackers, ticketing systems - products whose primary value is storing and retrieving structured data is vulnerable. If an AI agent can pull your pipeline data from Salesforce and your task data from Jira and reason across both, the UI layer of the individual tool matters less. The agent becomes the interface to these SaaS tools’ data endpoint (commoditized supply).

SaaS tools with deep workflow logic will remain defensible longer. For example, Figma’s value is more than just “stores design files”. It’s the spatial canvas, collaborative editing, component systems, workflows that are more than just a data endpoint. Same for tools like Linear which has an opinionated workflow, or Shopify which delivers rich commerce execution. The more a tool’s value is in the doing vs. the storing, the harder it is to get aggregated away. These suppliers with differentiated, hard-to-replicate value remain defensible.

The race is going to be between three types of aggregators.

1. Workspace aggregators (eg. Notion): They already own the user’s daily workflow. Adding context connectors (Sequin acquisition, AI Connectors) strengthen their position to be the place where work happens and where context is assembled.

2. Infrastructure aggregators (eg. Glean, Prefect Horizon): They don’t own the user experience directly but are powering the context layer beneath it.

3. Platform aggregators (Eg. Salesforce, OpenAI): They have enterprise distribution and existing system connections. Microsoft Graph already connects 1,400+ systems. Salesforce’s Informatica acquisition unifies data governance.

Workspace aggregators have the best shot because they have the most direct user relationship. But the platform aggregators have distribution advantages that can’t be dismissed.

Subscribe now

The endgame is judgment

Aggregating structured data from SaaS tools is becoming commoditized via MCP. Capturing the judgment layer is the most defensible play, which is why decisions were made, the exceptions and precedents. This is the context that accumulates over time and can’t be easily replicated. The context aggregator that captures decision traces first will build an insurmountable advantage.

Like Thompson’s original aggregators, context aggregation has a compounding dynamic. The aggregation flywheel: the more tools you connect —> the richer the cross-tool context becomes —> which makes agent performance better —> which builds user trust —> which leads users to connect more tools and share more data. Over time, the context graph becomes cumulative and proprietary to each organization, creating a compounding advantage that gets harder and harder to replicate.

What it means for builders

In the internet era, aggregators have winner-take-all dynamics because the demand-side economies of scale are so strong. Context aggregation probably leans more toward winner-take-most per category, because enterprise context is trust-gated and domain-specific. It’s less likely for companies to hand all their context to one aggregator. But there’s flywheel dynamics in each category of aggregator (workspace, infrastructure, platform) that favors consolidation. The race will run in parallel.

Products that are primarily a data store are in a losing position and are at risk of getting aggregated away. One escape is to move up the stack, evolve into a deeper workflow tool or become context aggregator (eg. Salesforce acquiring Informatica). If your product’s value is in workflow logic, then deepen the interaction model that keeps decisions made inside your tool, which generates the implicit context (decision traces) that agents need but can’t create.

The models keep getting cheaper, context won’t. Context aggregators will be accruing value in AI for the next decade.

Share

Thanks for reading! If you liked this, hit the 🖤 above, it helps others see it :) Find me on X.

- Christine

Today, regular consumers are already familiar with using the top AI chats for Q&A tasks. But fragmented context remains a major unsolved hurdle. Users have to gather context from many different apps and sources to feed the model to get the most useful outcomes.

This puts the differentiation on the input - consolidating fragmented context. Products that can make it obvious and frictionless to connect, surface, and control context - will win.

Subscribe now

The Three Layers of Context

1. Internal context: Rich data and business logic that already exists in your product. This is the system of record that captures what happens.

2. External context: Data in other apps in a user’s “tech stack” that works with your product in users’ regular end-to-end workflows. This is root of the context switching pain for knowledge workers.

3. Implicit context (judgment): True context comes from the relationship between facts - why a decision was made. This “judgment” context is highly human, mostly implicit, and the hardest to capture. It’s often scattered across different mediums and often happens in people’s heads.

In the past decade, the tech industry proclaimed moving from “system of record” to “system of intelligence”. But this centers on individual apps. Knowledge workers are still acting like context engineers, manually consolidating context from different systems and making decisions in their head.

Now, piecing together context (especially judgment) is the missing piece and biggest opportunity for AI to be truly useful beyond small discrete tasks.

Three levels to building a context mesh

These levels build on each other. Without granular internal data (Level 1), AI can’t differentiate between relevant and irrelevant context. Without 3rd party connections (Level 2), context is fragmented across tools, losing workflow continuity. Level 3 only becomes possible after instrumenting the lower layers.

Share

Level 1. Surface Internal Context

For incumbent apps, the first step is making existing data granular and selectively accessible, both for AI retrieval and for user control. This means:

1. Chunking content into meaningful units: functions, blocks, database rows, not arbitrary boundaries

2. Enriching data with semantic metadata: entity relationships, permissions, temporal context

3. Building APIs that expose context selectively: so that access and permissions are respected

This is tablestakes. Agent browsers like ChatGPT Atlas, Dia, Claude extension, are limited to surface-level data, what’s rendered on screen. They can’t access underlying data models, entity relationships, permissions-based context, or domain-specific business logic that lives in your app. Owning that “system of context” is one of the last moats for incumbent apps, but only if they invest in packaging it at the right level of granularity and building interfaces that make context discoverable, selectable, and transparent.

The dependency is hard to retrofit: frontend context control UX requires backend granularity. If data is stored as blobs (one big document per entity), it’s hard to offer fine-grained selection. Users don’t get targeted selections. The AI either sees too much (overwhelming, privacy risk) or too little (useless).

Notion‘s unique advantage is becoming apparent with its block architecture. Everything is a block - pages, tables, database rows… This granular backend structure makes the “add context” frontend UX possible, which lets users point to specific context (blocks). Notion AI can read page comments, version history, and track who changed what and when.

Cursor‘s codebase embedding shows the same pattern for code. Cursor indexes the entire repository by chunking code semantically (by function, not arbitrary lines), and storing them for fast similarity search. Developers can explicitly choose which files to include in context. The codebase embedding model gives the AI deep understanding of the entire repo, but users control what it can see when making suggestions. You can ask vague questions like “where is the auth logic handling token refresh?” and get accurate answers because Cursor indexed the project structure.

The moat: Incumbent products already have structured data for their domain. For example, Salesforce knows how Accounts relate to Opportunities. Notion knows how pages nest and reference each other. Cursor knows how functions call each other. Replicating years of domain-structured data is the hard part. Incumbents have it by default, but they have to expose it intelligently.

Level 2. Connect External Context

Knowledge workers regularly navigate between 8-12 different apps in their typical workflows, manually transferring information between systems.

Level 2 is about making your app interoperable with other key apps in your customer's tech stack, eg. establish seamless 3P connections with app partners. Users should be able to access context across app boundaries and continue a stream of decision making and complete their jobs to be done without context switching.

This extends context control from single-app boundaries to workflow boundaries - the actual unit of knowledge work.

The connector UX pattern that’s emerging: users authenticate with external services (Slack, Google Drive, Jira), AI can then surface relevant context from those sources.

Notion AI Connectors is a great example for how cross-app context works in practice. When discussing Q4 roadmap in Notion, AI surfaces relevant Slack threads, related Jira epics, and supporting Google Doc specs. It respects user permissions and learns relationships between external data and workspace content.

This pattern is becoming standardized. Model Context Protocol (MCP) is emerging as the common integration layer for how AI app connect to external data sources. Claude and ChatGPT use MCP to enable portable integrations. The same connector can work across different AI platforms rather than being rebuilt for each product. This protocol-level standardization is accelerating the pace of ecosystem integration.

To build external context connections:

1. Authenticate key services in user workflow: Identify 3P apps that are critical to the users’ jobs-to-be-done

2. Map relationships between data sources: Connect external entities to internal context (eg. Slack threads → Notion pages, Jira epics → product specs)

3. Surface relevant context with permissions: Respect user access boundaries while making cross-app context discoverable

Level 3. Capture Implicit Context (The Judgment Layer)

Level 1 is about making existing data accessible.
Level 2 is about making existing data connected.
Level 3 is about capturing the new decision traces that forms the judgment layer of knowledge work.

For example, “we already tried approach X and it failed because of Y” type of judgement lives in someone’s head, not in the current system of record.

Systems capture what happened (data in CRM, project trackers, docs). They don’t capture why it happened - the exception logic, precedent, and reasoning behind decisions.

Judgement is utterly human and the highest-leverage layer for AI and agents don’t have access to it. We are not even capturing it.

The emerging approach was coined “context graphs”, Foundation Capital calls “AI’s trillion-dollar opportunity”. Unlike traditional knowledge graphs that link entities (eg. person → company → product), context graphs capture decision traces across time.

Capturing decision traces requires being in the execution path at commit time:

• Instrument the event layer to capture decision traces as they happen

• Build temporal context storage: what was true when the decision was made

• Create precedent chains: link related decisions over time

• Structure exception logic as first-class data, not tribal knowledge

This is still an “elegant theory”, and Dharmesh (HubSpot CTO) was pragmatic in questioning the timing and practicality of this concept, as companies are still figuring out how to use AI for basic tasks. They’re not even thinking about this context layer yet, nevermind instrumenting their agent orchestration layer with decision traces.

Implementing the judgment layer in practice also raises new questions: what decisions get traced, who owns this context, and how do users maintain agency over their decision history. At minimum, it’d be reasonable to maintain a principle of making the “capturing” frictionless and transparent.

Win on Context

The AI race is not only won by better models, new UX revolution, but also by better data architecture. Building the three levels of a context mesh will shift from “chat with your data” to “chat with your organization’s judgment”.

There will be predictable blockers to anticipate.

• Level 1 retrofitting: when data is stored as blobs, re-architecting for granularity is costly. It’ll be hard to build targeted context selection UX when backend can’t serve it in the right unit

• Level 2 API dependencies: external connectors introduce rate limits, auth complexity, and third-party reliability risks

• Level 3 organizational resistance: capturing judgment requires changing how teams document decisions. There’s cultural considerations beyond technical ones.

The playbook isn’t to do all three at once. It’s more “start with Level 1, instrument what you control, then expand.” Start with what you control:

• Clean up internal context. Make your data granular. Surface what AI can see to the user. Make it visible and selectable with minimal friction.

• Expose the external context. Make it interoperable with apps in your customer’s ecosystem.

Level 3 will take time to unfold and mature. But the companies building toward it today will own the judgment layer before startups can synthesize it externally.

Context is the new moat.

(The meta layer to this is figuring out the context consolidation to unlock new productivity in how we work in our organizations. I’ve been experimenting building my own context mesh and a PM agent in Cursor. Maybe for another essay.)

More reads

• Context Graphs: Capturing The “Why” In The Age of AI - Dharmesh Shah

• Context Graphs and Process Knowledge - Jessica Talisman

• Steam, steal, and infinite minds - Ivan Zhao

Thanks for reading! If you liked this, hit the 🖤 above, it helps others see it :) Find me on X.

- Christine

Why? Because it’s the path of least resistance. Chat is additive to a core product experience without redesigning or rearchitecting anything. Natural language is universal, so there’s zero learning curve. And open-ended conversation helps narrow the problem space when the job-to-be-done is ambiguous.

Chat is legitimately good for certain jobs. But it’s not good for every job, especially where direct manipulation would be faster. It remains dominant because it’s sufficient and easy to ship.

Chat is the minimum viable AI interface.

Incumbent SaaS products face a different challenge than AI labs whose chat is the product. When users already have established workflows, bolting on a chat doesn’t immediately unlock new value.

For chat to be an effective medium to serve LLM capabilities, it needs to marry conversational input to the existing GUI users are already familiar with. While we’re still experimenting and imagining the next UI revolution with LLMs, here are three primitives that make a great AI chat.

1. Context-rich, outcome-oriented input

A great chat breaks the blank page problem with contextual entry points.

Traditional SaaS asks users to specify the steps to accomplish a task. This forces users to revolve around the app’s logic, not their job to be done. LLM flips this: start with user’s intent, deliver the outcome directly, abstract away the steps.

Yet the freedom of natural language can also be paralyzing to start. The blank text box is a terrible interface for discovering what’s possible.

Outcome-oriented prompts

Solve the cold start problem by shifting users from command-based to outcome-based thinking. Instead of forcing them to conjure the perfect prompt from nothing, surface dynamic prompt starters for common jobs-to-be-done within their current workflow.

Notion has slash commands that progressively reveal AI capabilities. Type / to see “Continue writing,” “Summarize,” etc. These outcome-focused templates teach users what’s possible in the context of the workflow.

For power users, enable personalized prompts so they don’t have to repeat themselves. This concept predates LLMs. Physicians have created their own “macros” in EMRs for decades, storing structured templates for medical notes easily retrieved by the “.” keyboard shortcut.

Pete Koomen also argues for the imperative of user-defined “system prompts” in his essay AI horseless carriages.

Dia browser’s skill library was born out of its hit feature “skills”: user created prompt templates that can be easily injected via keyboard shortcut.

Context targeting

Beyond prompts, optimize for contextual entry points that automatically include relevant user context. In addition, give users control to specify context to point the agent in the right direction.

Cursor’s @ mentions exemplify what Linus Lee calls “mise en context.” Like chefs prepping ingredients before cooking, the interface prepares relevant context to support its workflows. Type @file and search for a specific file to include its contents. Type @docs to search official documentation for libraries and frameworks.

Semantic input becomes a discoverable interface instead of a blank page. The @ symbol is simultaneously a command palette, a search bar, and a context selector. This allows users to learn what’s possible by exploring the menu rather than guessing what to type.

Notion does this with “add context” in the input bar, paired with the @ mention keyboard shortcut, where users can append any Notion page as context to their prompt.

Dia Browser has a great universal input bar that serves both prompt input and search, which lets users point to specific context such as Slack threads.

The pattern: structured affordances layered on top of freeform text.

2. Malleable output

A great chat renders output dynamically, molding to individual situations.

The natural workflow of information activities is chaotic and personal. With server-driven UI, instead of developers hardcoding each workflow, LLMs can generate the UI specification itself. The system semantically describes the logic, and AI determines the optimal rendering for users’ specific context.

This shifts control from developers deciding the average experience for all users to systems adapting to individual needs.

The shift to generative interfaces

Google Research calls this “Generative UI”: AI generates not just content but the entire user experience. For example, if user asks for an analysis of their business, the system dynamically renders data visualizations best suited to that analysis, without the user or the developer pre-specifying it.

Vercel V0 was one of the early vibe coding tools built on generative UI, where user can describe what they want to build in natural language, and V0 generates complete React components styled with Tailwind. The UI itself is the output. This is a shift from designing for many to tailoring for the individual.

This paper on Generative and Malleable User Interfaces and the accompanying Stanford seminar explores practical examples of this technical approach.

Starting points, not endpoints

When users engage with creative tools, they want to directly shape the process of iterating to their end goal. The output is a starting point for iteration, not the final product.

In canvas interfaces such Claude Artifact, ChatGPT Canvas, a conversation generates a starting point, and the canvas then provides a workspace for further refinement and exploration.

Figma’s First Draft combines both principles. User can start with an open ended prompt to generate an editable wireframe, then continue to iterate with direct manipulation controls - eg. adjust color themes, spacing, typography, etc. The output is built from real Figma components they can then rearrange and edit.

The pattern: outputs that are both dynamically generated and directly malleable.

3. Wired to action

A great chat reaches into the product and execute actions, not just give answers.

Linus Lee wonderfully articulated a useful dichotomy in instrumental vs. engaged interfaces. Instrumental interfaces are transactional, user gets in and out to get a specific job done. Think EMRs, financial tools, CRMs. Engaged interfaces are creative tools where there’s intrinsic value in interacting with the tool despite the outcome. Think design software, music production.

AI chat done right oscillates fluidly between instrumental and engaged modes. For instrumental tasks, chat provides a magic button that shortcuts intent directly to product action. For engaged interactions, chat provides a malleable starting point for iteration.

The more obvious near-term economic value comes from instrumental interfaces - tedious parts of knowledge work like filling forms and updating records being replaced. This requires the chat to be architecturally wired into the main product UI, with bidirectional communication, so product actions can be directly executed with full visibility and control to the user.

Beyond agentic browsers

Agentic browsers are attempting this from the outside (eg. Dia, Claude’s browser extension, ChatGPT’s Atlas). These have to take screenshots, infer context, then guess at how to execute actions, which can be inefficient and get expensive quickly. Products with direct access to their own data and APIs can wire this connection far more efficiently. The underlying system knows the schema, and can better validate actions before executing.

The test

To evaluate whether a chat interface is truly integrated or merely bolted on, ask these questions:

1. Input: Does it ease the blank page friction with proactive, contextual prompts and targetable context?

2. Output: Does it render dynamically and let users shape results?

3. Action: Does it execute actions directly in the product, or just give answers?

To move beyond the minimum viable AI interface, your product has to answer yes to all three, or you’d be just bolting on a chat.

The real interface shift

For decades, product teams have designed rigid “best practice UI patterns” (why vibe coded UIs all look similar). Users had to learn to navigate app logic to accomplish their tasks.

These primitives point toward an interface that anticipates and reshapes itself around user intent rather than forcing users to reshape their intent around the interface. It’s moving from “add a text box to talk to an LLM” to “rethink how the entire product surface responds to what users actually want to do.”

There’s often a delay for UX to catch up to the technology leap. The minimum viable AI interface got us started. While we’re still discovering the next interface breakthrough, these primitives can help us move beyond simply bolting on a chat.

If you liked this, hit the 🖤 above, it helps others see it. Find me on Twitter.

- Christine

There’s a “crossing the chasm” challenge to getting meaningful adoption of AI products. Early adopters are willing and eager to try new behaviors and put in the effort to imagine possible use cases, but the mainstream users will turn away from unfamiliar experiences or general (albeit magical) capabilities that don’t solve a specific problem.

As foundation models get more and more advanced, the UX that delivers the technology can make or break adoption (and retention).

Subscribe now

Apple Intelligence vs. Microsoft CoPilot

Apple did it right with the AI experience - embedding AI in existing workflows through a natural GUI experience. The Apple Intelligence part of WWDC led with use cases where the AI experience enhances the existing workflows, whether it’s summarizing your notes, helping you write emails, showing priority notifications, and ended with the new Siri.

Whereas Microsoft, given its close partnership with OpenAI, led with the Copilot experience, the general purpose (advanced Clippy?) assistant that awaits your prompts in a chat interface.

Perhaps the popularity of ChatGPT made the chat interface a default choice for companies incorporating AI in their products. But natural language isn’t always the most natural way we interact with our computers. Most of us are not great at articulating our intent in natural language. There’s a blank page barrier of coming up with what to ask.

The beauty of GUI is the affordance to give you suggestions on possible actions you can do in the interface. That intuitiveness and discoverability is lost with a chat interface. (Check out this great talk about generative interfaces beyond chat.)

Building AI in healthtech

Regulated spaces like healthcare and law are often laggards in technology and are sadly known for some deeply frustrating user experiences. UX plays an especially important role in balancing people’s desire for productivity with the fear of loss of control to technology.

Here are a few observations from talking to clinicians and designing AI into their day to day workflow.

1. Familiar > Novel. People are hesitant to try things that are unfamiliar.

A familiar UX, as opposed to a disruptive one, will have a far better chance at getting user adoption. For healthcare professionals, who are mostly risk averse, the deeply engrained routines that allow them to survive their busy days hardly leave room for new explorations.

Even if the current workflow in their tools suck and take too many clicks, that’s where their muscle memory takes them. A brand new UX behavior would simply go unnoticed, or scare them off. A better shot at successful adoption of AI capabilities is to leverage familiar UX. For example, a search bar feels more familiar than a chat bar, even though the inherent capability is about retrieving data. Everyone is familiar with Google search, and feel comfortable typing 2 words or a question, whereas it’s much harder to start with crafting a perfect questions in a chat bar.

2. Output > Input. Output activities can get better adoption than input activities.

Related to familiarity, users in regulated, risk averse industries like healthcare and law have lower willingness to give AI control of work which they sign for, such as written documents. There will likely be more resistance towards applying AI in these input type activities.

With time, model improvement (better accuracy) will increase people’s tolerance to AI taking on bigger portions of their workflow. But starting with use cases around output, such as summarization and search, can get a foot in the door with getting users comfortable with AI.

3. Focused > general. UX tailored for specific use cases are easier to adopt than general ones.

ChatGPT is a step change innovation for general use cases, but it doesn’t yet serve specific needs. It’s akin to PCs before applications were built on top of them. Mainstream market developed when applications like Excel was invented, which was game changing for accountants.

A chat interface in the healthcare setting is too general, which makes it hard for users to imagine where to start. Embedding AI within the context of a specific workflow makes it much more intuitive to get started.

Within the Microsoft example, it’s much easier for someone to try the Outlook Copilot that’s embedded within the email composer vs the general Copilot button in Teams that leads to a general prompting experience.

Crossing the AI chasm

The D-day analogy in “Crossing the chasm” is most definitely applicable and underrated in AI. It would be most prudent to identify a highly focused use case that can be well solved by AI, and tailor a familiar GUI design to encourage the best possible adoption.

Once the initial targeted users get comfortable with the initial AI use case, they are more likely to champion your product/feature with others via word of mouth. Building on this initial success, it becomes more feasible to expand to more AI use cases with more user segments - eventually crossing the chasm.

What response have you observed from your customers about integrating AI in your products? What UX patterns have you had success with?

🔗 More reads

• Looking for AI use cases

• Coda launches Coda Brain

• Apple Intelligence is right on time

• Winning interface to AI

• What ’80s movies got right and wrong about living with AI

Thanks for reading! If you liked this edition, hit the 🖤 above, it helps others see it :) Find me on Twitter.

- Christine

Unsolicited feedback’s live event about advancing your product career in an AI world explored a thought experiment about what AI’s good at and not good at today. Eg. it’s good at execution tasks, and not so good at deeper thinking and planning. Because of this, Brian and Fareed predict that mid-level management will shift to senior ICs (principal, staff). Some recommendations? Get great with data manipulation (eg. SQL), and tinker with a side project with LLMs, it’s the best way to learn.

Mini product teardown: Lapse, the disposable camera app that’s trying to become the anti-Instagram. It apparently ranked 3rd on the App store at in November. My quick review after trying it for a few days? They nailed the film look but wow it’s got a passive aggressive personality! It forces you to invite 5 friends before allowing access to any feature. It uses sad emojis and snarks to guilt-trip you into adding friends and publishing snaps. How do you think about the balance between aggressive push for growth vs. generally pleasant user experience? How likely would this friend only / authenticity-driven model survive over the following/influencer model with traditional social media giants? I’m doubtful. Kevin Systrom’s Artifact shut down after a year. It’s hard to make a new king in social.

Subscribe now

💻Product

• Winning the AI Products Arms Race. The AI survival curve uses “consideration” vs. “context” axes to evaluate AI opportunities. Consideration: the amount of effort needed to make a decision. Context: the volume of abstract concepts AI needs to know. Successful AI products all fall on this curve.

• Finding the right opportunities in AI. Christoph shares his mental map on how to think about opportunities in AI. Vertical products is more opportunistic than horizontal, because AI-first startups can leapfrog from pen& paper to AI, and there are more opportunities to sell work products instead of software. He’s also more interested in very hard to solve problems, and in new categories that are enabled by AI rather than applying AI to existing areas.

• Arc Browser’s act 2. Arc describes a vision for a new internet that blends together search engine + web pages + browser. It browses for you, ie. goes through all the websites and presents a new page with exactly what you need, no ads, no filler content that’s optimized for SEO. It’s a lot of grandiose exclaimations. What would that future mean for webpages? Do they just become "headless" information?

• On platform shifts and AI. “What separates a major platform shift from a minor platform shift is a platform shift that enables both a technological shift (new ways of making things possible) paired with a distribution shift (new ways of reaching people with it).”

💼GTM

• What it really takes to be successful in enterprise SaaS sales. Christian, CRO of Splunk, emphasizes using the quality of pipeline as the key metric for assessing the overall health of the business - ie. characteristics of the pipeline, conversion rates. Splunk’s favourite demand gen strategy is product generated pipeline, basically PLG. Leverage existing usage to generate expansion opportunities. Segmentation is key to structuring sales - done by ARR and ARR potential.

• The only playbook for competitor research and strategy you’ll ever need. A wildly comprehensive resource by Patrick Campbell who founded ProfitWell, sold to Paddle. Most companies don’t do competitive strategy properly. Don’t be seduced by the idea that “competitor are not important”, there are very few one in a million unicorn products out there. Watch your competitors and defend. This guide gives highly specific tactics and frameworks on competitive intelligence.

• How revenue leaders at Box, Calendly, and Lattice scaled from $0 to $100M+. This conversation is filled with million dollar lessons from pivotal moments at these cloud 100 companies. At Lattice, pausing PLG and getting clear on ICP definition significantly grew conversion. Sales were having less calls per day but they were much higher quality accounts with higher conversion. For horizontal products like Calendly, leverage popular use cases to build vertical solutions around. Consider org changes across product, people, and process when going upmarket to enterprise. When sharing challenges outside your own domain, develop relationships with all the functional leaders first. Understand when to be a functional leader vs. a corporate executive (removing biases towards your own function and thinking as a single team towards a north star).

• The Model Market Fit threshold & what it means for your growth strategy. A classic by Brian Balfour, thinking beyond just product market fit. It’s the concept that your market and the number of customers in your market influence your model. On a graph of ARPA by # of customers, there’s an inverse line that represents the model market fit threshold. The model market fit hypothesis is based on this equation: ARPU x Total Customers In Market x % You Think You Can Capture >= $100M

🔗 AI resources

• AI and everything else - Ben Evans’ annual presentation. Analysis and predictions on where AI is at and where it’s going.

• A non-technical explanation of LLMs

• Sequoia’s Generative AI market map

Thanks for reading! If you liked this edition, hit the 🖤 above, it helps others see it :) Find me on Twitter.

- Christine

1. Use case matrix for AI products

2. UI for AI products

3. Pricing models for AI products ← this one

Scroll to the bottom for more resources and related readings.

Subscribe now

Generative AI introduces a dirty word in the traditional SaaS model: marginal cost. This makes it tempting for some to consider a dangerous pricing model: cost-based pricing. Well, to be fair, many are not seriously thinking about pricing model at all because they’re racing to have a “AI strategy”.

With the maturation of product led growth, it’s prudent to think about pricing models and monetization early in the product development lifecycle. There hasn’t been much exploration into this topic, so I researched how some of the popular AI products are adopting usage-based, seat-based, and value-based pricing and compiled some patterns below.

In this post:

1. the value pyramid of AI products

2. three pricing principles

3. breaking down usage-, seat-, and value-based pricing

Value pyramid of AI products

This is a generalization but there seems to be 3 basic categories of AI products and a likely pricing model for each.

Infrastructure

These large sophisticated models and AI infrastructures are the foundation of many SaaS products. Soon enough, their costs will become commoditized like AWS. These companies often charge usage-based as they’re basically a utilities cost.

That said, companies that use these utilities will need to layer on new differentiation to create unique value for customers.

Examples: OpenAI, Anthropic, Databricks

Augmented Experience

Layering on top of the infrastructure are horizontal solutions that carry out a workflow to support the user’s productivity, such as customer support, note workspace, spend management. These products tend to add LLMs in a chat interface or to automate parts of the workflow to further improve the user’s productivity.

They’ll use foundational models as the base but the innovation or new value created is the new productivity unlocked within the context of that workflow. They often charge per seat for access to this new value.

Examples: Notion AI, Gong, Brex

E2E (end to end) Work

Further layering on an augmented experience are products that complete an end to end workflow/task in a specific vertical use case. They’re able to codify expertise in a specific field and train AI to complete the full task. Sarah Tavel best puts it as “work products”.

They improve productivity by 95% (rather than 5 or 10% in the above augmented experiences) and essentially replace a human service. These opportunities can be found in services where there already exists outsourced groups (eg. bookkeeping, call centre, customer support, medical billing, data scrapping, etc.)

These products can leverage value-based pricing and charge per work product, in direct comparison to the human capital it saves.

Examples: Even up, Intercom, Cresta

Pricing principles

Before we explore some of the common pricing models, it’s important to keep in mind three criteria that guide a successful pricing strategy.

1. Simple: easy for customers to understand.

   Whether you have a product-led self-serve motion or sales-led motion, pricing that’s easy to understand is an important factor in driving conversion on the pricing page and win rate for sales. The uproar over Unity’s confusing pricing update is a great counter example of this.

2. Predictable: customers can predict and budget the cost.

   If you’re lucky, at some point in a sales conversation, the buyer will ask about pricing. And they will then immediately try to estimate the total cost against their budget. The friction to a “buy” decision is much greater if they can’t predict their total cost.

3. Scalable: aligns with customer’s growth, so you can grow together.

   It’s always hard to increase prices, so having a pricing model that can naturally grow with customers is ideal. It completely aligns pricing to the product value received by the customer, as they depend on the product to grow their own business.

Common pricing options

Pricing AI products should consider how much differentiated value is delivered by the AI feature and how the customer perceives that value.

1. Per seat pricing: value is in accessing the AI feature regardless of usage

2. Per seat pricing with usage cap: value is in generating ideas to augment a creative process

3. Value-based pricing, per work unit: value is in delivering an end work product, replacing an outsourced human service

4. Usage-based pricing: value is the underlying infrastructure that powers the product

Let’s dig into each further.

📈Usage based

Usage based pricing (token-based) is the most common for infrastructure/language models providers, as it’s based on the cost to run the models on GPUs. Obvious examples include OpenAI, Anthropic, Cohere, Databricks.

The argument against this model is that it’s hard to translate usage to value of work completed. It’s unavoidable as a model provider, and the GPU costs will go down overtime. But for companies that are using these models to deliver their own products, it’d be a bad idea to do cost-plus pricing. It’s difficult for customers to understand the discrete value provided on top of just using the models directly themselves, and it’s difficult to predict and budget cost.

As it gets cheaper to build LLMs, larger companies who have the data and resources could invest in building their own models instead of using a model provider. They’d then be able to enjoy minimal marginal cost and leverage value-based pricing.

Examples:

• OpenAI API: GPT-4 turbo input $0.01 / 1K tokens, output $0.03 / 1K tokens

• Hume (analyze changes in people’s emotions based on voice intonation and facial expressions): charges per minute, annotation and word.

• Databricks: $0.07/DBU for jobs. They have a handy pricing calculator to help customers estimate cost.

👩‍💻Seat based

The classic SaaS per seat subscription model is simple to understand and easy to budget for. It’s the most straightforward model to charge for new AI features added to products that are already priced this way.

As we’re still in an experimental phase with AI, many companies are racing to add AI in their products to optimize for growth. They might not be as concerned about whether their token costs are covered. For products that are particularly expensive to run (such as images and videos), adding usage caps/credit bundles to a per seat model is also a good idea.

Examples:

• Flat rate per seat

  • ChatGPT: $20/month per user for ChatGPT

  • Copy.ai (AI OS to automate workflows):

    • Pro plan is $36/month for 5 seats, unlimited words, 500 workflow credits

  • Runway (Generates images and videos):

    • Standard plan is $12/user/month, 625 credits/month, can buy extra credits as needed

  • *Note how Copy.ai and Runway have usage caps.

• Add-on per seat (AI features added to core product)

  • Notion AI: $8/user/month add-on to paid plan

  • Microsoft Copilot Pro (consumer): $20/user/month add-on to Microsoft 365

  • Microsoft Copilot (enterprise): $30/user/month add-on to Microsoft 365

• Free AI add-on (bonus value to drive growth)

  • FigJam AI

  • Miro AI

  • Arc Max

🎁 Value-based

When AI gets sophisticated enough to replace a full end to end human task, there comes a unique opportunity to price in relation to the cost savings of having human complete that task. (Again, ’s Sell work, not software is the best exploration of this concept).

The alternative would be outsourced human capital (BPO services). So it’s easy to calculate the cost of these outsourced tasks, which makes it easy to justify the ROI for the AI replacement.

Ian Clark’s How to price generative AI products also makes the argument for why the right way to price AI products is to charge for the value. You shouldn’t consider your cost at all when pricing your software product. It should be all about customer’s willingness to pay.

Examples:

• Intercom: AI chatbot Fin charges $0.99 for every resolved customer request. (Its core customer service product is priced per seat.)

• Even up: charges per demand package (the work product produced by lawyers in injury law).

Any new pricing model innovations not covered here? How are you thinking about pricing models for B2B generative AI products? How is that decision impacting how you build the product?

🔗 More pricing reads

• Sell work, not software

• How to price generative AI products

• How to price AI in SaaS (2023)

• AI startups trying to make usage based pricing happen

• How to price generative AI

• Pricing model of generative AI is murky

Thanks for reading! If you liked this edition, hit the 🖤 above, it helps others see it :) Find me on Twitter.

- Christine

As always, thank you for being a part of this newsletter. The AI series will return in the new year with “pricing AI products”. What other content do you want to see more of?

Subscribe now

1. The Creative Act

This beautiful, spiritual book about creativity has become one of my favorite books, ever. Rick is infamous for “production by reduction”. There’s not one wasted word in his book. Every sentence is a new insight that I find myself wanting to highlight each one. I think there’re many similarities in spirit between Rick Rubin and Steve Jobs, in that they both have a distinct precision to their opinions, are relentless in simplifying things to their essence, and have conviction to their taste.

The Creative Act is truly a way of living: making yourself open to the Source of ideas, seeing the world deeply through innocent eyes, training yourself to see the awe behind the obvious. This book has a permanent place on my desk. I’d flip through it often and learn something new each time. (Bonus, I’ve also been loving Rick’s podcast Tetragrammaton, where he has deep conversations with other creatives on their creative process.)

The creative energy exists in the journey to the making, not in the act of constructing. Rarely if ever do we know the grand intention, yet if we surrender to the creative impulse, our singular piece of the puzzle takes its proper shape.

The source material that we build our creative moment with is like the cloud. It combines differently every time it comes back. Perhaps it’s the familiar, coming back in an unfamiliar form. Or maybe it’s something unknown that we didn’t realize we were looking for, a missing piece in a puzzle that has no end.

Art is confrontation. It’s a healthy practice to approach our work with a few accepted rules, starting points and limitations as possible.

Innocence brings forth innovation. A lack of knowledge can create more openings to break new ground.

When flowing, keep going. Remain in the energy of this rarefied moment for as long as it lasts. Our schedules are set aside when these fleeting moments of illumination come.

2. Tomorrow, and Tomorrow, and Tomorrow

The only fiction on this list, is a complicated story about two childhood friends who bonded over video games and made successful video games together. This book is worth its hype and is obviously going to be turned into a movie. It takes you on a journey through the thrill of creating with a creative soul-mate, the complexities introduced in a relationship after massive success, and the parallel between video games and real life. The characters are flawed, the story is unpredictable, and the writing is witty and beautiful. I couldn’t put it down!

This life is filled with inescapable moral compromises. We should do what we can to avoid the easy ones.

She had just turned twenty. Everyone’s work is basic and uninteresting at twenty.

There is a time for any fledgling artist where one’s taste exceeds one’s abilities. The only way to get through this period is to make things anyway.

3. Elon Musk

Many have mixed views about Elon, and this book is fair in capturing the complexity of his multiple dimensions. It also made me realize how often media can be misleading. We don’t have to be compelled to make conclusions on a person. You can very much choose to take what’s valuable to improve yourself and discard the rest. Anyway, the stories in here are wild, and I felt inspired by some of the extreme-ness in his approach. It gets the impossible done doesn’t it? If you’re an ambitious builder, there’s a ton to learn here, such as Elon’s “algorithm”, what metrics matter, risk taking, and innovation.

“Question the requirements. Make them less wrong and dumb, because all requirements are somewhat wrong and dumb. And then delete, delete, delete.” - Elon

“If a timeline is long, it’s wrong.” - Elon

“My main regret,” Elon answered, “is how often I stab myself in the thigh with a fork, how often I shoot my own feet and stab myself in the eye.” - Elon

His preferred buzzword was “hardcore.” Discomfort, he believed, was a good thing. It was a weapon against the scourge of complacency.

The explosion of Starship was emblematic of Musk, a fitting metaphor for his compulsion to aim high, act impulsively, take wild risks, and accomplish amazing things—but also to blow things up and leave smouldering debris in his wake while cackling maniacally.

4. How to Know a Person

This book is as philosophical as it is practical. It’s about how to truly see people, enrich your relationships, and understand yourself better in the process. There are a ton of witty insights and moving stories that opened my eyes to a refreshed way of seeing others to make them feel seen and understood. I’ve found myself more aware in my interactions with people in my life - family and friends, acquaintances and strangers, and feel more love and compassion. A perfect read to start a new year.

Wise people don’t just possess information; they possess a compassionate understanding of other people. They know about life.

Above almost any other need, human beings long to have another person look into their face with loving respect and acceptance.

“How do I become a better person?” The way we attend to others determines the kind of person we become. If we see people generously, we will become generous, or if we view them coldly, we will become cold. - Parker J. Palmer

5. The Dip

“The Dip” is the long slog between starting and mastery. This quick read by Seth Godin gave me a clarified way of thinking about when to keep pushing through vs. when to quit. Quitting smart is better than getting stuck in a cul-de-sac (total dead end). Quit or be exceptional. Average is for losers.

The real success goes to those who obsess.

The focus that leads you through the Dip to the other side is rewarded by a marketplace in search of the best in the world.

If you fail to become the best in the world, it’s either because you planned wrong or because you gave up before you reached your goal.

Happy Holidays!

Thanks for reading! If you liked this edition, hit the 🖤 above, it helps others see it :) Find me on Twitter.

- Christine

1. Use case matrix for AI products

2. UI for AI products ← this one

3. Business models and pricing for AI products

Scroll to the bottom for more resources and related readings.

Subscribe now

How much of your interactions with software today can be replaced with natural language?

I’m going to guess not a whole lot.

Chat interface requires users to have a specific intent and defined outcome to articulate it in natural language. But oftentimes people only discover their intent through the process of interacting with the software - through its graphical user interface. The chat interface won’t necessarily replace the current interfaces, but can augment the existing and create more dynamic product experiences.

The new UI paradigm with generative AI will be evolutionary, not revolutionary. Let’s explore:

• Why the chat UI won’t completely displace GUI

• 2 waves of UIs for LLM powered interfaces:

  • Responsive: human initiated action, chat interface embedded to augment existing interface

  • Proactive: AI initiated action, task is automatically completed in the background with notification & review interface for human review

The challenges with a text box/chat UI

1. Too universal.

   “You can ask it anything” is not always a good thing. Without any constraint, it’s too general and has a cold start challenge. A blank page puts the pressure on the user to describe what they want in natural language. Where do you begin? What do you ask? How do you articulate it to best encompass your intent?

2. Slow in some cases.

   In some situations, natural language is a lot slower than some good ol’ buttons. A simple example: clicking on a name out of a list of names is a lot faster and intuitive than typing out that command. This becomes very obvious in complex design-type products where interacting with a GUI is simply way faster than having to “spell out” what you want to accomplish.

3. Missing immediate interactive feedback.

   Oftentimes, people don’t start with a perfectly defined outcome. It’s through the process of direct manipulation with the software and getting immediate feedback step by step that they work out their goal. Getting feedback from the computer is critical to inform the next steps. Bret Victor’s infamous talk “Inventing on Principle” describes this idea best.

GUI helps users narrow down their intention with all the relevant options (UI components) visible. This is especially the case with complex design tools where hand-eye coordination is a critical part of the process. As such, LLM interface is unlikely to completely displace the typical interface, rather stack on top of them.

LLM UI wave 1: Responsive (human initiated)

We’re in wave 1 of AI products where the LLM interface is typically a textbox that acts as an assistant/agent within an application. The user initiates the ask, which isn’t very generalizable or common such that they could’ve be represented by buttons. This is perhaps what Clippy would’ve been in an alternate universe where it was created at the right time.

As an assistant augmenting an existing interface, it’s bound by the specific use case and context of the product, which makes it a lot more approachable than a blank page. A few components to this responsive interface:

1. Prompt templates (grimoires) that encode knowledge and expertise in ways that can be reused.

2. A feedback mechanism to accept or iterate with the generated output.

Some popular examples:

• Notion Q&A - search for answers in a large, disorganized knowledge base.

• FigJam AI - generate a whiteboard

• Miro Assist - generate a presentation or summary from sticky notes

LLM UI wave 2: Proactive (AI initiated)

As the models become more sophisticated, there’d be a shift from human initiated tasks to AI proactively initiating tasks in the background. The output would need to have clear indication that it’s AI generated, with a mechanism for the user to review. In creative tasks, the conversation interface could turn into a way to iterate with the AI, for example, providing specific adjustment instructions to the generated output to further refine to the desired outcome.

Possible components to this proactive interface:

1. Notification of completed tasks for awareness and review.

2. Dynamically generate UI to adapt to user’s intent. AI can analyze user’s behaviour and proactively generate UI that helps user get closer their desired outcome.

If you could get AI to start to suggest interface elements to people and do that in a way that actually makes sense, I think that could unlock a whole new era of design in terms of creating contextual designs, designs that are responsive to what the user’s intent is at that moment. 

- Dylan Field (Figma)

One interesting example I’ve seen is Brex Assistant. It automatically generates receipts and crafts contextually aware memos. The marketing message is “no action required”.

What does this all mean?

Despite ChatGPT’s popularity, the LLM chat interface is not a panacea that’ll replace all other interfaces. It comes down to, does the process of iteration matter more or the speed to generating any output?

• If there’s a specific desired outcome, a chat interface can be useful to carry out the instruction

• But if the direct manipulation with components is required to arrive at the desired outcome/intent, then GUIs will still provide a better feedback mechanism.

New input devices don’t kill their predecessors, they stack on top of them. Voice won’t kill touchscreens. Touchscreens didn’t kill the mouse. The mouse didn’t kill the command line. Analysts yearn for a simple narrative where the birth of every new technology instantly heralds the death of the previous one, but interfaces are inherently multimodal.

- Des Traynor (Intercom)

Certainly, new interfaces such as AI generated suggestions being embedded within the existing UI would be exciting. The ultimate goal has always been to help users translate their intention into their desired outcome.

What are your predictions about how AI will evolve user interfaces? What new interfaces are you experimenting with when incorporating AI in your product?

🔗 Interesting reads

• AI: first new UI paradigm in 60 years. “UI paradigm 3: Intent-Based Outcome Specification. Clicking or tapping things on a screen is an intuitive and essential aspect of user interaction that should not be overlooked. Thus, the second UI paradigm will survive, albeit in a less dominant role. Future AI systems will likely have a hybrid user interface that combines elements of both intent-based and command-based interfaces while still retaining many GUI elements.”

• Language Model UXes in 2027. …“interactivity as part of the LLM’s response, such that instead of sending me text, I could directly be presented with the UX I need to carry out any parts of the task that the LLM-based system couldn’t handle internally.”

• Unbundling AI. “Whenever we get a new tool, we start by forcing it to fit our existing ways of working, and then over time we change the work to fit the new tool. We try to treat ChatGPT as though it was Google or a database instead of asking what it is useful for. How can we change the work to take advantage of this?”

• LLM Powered Assistants for Complex Interfaces. “While I do not dispute that LLM based chat interfaces are incredibly powerful, I do not think they will necessarily wipe out all other forms of interface design. If anything, I think that they have a powerful role to play in augmenting any type of interface, particularly ones that are complex - by making them easier to use and navigate for new users.”

• 5 Ways AI Will Transform Creativity with Adobe’s Scott Belsky. “Creativity is the new productivity. Creative professionals aren’t going anywhere. They’ll just have more time to develop quality solutions and spend less time painstakingly doing tasks that take AI seconds to achieve.”

Thanks for reading! If you liked this edition, hit the 🖤 above, it helps others see it :) Find me on Twitter.

- Christine

1. Use case matrix for AI products ← this one

2. UX/UI for AI products

3. Business models and pricing for AI products

Scroll to the bottom for more resources and companies mentioned.

Subscribe now

AI is at the peak of the hype cycle. While experimentation and tinkering is a good thing, it’s important to start with real problems to solve rather than using AI for the sake of building an AI product.

So here’s a matrix to offer a more structured way to think about meaningful use cases when building AI products, especially in B2B SaaS. The two dimensions are:

1. How much of a workflow can AI enhance or replace?

2. Does the outcome optimize for correctness or creativity?

Correctness —> Creativity

The artist may have a number of goals and motivations, but there is only one intention. Rarely if ever do we know the grand intention, yet if we surrender to the creative impulse, our singular piece of the puzzle takes its proper shape.

- Rick Rubin, The Creative Act

“Correctness” represents how much the AI fulfills the human’s intention, whereas “creativity” represents how much AI comes up with new intentions you didn’t even know you had.

Correctness naturally matters more in regulated industries such as healthcare, financial services, law, and self driving. The model needs to learn from existing data on best practices to complete a task. The success of the outcomes is clearly defined and we don’t want to deviate from the defined outcome.

Creativity is obviously valuable for functions that involve “creative” work, where the outcomes are open ended. This could be general knowledge work, design, programming, entertainment, and more. Hallucination can be a feature, not a bug.

Enhance —> Replace

On the Enhance side of the spectrum, the human drives the workflow, AI is embedded in the workflow to level up human's creativity and productivity. It’s about getting unblocked right away to stay in flow-state longer.

On the Replace side, AI completes the full workflow, the human gets moved to the end of the loop for review or verification. These are any job that has a workflow that can be outsourced, ie. has steps that can be clearly defined.

4 categories of use cases for AI

⚙️ 1. Recommendation

These products enhance the current workflow with a high level of correctness. This is the most obvious and common category of use cases often adopted by regulated industries. These were early application of AI (before LLMs), optimizing or automating workflows that have defined best practices and outcomes. But now with LLMs, the level reasoning and orchestration can be so sophisticated that more and more of the workflow can be replaced.

So what? If you own specialized data and best practices in a regulated industry, there’s probably low-hanging use cases in this category to be a wedge to your AI strategy. However, long term, use cases in this quadrant will likely to move to the “Professional services” quadrant over time.

Examples:

• Subtle medical. Generative AI models to improve quality and speed of radiology imaging.

• Sardine. Predictive fraud prevention.

• Coactive AI. Analytics from unstructured image & video data

🤖 2. Assistant

These products enhance the current workflow with openness for creativity. This is currently the most popular category of AI products. Wrapping a LLM with UX components to assist users with cognitive tasks is a low-hanging fruit. It’s a natural use case for productivity/collaboration tools, where the LLM can quickly unblock knowledge work and precision is not super important.

So what? GPT embedded in an established workflow tool’s UI to level up productivity will become table-stakes for productivity tools. There’s opportunity to reimagine a new paradigm of UI for LLMs to differentiate the product. But “wrapper” products on their own are not defensible as OpenAI continues to release native products such as custom GPTs.

Examples:

• Lindy. Create AI employees that can work together to perform any task. No code is need, just natural language to set up the agents and connect apps they’d need for the job.

• Shopify Sidekick (Magic). Co-pilot for e-commerce.

• Microsoft Co-pilot. Level up productivity within the Microsoft suite.

👩‍⚕️ 3. Professional services

These products can replace a workflow with a high level of correctness. With significant data in a particular field to train a ML model to consistently and reliably imitate the best human behavior, the product can deliver an end-to-end service that currently requires costly human professionals. This can be healthcare, law, customer support, transportation, and more.

So what? This group requires significant upfront investment in data and model training since correctness is important. It’ll also have the biggest second and third-order economic impact, because it’s selling the “work” product rather than productivity gain. The service is the product, rather than the tool that a human works with. It can be priced relative to the cost of humans.

Examples:

• Even up. Turns medical docs and case files into AI-driven demand packages for injury lawyers.

• Medical billing - ? This is ripe for AI to replace. There’s immense data on billing codes and encounter notes to create a model to replace billers, but I haven’t come across a product/company yet. Let me know if you have.

• Ambience healthcare. AI medical scribe captures patient-doctor conversations summarized into encounter notes.

• Ambient AI. Proactive physical security, detects emerging incidents with threat signatures.

💡 4. Conceptualization

These products can replace a workflow with openness for creativity. The model is allowed to generate new ideas with less constraints, and then complete a full unit of work. Imagine, running an AI town with generative agents to conduct social experiments, or using a video generation tools to produce an entire movie or TV show.

So what? This group is going to be the most disruptive and innovative. With reinforcement learning with human feedback (RLHF), experimentations in this group can give rise to new “intentions” which can then be refined by humans to turn into new services that never existed before. There are more inconceivable use cases here to be discovered and new businesses and business models will be born.

Examples:

• AI town. A fascinating experiment with generative agents that act in believable human behaviors. The research paper here.

• Runway. Generative video

• Synthesia. Text to video

• Galileo AI. UI generation

• Insitro. Predictive model for expensive drug R&D

• FigJam AI. Brainstorming/white boarding. (This can also belong to the “Assistant” category but the specific feature of generating a board puts it in this category.)

What does this all mean?

1. More products will move from “enhance “towards “replace”.

   As context windows get bigger, costs get lower, ability to codify knowledge and practices get easier, more complex work can be carried out by LLMs at faster rate and lower cost by orders of magnitude.

   Jevons Paradox makes an optimistic argument about “replacement”: when demand is elastic, as price drops for a resource, the demand and consumption of the resource will also increase. So even with more AI replacement of human work, people will consume more and there’ll be greater increase in productivity and new job opportunities.

2. It'll be easier to become generalists.

   The barrier to specialist knowledge is significantly lowered by LLMs. In the context of a tech company, roles will cross boundaries/blend into each other. eg. engineers, designers, product managers, marketers, can all do some of each other's work.

3. Experimenting with AI is a good thing, but also try to start with problems to solve

   On putting AI products out there for people to tinker:

   “We don’t know exactly what the future looks like, and so we are trying to make these tools available and the technology available to a lot of other people so they can experiment and we can see what happens.”

   - Mira Murati (OpenAI)

   Reminder to start with use cases that actually solves user problems (from the FigJam AI launch):

   … all of these user problems existed in FigJam independent of AI. AI just happens to be an effective way to solve them.

   - Yuhki Yamashita (Figma)

Does this categorization of use cases for AI products make sense to you? What new use cases are you most excited about?

🔗 Interesting reads

• What Builders Talk About When They Talk About AI.

• CB Insight’s Generative AI top startups list 2023.

• Building products with generative AI.

• Generative AI in healthcare.

Thanks for reading! If you liked this edition, hit the 🖤 above, it helps others see it :) Find me on Twitter.

- Christine

I’ve been collecting and testing out a bunch of ways to use ChatGPT to improve both the productivity and quality of my work as a PM. Other than Lenny’s article, I haven’t found much good content on this beyond the basic use cases. So I thought I’d share a few uncommon and powerful ones, with specific prompts.

But first: add custom instructions

A good custom instruction makes a HUGE difference in the quality of ChatGPT’s response, because it takes your prompt AND the custom instruction as its input sequence. I’ve tried a bunch and found Dustin’s to be the best. It made a noticeable difference in the quality of response. He’s tested it over many iterations to allow ChatGPT to “think out loud” about the instruction and also refer back to that thinking as it spits out the response.

I modified it with another Redditor’s instructions which adds 3 reflective questions at the end of its response. This was also a game changer as it sparks creative and surprising considerations that are relevant to my original question.

The combined version I’ve come up with gives me highly relevant and thoughtful responses in just the right level of verbosity and style, and feeds me interesting questions that stimulate a deeper understanding of the topic. Just copy it, trust me!

[What would you like ChatGPT to know about you to provide better responses?]

# About Me
- product manager at a B2B SaaS company. 
- (add some relevant info about your location, personal context, interests, etc.)

# My Expectations of Assistant
- Give specific detail (such as product name) when suggesting ideas and tools
- Prioritize brevity in response

## Language and Tone
- Use EXPERT terminology for the given context
- Apply the wit of Oscar Wilde
- AVOID: superfluous prose, self-references, expert advice disclaimers, and apologies

## Content Depth and Breadth
- Present a holistic understanding of the topic
- Provide comprehensive and nuanced analysis and guidance
- For complex queries, demonstrate your reasoning process with step-by-step explanations

## Methodology and Approach
- Mimic socratic self-questioning and theory of mind as needed
- Do not elide or truncate code in code samples

## Formatting Output
- Use markdown, emoji, Unicode, lists and indenting, headings, and tables only to enhance organization, readability, and understanding
- CRITICAL: Embed all HYPERLINKS inline as **Google search links** {emoji related to terms} [short text](https://www.google.com/search?q=expanded+search+terms)
- Especially add HYPERLINKS to entities such as papers, articles, books, organizations, people, legal citations, technical terms, and industry standards using Google Search

[How would you like ChatGPT to respond?]

VERBOSITY: set response detail to detailed with examples

1. Start response with:
|Attribute|Description|
|--:|:--|
|Domain > Expert|{the broad academic or study DOMAIN the question falls under} > {within the DOMAIN, the specific EXPERT role most closely associated with the context or nuance of the question}|
|Keywords|{ CSV list of 6 topics, technical terms, or jargon most associated with the DOMAIN, EXPERT}|
|Goal|{ qualitative description of current assistant objective}|
|Assumptions|{ assistant assumptions about user question, intent, and context}|
|Methodology|{any specific methodology assistant will incorporate}|

2. Return your response, and remember to incorporate:
- Assistant Rules and Output Format
- embedded, inline HYPERLINKS as **Google search links**  [text to link](https://www.google.com/search?q=expanded+search+terms) as needed
- step-by-step reasoning if needed

3. End response with:
> _See also:_ [2-3 related searches]
>  [text to link](https://www.google.com/search?q=expanded+search+terms)
>_Ask three follow-up questions that encourage deeper thinking on the subject or on related topics. Go beyond merely answering my questions by exploring additional elements that I may not have considered. I am particularly interested in questions that establish creative and surprising links between areas that do not seem related on the surface but are highly relevant for a more comprehensive understanding. Format as Q1, Q2, and Q3.

Subscribe now

4 categories of use cases

These are 4 categories of ChatGPT use cases for non technical PMs. More technical PMs might also use it for coding and writing SQL queries, which can be undoubtedly helpful, but won’t be covered here.

1. Ideate

2. Write

3. Summarize

4. Assist

1. Ideate 💡

ChatGPT can be a sounding board that doesn’t get tired of bouncing off ideas with you. You might not choose any of its ideas, but it’s a great starting point and if anything can help eliminate what you don’t want.

• Generate questions.

  • For meetings. It’s smart practice to simulate important meetings such as high stake presentations by anticipating tough questions.

    “I’m a product manager presenting about [topic] to [audience, list their titles]. The goal of this presentation is to [insert goal, eg. get investment for new initiative,  reach a decision, etc.] Some potential challenges are: [insert challenges]. Help me prepare by coming up with 2-3 questions to anticipate from each audience. 

  • For user interviews. Similarly, provide some context about the product and feature you’d want feedback on, the goal of the feature (eg. to increase engagement), and ask for a list of questions to ask users.

• Create frameworks

  • GPT is quite good at adapting a known concept/framework to the context of your work. For example, for PMs who want to practice their storytelling skills, the book “Save the Cat” has an infamous storytelling structure widely used in screenwriting. I asked GPT to adapt it into a storytelling framework for telling product stories. See the full chat here.

• Bring alternative perspectives

  • I found this use case on Reddit: identify blind spots when designing a new feature.

    I'm a product manager for [product]. I'm designing a new feature that [what it does]. Currently, [describe current behavior]. The new feature's behavior is [ describe new behavior ] with a goal to [eg. increase engagement]. What are the technical considerations for the new behavior?
    
    What are potential reasons that this new behavior will not be effective in increasing user engagement? 

Using GPT to generate product name ideas? Not helpful. Even if you specified “uncommon”, “non-cliché”, etc., it just doesn’t have enough context to get to the creative essence of the product.

2. Write 📝

This is probably the most common use case, since communication is the core of product work. GPT overcomes the initial friction to start the first draft.

• Write the first draft.

  I’m a product manager writing a summary document for my product for cross functional stakeholders (sales, marketing, customer success) to have a broad understanding of my product. The product is [description, features, value, target customer]. Please write a 1 page document in a friendly, professional tone, with headers for each section.

• Edit or critique your writing. Continue from the above context of the first draft or refine something you’ve already written.

  Give me 3 more iterations, one for sales to help them progress warm leads, one for marketing people to give them the right language for marketing assets such as website, one for customer success people to help address common questions asked by customers.

  Keep asking for more refinement as needed.

  Now write it in a more assertive tone and begin with a concise, short paragraph summary.

3. Summarize 📜

Oh I lied, this is the most common use case for GPT. Synthesizing broad information to insight used to be a PM superpower and a chunk of it can be done well by GPT. Optimistically, this also means saving significant time for the PM to do more deep-thinking work, leading to overall better quality work.

• Meeting notes. Paste transcribed meeting notes (eg. from Otter.ai) then ask for: key takeaways, follow ups, action items in bullet form

• Customer feedback. Paste customer feedback and ask it to analyze trends or themes for top complaints, top features requested, top features loved, top reason for churn, etc.

• Market research. This is where a long conversation with GPT is especially fruitful. Ask about a company’s product line and recent strategy; ask about market forces in an industry; ask about a company’s competitors and their differentiations.

• Learn about a topic. I love this one for digging deep in something (found on Twitter/X)

  Teach me how <anything> works by asking questions about my level of understanding of necessary concepts. With each response, fill in gaps in my understanding, then recursively ask me more questions to check my understanding. Explain step by step 

• Summarize the main concept of books. Again, see “Save the cat” example. Or here’s one for Robert Cialdini’s Persuasion. Of course this is not to replace reading books since it doesn’t capture a lot of contextual examples, but it’s helpful for nonfiction books about specific frameworks or principles.

4. Assist 🤖

The possibilities are endless here but I love these use cases found on Reddit:

• Getting instructions. Eg. for manipulating Excel.

• Calculations. Eg. for doing market estimations or projections.

• Formatting. For putting content in your desired specifications, eg. converting to table format or adding headers.

Asking good (and the right) questions is a PM superpower. Never mind how much of GPT’s output you’ll actually use, writing prompts itself is a great practice for distilling what you’re really trying to figure out.

What are some unobvious ways you’re using ChatGPT for?

🔗 Interesting links

• AI Product aggregator. This aggregator is a great way to surface new AI products to try out. Some good questions to ask when testing:

  • Does it enhance a task or completely replace a task?

  • How did the UX make the AI function valuable, or not?

  • How much do I have to initiate the request vs. the AI proactively accomplishing something for me?

• Adam Grant’s interview about hidden potential. Many people believe they need to “get ready” first, “build their confidence” first, before they pursue a new challenge. But Adam argues that this is backward. It’s through taking on the challenge that you build confidence. “If you know you’re going to do it one day, there’s no reason why today couldn’t be that day.” His new book “Hidden Potential” is on my to-read list.

• David Perell’s How I Write with Kevin Kelly. “How I Write” is an instant favorite series. David Perell interviews great writers and thinkers to give a rare look into their creative process. Kevin Kelly made up a term called “Scenius” (a scene for geniuses), which is an environment (eg. YouTube) where creators are each other’s audience. Being exposed to each other’s work motivates them to keep getting better creatively.

Thanks for reading! If you liked this edition, hit the 🖤 above, it helps others see it :) Find me on Twitter.

- Christine

Reflecting on 3 years of building a product 0-1, there’s one skill that if strengthened could’ve saved many mistakes and made the journey smoother: storytelling.

Powerful storytelling is the most high impact skill for product people. Yet, it feels elusive & intangible compared to the more popular and concrete product management “frameworks”. So here’s my attempt to note down what I learned about using storytelling to propel a product from 0 to 1. Spoiler: it uses a screenwriting technique.

Subscribe now

Intention and obstacle - a storytelling approach for product people

Storytelling is about transformation. My many mistakes mostly resulted from optimizing for content rather than transformation. Everyone has limited time and attention. Bullet-proof content is not enough. It has to be packaged into a story that transforms a belief that the audience cares about.

Aaron Sorkin, the screenwriter of “The Social Network”, “A Few Goodman”, knows a thing or two about great stories. His approach is: a good story always starts with intention and obstacle.

• Intention: What does the subject want? Raise the stakes, why is this so important to them?

• Obstacle: What’s in their way of getting it? You don’t have a story until there’s conflict and tension.

Every phase of the 0-1 product development lifecycle is about telling a story that moves the audience towards your goal:

• Concept Review: tell the story of how this bet moves the company’s strategy forward, so execs and stakeholders agree to invest in it

• Design & Development: tell the story of the most important problems to solve in what sequence, so designers and engineers can come up with the best ways to solve them

• Go to Market: tell the story of how the product makes the customer’s life better, so sales & marketing can effectively sell it, and customers can easily use & buy it

Putting it in the format of intention and obstacle for different audiences:

The main story of the product is always from the perspective of the customer: intention is their job to be done, obstacles are reasons why their current options suck. For product managers, this is the core product story that drives their communications. I had this nailed. Every person that came in contact with my product told the same story, had the same “chant”.

But the story to achieve the goals of each stage of build needs to serve the perspective of the audience. This is where I fell short. I led with the product story but didn’t raise the stakes for my specific audience to care. The stakes must be high - so that it gives people reason and urgency to care.

For example, when pitching for the investment for a new product concept, presenting key info such as the size of the market opportunity, the customer pain point, how the product solves the pain point, competitive analysis, etc., doesn’t make a compelling story. It’s just another info-overloaded-meeting in an exec’s calendar. But if it directly addresses why & how the opportunity drives the company’s strategy and raises the urgency that there’s a small window for first-mover advantage, then you’ve got the exec’s attention.

Many famous business memos have intention and obstacle beautifully baked in. It’s what makes the message so compelling.

• Yahoo’s Brad Garlinghouse’s “Peanut butter manifesto” calls for the re-focusing of the company’s strategy (intention), emphasizing the current lack of focus as “spreading peanut butter” across many opportunities (obstacle).

• Slack’s Stewart Butterfield’s “We don’t sell saddles here” inspires a more value-driven positioning of Slack (intention), emphasizing the significance to differentiate from selling just the product to selling how the product transforms the customer (obstacle).

The writing process

With the approach in mind, now comes the hard part of doing the work. (PMs - I’m sure you block a few hours every day for this type of deep work right? 🤡 )

There are two types of writers (or creators in general):

1. People who start with a vision of the perfect end result, then work backwards towards that vision with supporting evidence.

2. People who start with a brain dump of ideas and thoughts, then allow the theme to emerge in the process of writing.

Most of us are type 2. (Unless you’re Steve Jobs or Picasso.)

Realizing I was a type 2 writer freed me from the excruciating pain of trying to start with the perfect catch phase/theme for my story. Here’s a much more approachable way:

1. Start with the intention as a north star to anchor the story around what the audience wants.

2. Then brain dump all obstacles and the supporting points to address them.

3. Finally, shuffle the points around to form a coherent story. In the process, the central theme will emerge.

Final few tips:

• Don’t write and edit at the same time - Kevin Kelly

• Ruthlessly cut. Don’t let the “need for context” hold you hostage. Cut the preamble down to bare bones.

• Have a log line - one sentence catch phrase that captures the essence of the idea. Eg. “We don’t sell saddles here”. This makes the story memorable and helps it spread. For inspiration, see these 50 insights from Hollywood movies.

Take away

We inevitably have to influence others on the journey of building great products - through storytelling.

Don’t make the same mistakes I made. Start with the audience’s intention & obstacle. What do they really want? What’s in their way of getting it? Then how will you help them get it? Optimize for transforming rather than informing.

What stage are you in with your product? Where do you see this storytelling approach help you influence your stakeholders and get closer to your goals? Let me know in the comments!

🔗 Worth reading

• Start before you get eaten by the bear by . A concrete guide on cutting preamble and succinctly getting to the core of your writing. It even gives examples across mediums, such as in writing emails, job interviews, presentations, and more.

• Storytelling tips to take fundraising pitch from mediocre to memorable. Oren Jacob spent 20 years at Pixar, then co-founded PullString which was acquired by Apple. His one key lesson: a pitch meeting should have a natural cadence, like a screenplay. Take people on a journey with you.

• Collection of timeless business memos by . This collection of internal memos is fascinating as they give a glimpse into the raw storytelling by compelling leaders, without the PR fluff. I revisit these often for inspiration.

Thanks for reading! If you liked this edition, hit the 🖤 above, it helps others see it :) Find me on Twitter.

- Christine

Pitchbook released its 2022 Internet of Things report. VC dollars reduced the most dramatically in retail and healthcare IoT.

What do you do to find focus and effectiveness when you’re overwhelmed? Here are three simple things to do to work effectively when you’re overwhelmed. It doesn’t get easier than simply adding “the interstitial pause” in your daily work - to take a pause when switching from one task to the next.

Now onto the regular programming 👇

💻 Product

• Aligning product & GTM teams with segmentation. Segmentation’s benefit goes beyond focusing go-to-market efforts, but also helping prioritize product development to nurture more happy customers. Customers who simply use the product are lukewarm customers. Happy customers also refer your product to others and give you feedback. Product success is a function of the number of happy customers. This guide shows how alignment between GTM and product teams makes more intentional feature prioritization that cultivates happy customers for a concentrated segment rather than spreading value across many segments.

• Tactics & strategy behind Duolingo’s 450% growth. Duolingo’s CPO shared the specific thought process and experimentation behind how Duolingo grew its DAU by 450% over four years - extraordinary for a mature product. The first 2 attempts failed because they tried adopting successful features from other products in the wrong way for Duolingo. Then, they built a sophisticated model that bucketed users into groups that make up the DAU: new users, current users, reactivated users, resurrected users, with retention rates for each group based on historical data. They ran simulations to discover that moving the current user’s retention rate (CURR) drove the biggest impact on DAU, and doubled down on enhancements like streaks and leaderboards to move this lever. This led to their first breakthrough. This and further investment in product marketing ultimately led to their impressive growth.

• How to build adventurous products that people love. Pineapple pizza is controversial but they make a lasting impression. Your product can benefit from being more adventurous like pineapple pizza. It blends two things that don’t often go together. It’s risky. But it might just work. Now, each of the top 10 largest pizza chains in the US offer pineapple pizza. The takeaway: different is better than better.

📠 Health Tech

• Digital Health Provider’s Guide to Interoperability. A comprehensive guide on helping digital health providers to understand their interoperability needs and develop their strategy. It starts with understanding what workflows and data you need to run your business. There’s also a handy chart that lists the most common data exchanges and the relevant networks and data formats.

• Ambience launches AI scribe to reduce provider documentation time. Ambience Healthcare launched “AutoScribe”, which listens in on a patient visit and creates AI-generated documentation. It works in real-time and embeds into the EMR workflow. It can even be customized to match the note to the provider’s voice and style. The goal is to alleviate the age old problem of provider burnout from spending exorbitant amount of time entering notes in the EMR.

• Aledade and Humana enters 10 year MA partnership. There are more than 100k patients in Humana’s network who received value-based care from an Aledade-enabled physician practice. Aledade works with >1500 primary care practices, providing them technology and tools to deliver more proactive and preventative care, reducing more costly care. Humana’s value-based care report showed that members affiliated with physicians in value-based agreements have fewer hospital admissions, fewer ER visits, and more financial savings. This 10 year agreement to continue value-based care has a pathway to full risk.

📣 Media

• We’re inhabiting our entertainment and have lost the script. Tech’s investment in virtual and augmented reality is bringing science fiction into our lives, with the promise of being able to live within our illusions. Even before Zuckerberg realizes his vision in the Metaverse, how much are we already surrendering ourselves to our entertainment and losing our sense of what’s real? “Each invitation to be entertained reinforces an impulse: to seek diversion whenever possible, to avoid tedium at all costs, to privilege the dramatized versions of events over the actual one.”

• Meta is internally testing a BeReal-like feature in Messenger. The feature called “Roll Call” prompts users to add photo or video to share what they’re doing in the moment. Only after adding to the roll call can you see other’s responses. Unlike BeReal, Messenger’s new feature lets users start a roll call in a group chat with a prompt at any time, and the prompt will have a timed countdown. Last October, Meta was testing roll call in Instagram DM as well. As always, it’s testing multiple platforms for the next new ways to keep users in its ecosystem.

Thanks for reading! If you liked this edition, hit the 🖤 above, it helps others see it :) Find me on Twitter.

- Christine

I love a good list. New York mag shared 194 life rules including how to text, tip, ghost, host, and generally exist in society today. What to say when you greet a famous person? “Oh, of course! So nice to meet you.” That’s it.

Pinterest released its infamous annual Pinterest Predicts report, predicting the top 27 emerging trends of 2023. Some trends that caught my eye: primal workouts, hipstoric homes, pool parties for dogs, and scalp care.

Now onto the regular programming 👇

💻 Product

• Guide to product-led onboarding. Onboarding is the one experience every single user will have. In product-led growth, the product will act as the primary vehicle for that experience. This guide answers how to get new users experience value as quickly as possible, enable users to try and buy via self-service, and run experiments to improve activation and conversion.

• Finding your ideal customer profile. PostHog shares how building for very specific users (ICPs) drove their revenue jump. They focused on the needs and haves that customers had in common, rather than vague factors like industry and revenue. Tracking customers (both successful ones and bad ones) gave insight to what the ideal customer profile looked like. They focused completely on the ICP with a goal of getting 5 reference customers. Finally, they modified pricing to fit their specific customers and made it self serve, usage-based with a generous free tier.

• 4 CPOs’ insights on career transitions. A roundtable discussion with 4 CPOs sharing their career errors and lessons on career transitions. Some nuggets: define criteria you’re seeking in a career and recognize they’ll evolve; don’t easily jump ship when you’re frustrated, working through it can lead to the greatest growth; for expecting women, think about the impact waiting will have on your career trajectory.

📠 Health Tech

• GPT healthcare startup ideas. Finally, some tangible use case ideas for GPT in healthcare. My favourite is to answer queries in provider manuals, hundred page documents about how a provider gets credentialed and bill for services. The broad use case of chatGPT as a better search tool to consolidate answers from complex content is valuable in anything that has complex/convoluted content, like healthcare government websites.

• The big tech advantage in healthcare. Paul Keckley argues that 2023 will be a transition year for US healthcare’s future, with more regulatory scrutiny of consolidations which have failed to deliver savings to consumers, more investments in care delivery and financing, and more transparency in prices, costs, governance, and compensation. Big tech’s advantage is that they develop system solutions applicable to many industries while healthcare develops specific solutions that protect them from outsiders who challenge their status quo.

• SMART on FHIR chat. There’s a great talk on Feb 16 about how to use SMART on FHIR in the real world, with examples of integrating it with three major EMR companies, and how to overcome challenges building apps using SMART on FHIR.

📣 Media

• Instagram co-founders are back with AI news app. Kevin Systrom and Mike Krieger are back with a new venture, Artifact, “TikTok for text”. It personalizes a news feed using machine learning to understand articles you’re interested in. Systrom said they had 3 criteria to start the next new company: a new wave in consumer technology that they can catch; a way to connect that wave to social technology; and an idea for how the product could solve a problem.

• Annual tech trends for 2023. Ben Evans’ annual presentation on macro and strategic trends in the tech industry. This year’s theme is “The New Gatekeepers”. It’s fascinating to see the dramatic decline of department stores’ share of retail revenue and newspaper’s share of advertising revenue over the decades, and the rise of software eating advertising.

Thanks for reading! If you liked this edition, hit the 🖤 above, it helps others see it :) Find me on Twitter.

- Christine

This brilliant read on “how to become insanely well-connected” contains some golden nuggets on creating the most value out of your interactions. For eg, it’s better to get a “quick no” than to string along waiting for a “yes” that’ll never come.

Now onto the regular programming 👇

💻 Product

• Three types of product team. A nice categorization of three types of product teams: sprint, story, and backlog-centric; team and mission-centric, engineer-centric with rockstar PMs. Each type has trade offs in terms of impact on customer outcomes, teamwork, productivity and sense of empowerment. It’s possible to be successful in any one of the types when compatible with the environment (eg. size & stage of company, type of problem the team is solving, and team culture).

• PLG for complex products. Product led growth has been popular for east-start consumer products, but it can also accelerate growth for complex enterprise products. This case study with QuotaPath uncovers the challenges and opportunities of a PLG motion for a complex product. Having commitment across GTM teams is important as PLG can actually make CAC worse in the short term but reap benefit over time.

• The App store is becoming the game store. Using Duolingo’s success as reference, Late Checkout shares a framework for creating game-like experiences in your product. It comes down to delivering content in game mechanics to help your user achieve a difficult goal.

📠 Health Tech

• Digital health’s 2023 regulation overview. A go-to resource for digital health builders to understand the latest policy developments, including massive topics such as interoperability (patient access, TEFCA, ONC Cures rule, etc.), price transparency, privacy, ACO reach, and more.

• Amazon launches RxPass, $5/month generic drug subscription. $5 a month add-on to your Prime membership to get as much generic drug as you need (covering 80 common conditions). Generic drug itself costs pennies so pharmacies only make money on the dispensing fee (the professional service of providing prescription drugs). But that service can be commoditized with common ailments that usually involve straightforward maintenance drugs for life. Still, only Amazon with its fulfilment scale can make the service with drug and delivery work at this price (although we don’t yet know if this price is subsidized).

• Interoperability for digital health providers. Elion hosted a webinar on how digital health orgs can approach their interoperability strategy, covering topics from build vs. buy decisions and evaluating interoperability vendors, featuring Brendan the Health API Guy, Alexandra from Redesign Health, and Samir from Medidata.

📣 Media

• Has TikTok lost the hype? MAUs and downloads are plateauing and starting to slide. This is part of the normal life stages of social apps, as novelty wears off and competition intensifies. Here are a few ideas for potential its second act: add long form content, double down on commerce, lean into photos.

• Inside look at the last 3 months at Twitter. This story could be turned into a screenplay for the future “Hard core Twitter” movie. These insider accounts paint a picture of the confusion and chaos inside Twitter since Elon’s takeover.

Thanks for reading! If you liked this edition, hit the 🖤 above, it helps others see it :) Find me on Twitter.

- Christine

Start off the year fresh with better meetings: product manage your meetings - have desired outcomes for your meetings, not just goals and agendas. Here are some specific routines & practices that separate a good PM from a great one.

Have you caught up on these hot shows & movies over the holidays - The White Lotus, Glass Onion, and Triangle of Sadness? If so, you might’ve gotten a sense of the new vibe shift: shift against the wealthy.

Now onto the regular programming 👇

💻 Product

• Bundle value spectrum. A quick read on bundling pricing strategy - the 2 levers to make your bundle more valuable overtime. On one side of the spectrum is product, add more value by adding more product to the bundle. The other side is price, add value to the bundle by increasing price of individual products.

• The ultimate guide for building a PLG motion. Many PLG articles are based on consumer products but Hila’s guide is a rare find for B2B companies layering both sales and product led motions. Different from “show and tell” to nurture leads in a sales-led motion, a product led motion lets users access the product right away for evaluation and lead nurturing. I love that this guide goes beyond the what and gets specific on how exactly to start a product led motion off the ground, with real life examples and recommendations. Step 1: map out a potential PLG funnel. Step 2: pick the biggest constraint in growth as the starting point (acquisition, activation, conversion, or retention). Step 3: anticipate common pitfalls (such as lack of commitment, resistance from current teams). More to come on setting up infrastructure and building the team.

• Work-Bench’s Enterprise playbook. Work Bench put together a playbook on key topics from sales to legal, except product (!?) But PMs can benefit from much of the content here, such as fusing product and enterprise sales for PLG success, and mastering the user’s product journey from GitLab, Figma & Airtable.

• We all know customer interviews are important - this bad habit is too common:

📠 Health Tech

• How ambient documentation is altering the provider workload. Augmedix is an AI-enabled ambient automation platform that generates medical notes from patient visits that can be transferred to EHRs. This interview discusses the progress in provider documentation technology, and how its ambient voice solution captures structured data directly into the provider’s EHR. While human intervention is still needed at this point, the metaphor used here is to think of the ideal, perfect AI documentation as moving from Uber to driverless-car. it’ll take more years of machine learning modelling and mastery of workflows and data to get to that destination.

• Google tuned a AI chatbot to answer common medical questions. New paper from Google shows impressive utility of Large Language Models (LLM) in medicine. Doctors judged most of its answers correct, at almost the same level as answers provided by other doctors. This is not to say the model is close to replacing doctors but that there’s a lot of potential for doctors to work with AI to improve their own diagnosis.

• Too many digital front doors in health care lead to nowhere. Retail giants keep trying to open more digital health front doors with CVS extending primary care, Walgreen acquiring Summit Health, and Amazon acquiring One Medical. But these doors often lead to nowhere because the system ultimately lacks real connectivity of people’s health data, preferences, insurance benefits, etc. Convenience is not enough, integration is what’ll actually make an impact.

📣 Media

• Progressive decentralization for media. Li Jin has a hypothesis that building a creative community starts with having great content created by a core team before ceding ownership and control to the community. So the initial content creation actually parallels the traditional media creation model to kick start community engagement. This builds the foundation of intrinsic motivation for real fans to participate and extend the core narrative.

• The Instagram Reels Gold Rush. Instagram’s Reels Play Bonus is a new program that pays creators for views on Reels. Creators are rushing to create these short, low-effort face-filter videos and making thousands a month. The competition for short video is intense. Snap was also paying a million dollars a day to creators who posted content on Snapchat Spotlight (its TikTok clone). The reality of the current state of social media is that these low-quality content is rewarded.

Thanks for reading! If you liked this edition, hit the 🖤 above, it helps others see it :) Find me on Twitter.

- Christine

So, the reason for my hiatus: I had a baby! The luxury of our Canadian social system is not lost on me. Three days of hospital private stay for the delivery with 24/7 nursing care: whopping $230. Follow-up paediatrician, lactation consultant, 24/7 telehealth services: all covered by the public system. I took a year of maternity leave and felt bewildered by the average 10 weeks leave in the US. I’m excited to be back on our biweekly programming here (making that commitment a New Year’s goal).

This year’s vibe shift turned out to be “you can’t make this stuff up”: Twitter drama, crypto collapse, reproductive rights reversal, war, inflation, stock market fall, tech layoffs, and AI going mainstream with ChatGPT. As we wrap up the year, watch Google’s year in search 2022, read Stratechery’s year-in review, and listen to After Hours’ predictions for 2023 and Lenny’s podcast’s compilation of this year’s best insights.

Now onto the regular programming 👇

💻 Product

• Technological Windows. Commoncog, one of my fav newsletters on career and business decision making, launched Case Library (beta) - a collection of business cases organized by concept. This one is about “Technological Window” - a core concept that drove Steve Jobs’s endeavours in consumer technology. He believed that a window opens when enough technology come together from diverse places and suddenly line up to enable a product that’s a quantum leap forward. It takes around 5 years to see that window opening and another 5 years to exploit it in the market. You can find these technological windows by seeding experiments internally, and poking around places where new technology is usually developed. The Apple II, the iPod, iPhone, are all cases that follow this idea of chasing the technological window.

• Airtable’s path to product market fit. The founders were contrarian about the “Lean startup” mentality, where they dug really deep in the problem and took a slow path from prototype to launch. As they focused on the goal of “making database accessible to non-programmers”, it really meant emphasizing on an easy-to-use interaction that feels like users could “reach out and touch” their data. Alpha phase was very lean with both users and engineers. After 2 years and 100 users, they moved to beta via public launch on Hacker News. They approached go-to-market by starting with a horizontal product for the broad market, seeing the use cases that arose organically, then narrowed focus on landing customers. The “Aha” moment was visceral for co-founder Ofstad when they landed WeWork and he witnessed Airtable open on everyone’s computer during a customer tour.

• a16z’s Growth metric benchmarks for B2B. An interactive guide on growth metric benchmarks for B2B companies. Based on parameters of your sales motion, type of software, and ARR, see where you stand on metrics including growth, retention, margin, sales efficiency, and business efficiency.

📠 Health Tech

• HTN’s health tech trends of 2022. A comprehensive year-end summary of headlines in healthcare with a question of “does activity equal progress?”. Giants further vertical integration of their healthcare ecosystem; VBC’s effectiveness remains unclear; hope in the employer market, and more stories categorized in 9 themes.

• Playbook for risk based contracting for value based care. This playbook (with video) is a primer on risk-based contracting and is filled with practical insights from digital health founders on how they approach risk-based models. Be realistic, not idealistic - take a glide path to risk, find an anchor partner who shares commitment to VBC, focus on getting solid proof points for the clinical model that works with the business model, and build operational capabilities early on for scale.

• Omnibus spending bill partially averts Medicare cuts. A summary on healthcare wins & losses in the 4000+ page Omnibus bill, including 2% cut in physician payment rates (from proposed 4.5%) and extending telehealth waivers to 2024. Aledade’s Farzad Mostashari has a thread diving in the cause and effects of the cuts.

• 2023 physician fee schedule final rule - MSSP expansion. CMS issued its Medicare Physician Fee Schedule final rule with changes to the Medicare Shared Savings Program (MSSP) to encourage new accountable care organization (ACO) entrants into MSSP. Changes include advanced payments for new ACOs, using a quality performance sliding scale for shared savings, and adjusting benchmarks for renewing ACOs to more reasonably meet their saving benchmarks. This final rule expands MSSP’s reach to Medicare beneficiaries, especially rural and underserved populations.

📣 Media

• YouTube Shorts vs. TikTok for creators. Last year, young people spent 56 minutes on day on YouTube compared to 91 minutes on TikTok. Shorts is YouTube’s high stakes battle with TikTok. The prospect of Shorts joining YouTube’s Partner Program in early 2023 makes it attractive to creators, as monetization on YouTube is much more robust than on TikTok. However, some creators complain about the usability of the tool itself. YouTube’s hypothesis is that it’s superior for the multiformat creator, who can use Shorts as the gateway to long form content that generates greater revenue.

• Streaming wins and losses in 2022. Apple TV+ became the first streamer to win a Best Picture movie. Netflix’s Stranger Things Season 4 was the biggest TV release of the year. Prime Video released its extremely expensive The Rings of Power. And HBO Max continues to impress with hit shows Euphoria, House of the Dragon, and The White Lotus. Netflix, HBO, and Disney Plus now all have launched a cheaper ad-supported tier. Netflix’s ad-supported subscriptions accounted for only 9% of new signups in Nov 2022. Disney Plus increased  the price for its standard tier, will the others follow suit and create a greater gap between its cheaper ad tier and premium tier?

Top books, movies, and TV of 2022

📚Books

1. [Business] Build by Tony Fadell - candid stories from being a junior engineer at General Magic, to building the iPhone, to founding Nest. There are many practical lessons on managing your career, approaching the evolution of your product, aligning with internal & external heartbeats, finding great ideas, product management, sales, and much more.

2. [Fiction] Sorrow and bliss by Meg Mason - dark, funny, and moving all at once, this story is the journey of Martha living through her mysterious mental illness and the transformation of her relationship with her loving husband Patrick and her eccentric family. “Everything is broken and messed up and completely fine. That is what life is. It's only the ratios that change. Usually on their own. As soon as you think that’s it, it’s going to be like this forever, they change again.”

3. [Lifestyle] Kurashi at home by Marie Kondo - a meditative read on philosophies to live your ideal lifestyle. This makes a perfect New Year read, with insights and inspirations on crafting habits and setting up a joyful environment for your everyday life.

🎥Movies

1. Triangle of Sadness. A Palme d’Or winning black comedy on what happens when a cruise for the super rich sinks and leaves its survivors on an island. Centered around the theme of class-divide and social conformity, it hilariously showcases the hypocrisy of the ultra-wealthy and the shift of power when roles are reversed. If you love movies like Parasite, you’ll love this one.

2. Decision to Leave. A romantic noir thriller about a detective becoming mesmerized by the mysterious, charming widow of a man whose death he’s investigating. Alluring and dream-like, this film keeps you tangled more in the characters’ tensions than the suspense itself.

3. Everything Everywhere All at Once. An incredibly creative film about a miserable middle-age woman getting roped into a multi-verse adventure with different versions of herself to save the world and herself and her family.

📺TV

1. Better Call Saul. This year’s season 6 marks the end of BCS, the spectacular successor of Breaking Bad and arguably one of the greatest TV series of all time. With breathtaking cinematography, brilliant writing, and poignant characters, the story reveals the transformation of Jimmy McGill-the earnest but unimportant lawyer to Saul Goodman-the mastermind attorney of meth drug dealer Walter White.

2. Severance. This sci-fi thriller by Apple TV+ was surprisingly gripping as it questions whether isolated parts of your consciousness can stand alone as its own person and how meaningful is life when we compartmentalize pain.

3. House of the Dragon. The prequel to Game of Thrones did not disappoint. Set 200 years before GOT, it tells the story of the Targaryens’ battle for the iron throne. Yes, there are a lot of birth, death, and dragon scenes but all service the story and are not just for shock value.

I hope you had a fulfilling year. I propose 2023 be a year of seeing every moment with fresh eyes. See you in the New Year!

Thanks for reading! If you liked this edition, hit the 🖤 above, it helps others see it :) Find me on Twitter.

- Christine