Let me tell you a story about one very expensive click.

In 2003, someone at Canadian energy company TransAlta made a copy-and-paste mistake in an Excel spreadsheet.

The bids were matched to the wrong contracts.

By the time anyone noticed, that ordinary mistake had cost the company $24 million.

Most businesses will never lose $24 million from one click.

They lose it slowly.

*The TransAlta incident was reported by CIO, citing the company’s CEO and Reuters.

That is where an AI consultant begins.

Not by selling AI.

By finding the quiet problems costing a business time and money, then building a better way to work.

The problem arrives first.

AI comes later.

That order matters, because a lot of people online are working in the opposite direction. They learned a few tools, added AI consultant to a profile, and went looking for somewhere to put them.

Clients can feel the difference inside ten minutes.

One person talks about models and agents and automation platforms. The other sits down next to an operations manager, draws the process out, points at the two steps where the time is disappearing, and says what should change. The second person is the consultant, and the gap between them has surprisingly little to do with technical skill.

It’s also why there’s no single course called Become an AI Consultant that finishes the job. The field now covers strategy, automation, governance, security and roughly twenty other specialties. Too broad for one certificate, too practical for credentials on their own.

The good news is that you don’t need most of it.

You need one useful intersection:

An industry you understand + a painful workflow + enough AI skill to improve it + evidence that your solution works.

Smaller target than becoming an AI expert. Worth considerably more.

Courses shorten the trip. They’ll give you vocabulary, structure and guided practice. What they can’t give you is proof that you can protect private data, handle an output that comes back wrong, win over a team that’s nervous about being automated, or leave behind a system that still works six months after you’ve gone.

Proof does that.

The master class below maps 25 real consulting paths. Each one gives you what you can sell, what to learn, a project you can build tonight without waiting for a client, and the test that turns that project into credibility.

Before you pick one, here’s the rule that will save you months of wasted study:

Don’t choose your path by the AI tool you enjoy. Choose it by the problem you’re qualified and willing to solve.

When you’ve found it, you should be able to finish this sentence:

I help [specific client] improve [specific workflow] by [measurable outcome] using a controlled AI system.

If the best you can manage is I help businesses use AI, you haven’t chosen yet. That sentence tells a buyer nothing, and it’s the main reason new consultants get ignored.

I don’t expect you to learn all of this today. I don’t expect you to finish 25 courses, master every tool, or start calling yourself an AI consultant by Friday. I expect one honest first step.

I spend a lot of my week on this. Testing tools, building workflows, breaking them on purpose, writing up whatever survived. I keep doing it because ordinary people can now build serious careers using capabilities that simply weren’t available to them a few years ago, and because most of the advice out there is written by people who’ve never had to hand a system to someone else and watch them use it.

Opportunity doesn’t count for much until somebody sits down and does the work.

I take AI consulting projects through Upwork and directly with companies, and that work taught me something the career guides leave out. You don’t need a large office, an impressive title, or permission from a major technology company. Most of this gets delivered remotely, from a home office or a kitchen table. Your age and your postcode matter far less than whether you can understand a problem, build something useful, and show that it works. [* There are some other platforms as well such as; Leland, Toptal, Contra, Freelancer, Fiver, Linkedin, X, and Local Market/Facebook pages]

Some of that proof is my own. In my day role I run security operations for a public post-secondary institution, and one of the projects I owned was the data-governance procedure behind a Microsoft Copilot rollout. It set out in writing which categories of health and financial information could reach the model, which couldn’t, who signed off on an exception, and how every one of those decisions was evidenced for an auditor.

It wasn’t glamorous. No agent. No demo video. Nothing you’d put on a landing page.

It was a workflow map, a data classification, a set of approval boundaries, and a document a reviewer could inspect without taking my word for any of it.

I mention it for one reason. This guide is about to ask you for evidence instead of claims, repeatedly, and I’d rather not ask you for something I haven’t done myself.

I didn’t want you taking the demand side on faith either, so I went back to Upwork and captured what clients were actually posting.

*The snapshot of ten Upwork listings captured in September 2026. Listings, budgets and availability change constantly. These show demand, not guaranteed income.

What matters isn’t any single advertised rate. It’s the spread of problems sitting behind the job titles.

One client wanted help getting a cautious business into AI without over-engineering it. Another wanted strategy for a marketing agency. A clinic wanted its website and its internal workflows fixed together. An insurance company wanted agents designed around claims and email. The rest were after product leadership, operational frameworks, creative direction and market research.

That spread is why I wrote this.

The 25 paths below are not 25 jobs you have to learn. They are 25 doors. I researched the courses, skills, tools and practice projects behind each one so you can find the door that matches what you already know, and start, instead of losing three months learning the wrong things.

Some of you will begin with automation. Others will spot the opening in finance, cybersecurity, healthcare or governance.

You only need one place to start.

What employed AI consultants can earn

Not everyone reading this needs to become a freelancer. The same skills lead to salaried roles inside consulting firms, technology companies and enterprise AI teams.

Published figures vary wildly, mostly because AI consultant describes everything from a general business-advisory role to a deeply technical position at a large firm. They also come in lower than what people quote on LinkedIn. ZipRecruiter puts the average near $114,000. Glassdoor has Data and AI consultants between roughly $104,000 and $186,000 across its 25th and 75th percentiles.

Treat that as the middle of the market, not the ceiling.

Directional 2026 bands, drawn from the sources above and from advertised roles. Not guaranteed salaries. The senior band is the softest of the three, because it swings hard on firm, specialization and city, and because the largest advertised numbers usually fold in bonus or equity rather than base pay. Check live listings in your own market before you plan anything around this.

Employment and consulting are different routes, and the honest difference shows up in year one.

A salary pays you from month one for holding a role. Consulting income starts at zero, arrives unevenly, and depends on the value of the result, the clients you win, and whether you deliver consistently enough to get asked back.

A realistic first milestone for a new consultant with one specialization and a solid proof project is a few thousand dollars a month across one or two engagements. That’s below an entry-level salary, and it should be. You’re being paid for a narrow result while you build the evidence and references that let you charge for a bigger one. The people who get past that stage do it by raising the value of the outcome, not by working more hours.

Don’t start by chasing the income.

Start by getting useful enough that someone will pay for the result.

This guide will show you how to choose that result, learn what it requires, build the proof and turn it into an offer.

You don’t have to change your entire life today. But you shouldn’t finish this article without knowing your first move.

Inside This Master Guide

Read more

I did not call this Vibe Coding 3.0 because another model was released.

I called it that because vibe coding changed what I believed I could become.

I was not a working programmer. I had studied computer science, but I had never built software for a living. I was not spending my evenings writing code, contributing to open-source projects or preparing to launch a startup. I never planned to become an app developer.

One evening, I opened Claude and described a small app I had no idea how to build.

A screen appeared.

It was rough. The design was ordinary. Some of the buttons did not work. But something had changed. An idea that had lived only in my head was now sitting in front of me.

That first build felt like a magic trick.

Then I asked the model to move one button.

The login stopped working.

I asked it to repair the login. It changed the database. I asked it to put the database back. It rewrote a page that had been fine ten minutes earlier.

By midnight, I had spent most of my usage limit trying to recover the version I already had at dinner.

I should have stopped there.

Instead, I built another project.

Then another.

Those projects did something a tutorial never could. They gave me proof.

I could show people what I had made. I could explain the problem, demonstrate the workflow and point to a working product instead of another certificate. That proof helped me begin finding work through Upwork. Now I take AI projects connected to mobile applications, security and practical automation.

Vibe coding did not turn me into a traditional programmer.

It gave me a different way into the work.

That is why I think this moment matters. A person with a useful idea can now build the first version without spending years waiting for permission. The skill can become freelance income, a product, a better job or the beginning of a small business. It can improve your living and give you more control over how you work.

It is not automatic money. You still need a useful problem, patience, judgment and the willingness to finish what you begin.

But the old barrier has moved.

You do not need to understand every line of code before you start. You need to sit down, choose one tool, choose one problem and stay with it long enough to produce something another person can use.

My fifth project felt nothing like my first.

The model was not dramatically smarter. I had simply stopped treating it like a magician and started giving it a workshop.

There was a written plan on the wall. Git gave me an undo button. Small files carried decisions from one session to the next. Tests decided whether a feature worked. The model handled one job at a time. When it made a mess, I restored the last clean version instead of arguing with it for an hour.

The work became calm.

That is the 3.0 part.

The first version of vibe coding was pure conversation: describe an app and keep prompting until something appears.

The second version added better models, tools and planning.

Vibe Coding 3.0 adds the part that turns a promising demo into dependable work: a clear specification, project memory, database permissions, recovery points, tests and a finish line that does not depend on the model saying, Looks good.

Before you finish this article, you should have a crystal-clear answer to one question:

What useful thing are you going to build first?

Not someday. Not after another course. Now.

Below, we are going to build a small booking app from an empty folder to a public URL. Customers will create accounts and request appointments. A staff member will see every request and approve or reject it. The data will live in Postgres. Row Level Security will decide who may read or change each row. Playwright will test the important path in a real browser.

It is boring on purpose.

A booking app contains the bones of most useful software: identity, data, roles, permissions, state changes, tests and deployment. Learn this shape once and the next idea stops feeling like a completely new world.

You do not need to memorize the code. You need to learn how to keep the work moving without losing control of it.

By the end, you will have either a working app or a precise failing check that tells you what remains. Both are more valuable than a beautiful screen you cannot trust.

Before we begin: what this guide can honestly promise

AI coding is not a photocopier. Two readers can paste the same prompt and receive slightly different code. Package versions change. Interfaces move. A model can misunderstand a clear instruction.

So this is not a promise that every screen will look identical after exactly ninety minutes.

It is a reproducible working method. Every stage produces something you can inspect: a file, a database table, a Git commit, a passing test or a live URL. If a stage fails, you stop there and repair that stage. You never pile new work on top of an unknown problem.

That is how real progress feels. Less magic. More proof.

Read more

My mother came into the kitchen holding her phone at arm’s length, the way she does when she has already decided the screen is at fault. There was a small charge on her account from a company neither of us recognized.

She asked me what it was.

Without looking up, I said, Mom, did you ChatGPT it?

My father started laughing :)

We all laughed.

I thought about it for the rest of the day.

Because she did ask ChatGPT. She showed me the answer later. It was clean, confident and perfectly organized. It explained what that kind of charge might be and gave her a sensible list of things to do next.

Then she called the bank anyway.

Of course she did.

The answer was never the work.

For two years, ChatGPT has been part of almost every working day for me. It is usually one of the first tabs I open and one of the last I close. Other tools arrived, stayed for a week and disappeared. ChatGPT remained.

But my working loop was almost identical to my mother’s.

Ask a good question. Receive a good answer. Then go and do the work.

The paragraph still needed verification. The number still had to be traced to its source. The summary still had to become a document carrying my name that another person could safely act on.

ChatGPT made me faster in the middle of the job.

I remained slow at the end.

That is the part of AI productivity we rarely measure. Generating an answer may take thirty seconds. Turning it into dependable work can take another hour.

The problem is not a shortage of intelligence. It is that the method still lives inside our heads.

We remember which sources to trust. We decide the order of operations. We know the format the final file needs. We perform the last check that separates something impressive from something finished.

Then we open another conversation and reconstruct the method again.

Codex showed me that the method did not have to disappear when the tab closed.

On June 2, OpenAI released six role-specific Codex plugins: Data Analytics, Creative Production, Sales, Product Design, Public Equity Investing and Investment Banking. Together, they package 110 skills and connections to 62 apps.

Each plugin carries more than a job title. It contains instructions, workflows and tool connections built around a particular kind of work. Instead of describing the profession again, you begin with a working method.

Subscribe now

The plugin produces the work.

The check decides whether the work is finished.

My mother still called the bank.

She was right to.

Start here before installing anything

Six plugins, six projects. Do not attempt all six. Choose one row from this table, finish it, and check it before you open a second.

Plugin availability may vary by plan, region, rollout and workspace policy. Connected apps also depend on what your organization permits. That is why each project below includes two possible routes.

The connected route allows Codex to retrieve authorized information from supported services such as HubSpot, Snowflake, Figma or approved financial-data providers, where available. The file route lets you complete the same exercise using your own exports, spreadsheets, PDFs, screenshots and notes.

Choose the workflow that matches your role and follow it step by step. Each section includes the required inputs, setup instructions, exact prompt, expected deliverable and acceptance check needed to test the result yourself.

To install a plugin, open Plugins from the sidebar, find the plugin you need and select Install. Start a new Work or Codex session, type @ in the prompt box and select the installed plugin. Connect only the apps and information sources required for that project.

If the plugin does not appear, check your plan, regional availability and workspace permissions. Search for it by name inside the Plugin Directory. If you are installing a plugin from an open GitHub repository, review its instructions and configuration before using it. Some templates may contain placeholder connector or app IDs that must be replaced with services available in your workspace.

For every exercise, ask Codex to show its plan before it makes consequential changes. Researching, analysing, drafting and preparing prototypes can usually proceed without intervention. Sending messages, publishing content, changing business records, deploying software or acting on an investment decision should always require your approval.

This version is honest because it no longer promises files readers cannot download. It also matches OpenAI’s current guidance for finding plugins and invoking a specific one with @. OpenAI’s ChatGPT Work guide.

1. Data Analytics: find the number behind the problem

Install the Data Analytics plugin.

For the connected route, select an approved warehouse or analytics service. For the file route, open the starter pack and upload 01-data-analytics/sales-data.csv with 01-data-analytics/business-context.md. The CSV contains eight complete weeks and the field definitions needed for the check.

Paste this prompt:

Codex should inspect the data, confirm the date range, calculate the funnel rates, identify the largest movement, and create a readable report.

Here is the arithmetic the report has to survive. If week one had 40,000 visitors, a 5 percent signup rate, a 20 percent purchase rate, and a $60 average order value, revenue is 40,000 x 0.05 x 0.20 x 60, or $24,000. If week two brings 34,000 visitors with all three rates unchanged, revenue is $20,400. Traffic explains the entire $3,600 decline and there is no conversion story to tell. Hold traffic flat and drop purchase conversion to 17 percent instead, and revenue lands at $20,400 again from a completely different cause. Same total, opposite action.

That is why the decomposition matters more than the summary. The result passes only when the revenue total matches the source file, every percentage can be recalculated by hand from two columns, the chart covers the correct dates, and every hypothesis is labelled as one. Recompute one week yourself before you read the conclusions.

If something fails, use:

The public Data Analytics template carries eleven example workflows, from metric diagnosis and KPI design through dashboards, validation, and data-quality assessment.

2. Creative Production: turn one brief into a reviewable campaign

Install the Creative Production plugin.

Connect an authorized creative tool when available. For the file route, upload all three files from 02-creative-production: campaign-brief.md, brand-rules.md, and product-reference.svg. They supply the audience, approved claim, visual rules, sizes, and product details Codex must preserve.

Use this practice scenario: an insulated bottle for people who commute.

Success is not six attractive images. Four checks decide it, and none of them require a second opinion.

View each ad at its delivered pixel size at 100 percent zoom and read the copy without enlarging it. Sample the hex values in the output and compare them to the brief instead of judging the colours by eye. Compare the rendered product against your source photo for lid, cap, proportions, and logo placement. Then write each claim beside the line in the brief that supports it, and delete every claim with nothing beside it.

The approval pause is the point of the prompt. It prevents Codex from polishing the wrong idea for twenty minutes.

3. Sales: walk into the meeting already prepared

Install the Sales plugin.

The connected route can use an approved CRM, calendar, email, meeting transcript, or document source. For the file route, upload all four files from 03-sales: crm-export.csv, customer-emails.md, meeting-notes.md, and proposal.md.

The check is mechanical. Pick any three claims in the brief at random and try to reach their source in one step. If a claim sends you hunting, it is an assumption wearing a citation. Beyond that, customer statements must be separated from seller assumptions, and every open commitment needs an owner and a date.

After the meeting, the same plugin can draft the follow-up and CRM-ready text. Review both before anything is sent or saved. The open Sales template covers fifteen workflows, from meeting prep and call follow-up through forecast review, account prioritization, and rep coaching.

4. Product Design: move from complaint to clickable proof

Install the Product Design plugin.

For this exercise, upload 04-product-design/current-page.svg and 04-product-design/redesign-brief.md from the starter pack. The files provide a deliberately weak newsletter subscription page, verified membership details, brand colours, and restrictions.

Most of the acceptance check is measurable. Every interactive control does what its label says. Body text renders at 16 pixels or larger at a 375 pixel width. Tap targets are at least 44 pixels square. Text and background clear a 4.5 to 1 contrast ratio. Free and paid membership sit in a single comparison with no hidden terms. No testimonial, subscriber count, or press mention appears that you did not supply.

One check needs another person, and it is the one that matters most. Show the page to someone unfamiliar with it, take it away after five seconds, and ask what the publication is about. If they cannot answer, the prototype failed regardless of how well it scored on the rest.

The official template supports nine workflows, including audits, visual exploration, screenshot-to-code, live-page recreation, and sharing after approval.

5. Public Equity Investing: make the thesis face the evidence

Install the Public Equity Investing plugin.

Use an approved financial-data connection or upload all three fictional files from 05-public-equity: investment-thesis.md, previous-earnings.md, and latest-earnings.md. This exercise organizes evidence. It is not personal financial advice.

The work fails if it mixes currencies, periods, adjusted figures, and accounting figures as though they were the same. Test it by choosing five figures at random and opening the cited page for each. If all five land where the citation says, and each one is labelled reported, adjusted, estimated, or commentary, the analysis holds. It succeeds when another reader can trace every number and understand why the classification follows from the evidence.

6. Investment Banking: build the comparison before the presentation

Install the Investment Banking plugin.

Use trusted connected data where authorized. For the file route, upload all three files from 06-investment-banking: target-company.md, peer-data.csv, and assumptions.md. They provide one valuation date, one currency, the target profile, a proposed peer set, and explicit calculation rules.

Every multiple must be rebuildable from two numbers you can point to. Pick one peer and reconstruct its EV/Revenue by hand from the stated enterprise value and the stated revenue, then confirm the fiscal period behind both. Check the valuation date, the currency conversions, the enterprise-value calculation, and the non-recurring adjustments. A polished table built from mismatched periods is still a failed analysis.

One repair loop for all six jobs

The most valuable prompt in this guide is not tied to any role. Use it after Codex creates the first deliverable:

This is the difference between using a plugin as a shortcut and using it as a working system. The first produces output. The second defines what finished means.

What I would try first

Do not install all six and connect every account on the first day.

Choose the plugin closest to a task you already repeat. Use files before live systems when possible. Run one narrow project. Check every important number, source, action, and file. Then decide whether the workflow deserves deeper access.

Keep the failures. The list of what broke on your own data is worth more than any launch summary, including this one, because it tells you exactly how much authority the system has earned.

For most individual readers, Data Analytics or Product Design offers the easiest first test because both can begin with one file or screenshot. Sales becomes especially useful when customer information is scattered. The two finance plugins need the strictest source discipline. Creative Production works best when the brief and approval point are clear.

The larger shift is not that Codex learned six professions overnight. It did not.

The shift is that professional methods can now travel with the tool. The instructions, sources, checks, and deliverable format no longer have to be reconstructed in every conversation.

Your role is still larger than the plugin. You decide what matters, what evidence is acceptable, what authority the system receives, and whether the result is good enough to leave the workspace.

Codex can carry more of the work. Judgment is still the part you should be careful not to outsource.

Most people will build their first AI company in about ten minutes.

They will open Grok Bot, create a Researcher, a Writer, and a Reviewer, then give each one a professional job description. By lunchtime, the Researcher will have collected the information, the Writer will have produced a clean report, and the Reviewer will have marked it ready.

It will feel like three people just finished a day of work.

Then the human will open the report.

One source is old. Another does not support the claim beside it. The Writer has filled a missing fact with a reasonable assumption, and the Reviewer has improved the wording without checking whether the argument is true.

Now the human has to reopen the sources, correct the claims, explain the assignment again, and repair the report that three AI workers supposedly completed.

The work moved through three Bots.

The responsibility never moved with it.

That is the part of Grok Bot I think people will misunderstand. The attraction is obvious. You can create persistent AI workers, give each one a role, connect them to files and tools, schedule repeated jobs, and let them hand work to one another while your laptop is closed.

It sounds like a company.

But companies do not work because everyone has a title. They work because each person knows what they own, what must be handed forward, who can reject weak work, and which decisions must return to the person in charge.

Without that structure, adding more AI workers can create more work for you. One Bot produces information. Another makes it sound convincing. A third approves it. You become the unpaid manager who discovers the mistakes at the end.

The real opportunity with Grok Bot is not building the biggest team.

It is building the smallest team that can complete one useful job without quietly handing the unfinished thinking back to you.

I would start with three roles.

A Scout finds the information and proves where it came from.

A Maker turns that evidence into the report, plan, article, analysis, or other finished work.

A Controller checks whether the result meets the standard and sends weak work back to the person who owns the mistake.

The three Bots share five working files, one definition of finished, and one handoff format. They can research, organize, draft, compare, and review. They stop before sending, publishing, purchasing, deleting, or making another decision that should still belong to you.

This matters even more because Grok Bots on one account share a persistent cloud computer, including files, browser sessions, and logins. That shared environment makes collaboration easier, but it also means one careless connection can give the entire team more access than you intended.

So this article is not about creating three Bots and hoping they behave like employees.

It is about building the missing structure around them.

Below, I am going to build the complete system from the first folder to the first accepted result: the exact Bot roles, files, prompts, handoff contract, approval rules, repair loop, test process, cost tracker, and weekly safety check.

Subscribe now

Read more

You probably do not want another AI subscription. You want to know whether you can cancel one.

That is where I would start with Fable 5 and GPT-5.6 Sol.

It is easy to find reasons to keep both. One gives you an answer you like. The other handles a difficult question differently. Before long, you are pasting the same request into two windows, comparing the results, and wondering which one to trust.

That can be useful for an important decision. It is a tiring way to do everyday work.

I think a comparison should help you settle that decision, not give you another list of features to think about.

If you mainly write, you need to know how much editing remains. If you research, you need sources that support what the answer says. If you code, you need changes that work inside your project, not just an impressive demonstration.

Those are different needs. There is no reason to assume the same model will be your best choice for all three.

There is also a difference between paying for a subscription and paying for each API request. A cheaper token price does not tell you which subscription suits you. A benchmark score does not tell you how many corrections your particular task will need.

My advice is to give each tool a reason to stay.

Start with the work you do every week. Decide what an acceptable result looks like. Then ask whether the second model improves that result enough to justify its cost and the extra switching.

For Fable and Sol, that is the comparison worth making: which one should you open first, and what would make you reach for the other?

Read more

I asked ChatGPT to research a new feature and explain it in 250 words. No fluff. No made-up facts. Just useful information, accurate details, and reliable sources I could verify myself. Normally, that sentence means I expected a long answer in the chat window. Then I would copy it into Word, check the links, rebuild the structure, fix the formatting, and turn the answer into something usable.

This time, ChatGPT did something different.

It researched the feature, opened the official documentation, created the article, built a formatted Word file, rendered every page, found a numbering problem during review, corrected it, and returned the finished document.

I did not ask for a better answer.

I asked for completed work.

That small difference is the entire idea behind ChatGPT Work, a new experience OpenAI introduced in July 2026. It brings the agentic technology behind Codex into everyday ChatGPT, then points it at research, documents, presentations, spreadsheets, websites, connected apps, and recurring business workflows.

Most people will see a new Work button and treat it like another model selector.

That would miss the real shift. Because ChatGPT Work is not just another feature to try. It changes what you can ask ChatGPT to handle for you. In the rest of this article, I break down exactly how it works, how to set it up, what to connect, what to avoid, and how to turn it into a reliable system for real work.

Read more

There is a very good chance you are about to spend money on the wrong AI coding tool.

Not because Codex is bad. Not because Claude Code is better. The problem is that they have changed so fast that the old advice around them already feels outdated. And once you start building your projects, habits, memory and workflows around one of them, switching is not as simple as cancelling a $20 subscription.

Right now, $20 gets you into either side. Claude Pro includes Claude Code. ChatGPT Plus includes Codex. Go higher and both quickly move into the $100–$200 range.

But the monthly price is almost the least important number here.

You are no longer buying an AI that writes a few lines of code. You are choosing something that can enter your project, read hundreds of files, run commands, use tools, remember rules, call other agents, test its own work and keep working while you are doing something else.

And here is what makes the choice harder: Codex and Claude Code are becoming much more alike than they were even a few months ago.

Claude Code now has memory, CLAUDE.md, Skills, MCP, hooks, subagents, background work and isolated worktrees. Codex has memories, AGENTS.md, Skills, plugins, MCP, hooks, custom agents, cloud work, scheduled tasks and worktrees.

So the old “Claude is better at this, Codex is better at that” comparison is becoming less useful.

The actual question now is where your money and more importantly, your time will compound better.

Read more

Nine Doctors, One Family, and One AI.

Jennifer Erickson’s husband, Carl, is a cancer survivor living with a disability. A few days at the Mayo Clinic can put them in front of nine different health care professionals, each bringing tests, recommendations, follow-ups and another piece of the puzzle. Jennifer started doing something smart. After the appointments, she dictates her notes into . It organizes what everyone said, compares treatment options and lays out who is responsible for what comes next. Hours of mentally sorting through information can become a few structured minutes.

But when I read Jennifer’s story, one thought stayed with me:

What if her AI could do that part too?

Browse. Click. Type. Prepare. Then stop when Jennifer’s decision was needed.

And then I realized this isn’t Jennifer’s problem.

It’s ours.

Read more

On August 26, 2026, OpenAI finally released its technical report on one of the most unusual AI security incidents we have seen so far.

The report is long, technical and filled with cybersecurity terminology. So I have broken the entire story down into clear, simple language that anyone can follow.

In this article, you will learn what happened, how the AI agents escaped, why hundreds of them joined the attack, how the intrusion spread across different companies and what must be done to prevent something similar in the future. You do not need a background in AI or cybersecurity. By the end, you will have a clear understanding of the complete incident.

Thank you for reading , where you will find AI by the people, for the people, with simple, smart and clear guides. If you want more independent AI breakdowns like this, you are warmly invited to subscribe.

Here is why this story deserves your attention.

At first, it looked like a sandbox failure. An AI agent had escaped a locked digital workspace that was supposed to keep it away from the wider internet.

Then investigators found the message board.

Roughly 1,200 AI agents that were supposed to work alone had discovered a way to communicate through an internal software storage system. They exchanged more than 70,000 messages and files. They created mailboxes, assigned work, shared credentials, issued HOLD and VETO commands, verified who had sent important messages and preserved discoveries for agents that had not started yet.

About 700 of them went on to participate in an intrusion into Hugging Face.

According to OpenAI, METR and Hugging Face, agents in an internal cybersecurity evaluation escaped intended restrictions, reached the internet and compromised parts of Hugging Face and OpenAI’s research infrastructure.

The obvious lesson is that AI models are becoming dangerous hackers. That is true, but incomplete.

The deeper lesson is that a capable model does not need to become a single rogue superintelligence. Give many copies enough time, tools and shared infrastructure, and they can build something far more operational: an organization.

That is the headline. Now let us reconstruct the incident step by step, beginning with the broken cybersecurity challenge that gave the agents a reason to search for another path.

Read more

Everyone of us already have access to more AI intelligence than they know what to do with. ChatGPT can research. Claude can work through large projects. Gemini can reason across documents. Coding agents can inspect files, write software, use tools, and keep working for hours.

And yet most of us still use AI almost the same way we did three years ago.

Open a box.

Type a prompt.

Get an answer.

Close it.

Tomorrow, open another box and start again. That is the strange thing about the AI boom. We have spent enormous energy asking which model is smartest while giving much less attention to a more useful question:

I do not mean training a 100-billion-parameter model. I mean something far more achievable.

An AI that knows what you are working on. It has access to the right information instead of everything you have ever written. It remembers decisions without preserving every obsolete idea. It knows how you prefer recurring work to be done. It can search, read files, use tools, and complete workflows. It knows which actions it may take alone and which require your approval.

Most importantly, when it fails, the failure does not disappear into yesterday’s chat.

The system changes. That is what we are going to build.

The model is not your AI

• This is the first idea to get right.

• Imagine hiring an extraordinarily intelligent employee on Monday morning.

• You give them no company files.

• No previous decisions.

• No procedures.

• No access to your systems.

• No idea what they are authorized to change.

• No examples of good work.

• No record of previous mistakes.

• Then you ask them to run the business.

• Their intelligence would not be the main problem.

• The missing system would be.

• AI works surprisingly similarly.

Anthropic describes the basic building block of agentic systems as an LLM enhanced with augmentations such as retrieval, tools, and memory. OpenAI describes an agent as three core components working together: the model that reasons, the tools it can call, and the instructions and guardrails that define how it behaves.

More recent work goes further. OpenAI’s engineering team now writes about the harness as the thing worth designing, and warns that a giant instruction file crowds out the task itself, so that too much guidance becomes non-guidance. Anthropic’s agent safety framework names four components that determine how an agent behaves: the model, the harness, the tools, and the environment it operates in.

Three of those four are yours to build.

Read more

There is a strange point in almost every serious AI project. At first, the AI feels brilliant. You explain what you are doing and it understands the audience. You correct a few assumptions and it learns the shape of the project. You make decisions together.

Then you open a fresh conversation.

And somehow you are back at the beginning.

You explain the project again.

You paste the same constraints.

You remind it which option you already rejected.

You tell it that the number in yesterday’s draft was wrong.

You explain why the tone changed.

You reconstruct the work before you can continue the work.

The obvious reaction is:

AI needs more memory.

That distinction changes everything.

A giant conversation is not memory

Inside one long chat, an AI can appear to remember surprisingly well. But much of that continuity comes from context. Earlier messages can be made available again when the model produces its next response. It is less like a person remembering yesterday and more like someone being handed the relevant pages of a notebook before answering.

This matters because context is finite. In Effective context engineering for AI agents, Anthropic frames the work as curating and maintaining the best possible set of tokens during inference, so the model has the strongest chance of producing the behaviour you want. As agents run longer, they generate more messages, tool outputs, plans, files, observations, corrections, and failed attempts.

Simply keeping everything available creates another problem:

noise.

The current plan sits beside three obsolete plans. The approved decision sits beside the option you rejected. Yesterday’s assumption sits beside today’s correction. The useful lesson is buried somewhere inside 40 pages of conversation.

Nothing is technically missing. But the signal is getting harder to find.

This is the first mistake people make with AI memory:

They try to solve forgetting by giving the AI more things to remember.

Eventually, more context stops behaving like better memory.

Even the products are moving toward memory management

ChatGPT already has meaningful built-in memory. Projects can keep related chats, files, and instructions together, giving ongoing work more continuity than starting every conversation from scratch.

Read more

Why most people have already hired their first AI employee? And then they gave that employee every job in the company such as;

• Research this market.

• Compare the options.

• Write the article.

• Plan the project.

• Check the facts.

• Fix your own mistakes.

• Remember what we decided last week.

• Monitor what changes.

• Tell me what to do next.

It sounds ridiculous when you describe it as a job description. Yet that is how millions of us use AI. One model. One conversation. Every responsibility.

Read more

Something strange is happening in how most people use ChatGPT.

The model keeps getting better.

The workflow does not.

You open a blank chat.

Explain the project.

Describe the audience.

Upload the document.

Restate the decision you made yesterday.

Explain what ChatGPT should not do.

Then, finally, useful work begins.

Two days later, another blank chat appears and half that context gets rebuilt.

For the first few years of generative AI, the advantage went to whoever could write the better prompt.

That is no longer the main bottleneck.

The bottleneck is continuity.

A good prompt can improve one conversation.

A good system can improve the next fifty.

And the pieces for that system are now sitting inside ChatGPT.

The blank-chat tax

Imagine you are deciding whether to launch a consulting service.

Monday, you ask ChatGPT about the market.

Tuesday, you start another conversation about competitors.

Wednesday, you upload notes and compare three offers.

Thursday, another chat drafts the landing page.

Friday, you ask what to do next.

Nothing is necessarily wrong with any individual conversation.

The failure happens between them.

The research chat does not know your final decision criteria.

The writing chat does not know why one offer beat another.

The Friday planning chat does not know what changed on Wednesday.

You are the integration layer.

You are carrying context from one AI conversation to the next in your own head.

That is the tax.

And it grows with the size of the work.

A five-minute question can survive a blank chat.

A five-week project cannot.

ChatGPT changed underneath the workflow

Three current ChatGPT capabilities make this problem much more interesting.

Projects can keep related chats, files and instructions together so long-running work has persistent context. Projects now support default or project-only memory, and the memory setting can be changed later in Project settings. Shared Projects automatically use project-only memory.

ChatGPT Work, introduced in July 2026, is designed for longer, more involved tasks. OpenAI says it can research and analyze information, work across connected apps and files, and produce finished documents, spreadsheets, presentations, reports and Sites.

Scheduled Tasks add time to the system. They can handle reminders, recurring work and monitoring, including checks that notify you when something meaningful changes.

Separately, these look like features.

Together, the architecture changes.

Old model:

Prompt → Answer

Better model:

Context → Research → Decision → Execution → State → Follow-up

That second model is much closer to an operating system.

Six jobs, not one giant prompt

A Personal AI Operating System does not mean uploading your whole life into one enormous Project.

More context is not always better context.

A useful system performs six separate jobs:

Notice what is missing.

There is no universal super-prompt.

Research should not behave like Writing.

A client Project should not silently inherit unrelated personal context.

A monitoring task does not need your complete project history simply to check whether one condition changed.

The goal is not making ChatGPT remember everything.

The goal is putting the right context next to the right job.

There is also one configuration decision worth understanding before you build.

Project-only memory keeps a Project self-contained. Chats can reference conversations inside that Project but not conversations outside it, and outside chats cannot reference conversations from the Project.

That is useful when boundaries matter.

But there is a trade-off:

ChatGPT Work is currently unavailable inside Projects using project-only memory.

So there is no universal best setting.

If isolation matters most, project-only memory may be right.

If you want Work available inside that Project, default memory may be the better design.

Design around the work you need done, not whichever feature launched most recently.

Use the smallest capable mode

Before building anything, save this table.

Do not use Deep Research to rewrite an email.

Do not create a Scheduled Task for something you will check manually once.

Do not start a long Work task when you need a three-sentence answer.

Use the smallest capable mode.

Scheduled Tasks currently cannot run more frequently than once per hour, and active-task limits vary by subscription.

The individual limits matter less than the principle:

Automation should come after you know exactly what deserves repeating.

The part almost every AI workspace gets wrong

Creating four Projects is easy.

The hard part is the seams.

Research finishes with thousands of words, and the decision process has to rediscover what matters.

A decision produces generic pros and cons, but the Project never records why the choice was made.

A work session ends inside twenty chat messages, and Monday-you has to reconstruct Friday-you.

A monitoring task reports everything that happened, so eventually you stop reading it.

Every component can work perfectly while the system still fails.

Because the value of an operating system is not in the parts.

It is in the transitions between them.

Here is the test.

The Monday Reconstruction Test

Close everything Friday afternoon.

Monday morning, open one Project.

Can you resume meaningful work in under three minutes without:

opening several old chats,

re-explaining the project,

reconstructing previous decisions,

or reading an entire conversation from the beginning?

If not, you have folders.

You do not yet have a system.

Below is the layer that fixes it.

Read more

Most tutorials end when ChatGPT produces a beautiful homepage.

That is not a business website.

A business website must help a real customer do something:

• Reserve a table

• Request a roofing estimate

• Compare window-cleaning packages

• Call the business

• Find its location

• Trust its claims

• Receive a response after submitting a form

It also needs a real domain, professional email, working mobile layout, accurate information, security controls and a reliable update process.

This tutorial covers that complete journey.

We will use three fictional businesses:

1. Ember & Olive Kitchen, a neighbourhood restaurant

2. Summit Ridge Roofing, a residential roofing company

3. ClearView Window Care, a window-cleaning service

You do not need to know how to code. You do need to know your business, review what ChatGPT creates and test everything before customers see it.

By the end, you will have a repeatable system for turning a business idea into a professional live website.

First, understand what ChatGPT is doing

In the early GPT-4 era, people could ask ChatGPT to generate HTML, CSS and JavaScript. They still had to organize the files, find hosting, fix bugs and deploy everything themselves.

One experienced developer reported building three projects with ChatGPT producing an estimated 85% to 95% of the code. Those projects still took approximately 30 to 60 hours each.

The time went into:

• Debugging

• Improving the design

• Reading documentation

• Fixing mobile problems

• Connecting APIs

• Configuring hosting

• Adapting the project when requirements changed

Treat those figures as one practitioner’s report rather than a measured study. The proportions will be different for your project. The pattern is the part that holds.

This remains an important lesson.

ChatGPT can reduce typing. It cannot remove judgment.

The newer ChatGPT Sites workflow can create, host, refine and publish supported websites inside ChatGPT. It can also connect a domain the owner already controls. However, you remain responsible for the business facts, design decisions, forms, privacy, testing and final decision to publish. OpenAI’s Sites documentation.

The best mental model is:

ChatGPT builds the first version. You build the truth, trust and customer experience around it.

Before you build, know what you are standing on

ChatGPT Sites opened in public beta on July 9, 2026.

Beta is not a footnote. It is the ground your business address will sit on.

Confirm these five facts before you plan anything around it.

1. It requires a paid plan

Sites is available with ChatGPT Plus, Pro, Business, Enterprise and Edu plans. It is not available on the free tier.

2. Usage limits apply during the beta

Plan-specific usage limits apply across all of your Sites. ChatGPT shows the current limits and notifies you as you approach one.

Reaching a limit can prevent you from creating a Site, adding storage, or keeping a high-usage Site public. You can still edit and manage existing Sites.

Read that again as a business owner. A successful local campaign that finally sends real traffic to your website is exactly the condition that pushes a Site toward a limit.

3. Regional availability is not universal

At launch, public publishing was not available in the European Economic Area, the United Kingdom or Switzerland. Confirm current availability for your country before you commit to this route.

4. Custom domains are conditional

Custom domains are available where supported, and were not available in Enterprise workspaces at launch. Confirm that the option appears in your own Site settings before you buy a domain specifically for this purpose.

5. Data residency is not supported at launch

This covers the deployed Site, its code, its stored data and files, generated artifacts and logs.

If your website collects customer names, addresses or photographs, that fact belongs in your planning and in your privacy notice. Part 17 covers it properly. OpenAI’s Sites documentation.

None of this means you should avoid Sites. It means you should not print five hundred business cards pointing at it without a fallback plan.

What we are building

A useful small-business website has four layers.

Skipping one layer creates problems later.

If you design before defining the customer, the website looks generic.

If you buy the domain before checking the business name, you may buy the wrong address.

If you publish before testing the form, customer requests may disappear.

If you create a website without an update system, its prices, hours and services eventually become unreliable.

We will build these layers in order.

Part 1: Decide what kind of website you need

Before opening ChatGPT Sites, identify the primary action you want visitors to take.

Do not say you want more customers. That is too broad.

Choose one main conversion:

• Call the business

• Request an estimate

• Make a reservation

• Order through an external platform

• Book an appointment

• Visit the location

• Join an email list

A website can contain several actions, but one should be clearly more important.

For our examples:

This decision controls the navigation, homepage order, button text and forms.

Choose your build path

ChatGPT Sites is a strong fit for information-led and lead-generation websites. OpenAI documents support for content-led sites, persistent structured data and file storage, depending on the site’s requirements. Some frameworks, private networks and hosting patterns remain unsupported.

Use this decision table:

Important limitation: ChatGPT Sites must not be used to process payment-card data or enable financial transactions. A restaurant can link to an approved external ordering system, but the Site itself should not collect card information.

The same policy also prohibits using Sites to process Protected Health Information, to target children under 13 or the applicable age of digital consent, to distribute malware, to enable phishing, or to impersonate people or organizations.

That first item matters more than it looks. A clinic, dental practice, physiotherapy studio or counselling service should not treat a Site as a patient intake channel. Send those customers to a compliant booking system instead.

Part 2: Set your budget before choosing tools

Your website can become expensive when you buy disconnected tools without a plan.

Start with three numbers:

1. Setup budget

2. Monthly operating budget

3. Amount you can afford at renewal

Use this prompt:

Act as a small-business website launch planner.

Business type: [BUSINESS]
Location: [CITY AND COUNTRY]
First-year budget: [AMOUNT]
Maximum monthly budget: [AMOUNT]
Technical experience: zero
Primary website objective: [OBJECTIVE]

Create three options:

1. Minimum viable setup
2. Professional setup
3. Growth setup

For each option, estimate:
- Domain registration
- Domain renewal
- Website platform
- Professional email
- Booking, ordering or form services
- Photography
- Logo and business-card preparation
- Optional professional assistance

Separate essential costs from optional costs.
Show both the first-year cost and expected annual renewal cost.
Do not assume promotional prices continue after the first year.

A realistic starting budget

Prices vary by currency, promotion and provider. Check them again before buying.

At the time of writing, Microsoft lists Business Basic without Teams at CAD$7.30 per user per month when paid yearly. The plan includes custom-domain email. Its version with Teams is CAD$9.50 per user per month. Prices exclude applicable taxes and can change. Microsoft 365 Business pricing.

One cost is missing from most budgets: the ChatGPT plan itself. Sites requires a paid plan, so add that subscription to the monthly column rather than treating it as a tool you already happen to own.

Do not buy every tool on day one. Build the core customer journey first.

Part 3: Create the Business Source of Truth

This is the most important document in the project.

ChatGPT cannot know which claims are true unless you supply and verify them.

Create one file containing:

• Legal business name

• Public brand name

• Owner-approved description

• Phone number

• Email address

• Physical address or service area

• Opening hours

• Products or services

• Approved prices

• Booking or ordering links

• Verified licences and insurance

• Real warranties

• Approved testimonials

• Social profiles

• Logo files

• Brand colours

• Claims that must never be invented

Use this interview prompt:

Help me create a Business Source of Truth for a new website.

Ask me one question at a time about:

- Business name
- Location and service area
- Ideal customer
- Main customer problem
- Products or services
- Prices
- Opening hours
- Contact details
- Differentiator
- Licences, insurance and certifications
- Warranties
- Reviews and testimonials
- Booking or ordering process
- Privacy requirements
- Primary website action

Do not invent missing information.

When I cannot answer something, mark it:
OWNER MUST CONFIRM BEFORE PUBLISHING.

At the end, organize the approved information into a document that can
be used as the only factual source for the website.

Save this document.

When you make future website changes, attach it again or keep it with the Site project.

Part 4: Choose the business name and domain

A good name is not simply creative.

It must also be:

• Easy to pronounce

• Easy to spell

• Easy to remember

• Appropriate for the local market

• Different from nearby competitors

• Suitable for a website and email address

• Available for lawful business use

Use ChatGPT to generate possibilities, not to make the final legal decision.

Using the attached Business Source of Truth, generate 20 possible names.

Organize them into:
1. Clear and descriptive
2. Premium and memorable
3. Local and community-focused

Each name should:
- Be easy to pronounce
- Be easy to spell after hearing it once
- Avoid hyphens and numbers
- Work on a business card and vehicle
- Allow the business to expand later

Suggest possible .com and Canadian .ca versions.

Do not claim that any name, domain, trademark or social handle is
available. Give me a verification checklist for the five best names.

Verify the name

Before buying anything:

1. Search the name on Google.

2. Search it in Google Maps.

3. Check your provincial, state or national business registry.

4. Search the applicable trademark database.

5. Check the .com and local country domain.

6. Check major social platforms.

7. Search for similar spelling and sound.

8. Check whether the name could be confused with a competitor.

If you are in Canada, consider both .com and .ca. A .ca domain has Canadian presence requirements.

ChatGPT can help organize the search. It cannot provide a guarantee of legal clearance.

Part 5: Buy the domain without getting trapped by the first-year price

The domain is the address of your website.

For example:

• emberandolive.ca

• summitridgeroofing.ca

• clearviewwindowcare.ca

You purchase it through a domain registrar.

Common choices include:

As of August 19, 2026, Porkbun displays a standard .com price of US$11.08 and includes WHOIS privacy, SSL, DNS management and email forwarding. Prices and availability can change. Porkbun .com pricing

Cloudflare says it charges the registry’s wholesale price without adding a registrar markup. It also includes WHOIS privacy and DNSSEC support. Cloudflare .com domains

Compare the real cost

Do not compare only the promotional headline.

Check:

• Registration price

• Renewal price

• Transfer price

• Currency

• Taxes

• WHOIS privacy

• DNS management

• Two-factor authentication

• Domain lock

• DNSSEC

• Support

• Email add-ons

• Premium-domain designation

A domain that costs $1 during the first year can become expensive at renewal.

Purchase process

1. Create the registrar account yourself.

2. Use a permanent recovery email.

3. Enable two-factor authentication immediately.

4. Search the exact domain.

5. Check the spelling character by character.

6. Confirm the renewal price.

7. Verify that it is not a premium domain.

8. Register it in the business owner’s account.

9. Enable domain lock.

10. Enable automatic renewal.

11. Save the receipt and recovery codes.

12. Record the renewal date.

Never allow a freelancer or marketing company to permanently own your domain. They can be given limited access if necessary. The business should remain the owner.

The domain is also your escape route. Whatever happens to the website platform, an address you control can be pointed somewhere else in an afternoon. That is the single strongest reason to keep it in the owner’s own account.

Part 6: Create the visual identity

The website, logo, business card and social profiles should feel like the same business.

You need:

• Primary logo

• Horizontal logo

• Icon-only logo

• Black-and-white version

• Colour palette

• Heading font

• Body font

• Button style

• Photography direction

• Brand voice

Start with three directions:

Using the Business Source of Truth, create three distinct brand directions.

For each direction, provide:
- Brand idea
- Customer impression
- Logo concept
- Colour palette with HEX codes
- Heading font
- Body font
- Photography style
- Button style
- Icon style
- Writing voice
- What to avoid

The system must work across:
- Website
- Business card
- Invoice
- Social profile
- Email signature
- Vehicle or storefront signage

Do not create a final logo yet.
Explain why each direction fits the target customer.

Select one direction, then request the logo.

Do not judge it only at a large size. Test it:

• In black and white

• At favicon size

• On a phone

• On a business card

• On a dark background

• On a light background

• From a distance

AI can create a useful concept, but verify spelling, originality and commercial suitability before adopting it permanently.

Keep the source files somewhere you control. A logo that exists only inside a chat thread is not an asset you own in any practical sense.

Part 7: Create the business card

Use the same brand system.

The front can contain:

• Logo

• Business name

• Positioning line

The back can contain:

• Person’s name

• Role

• Phone

• Business email

• Domain

• Address or service area

• QR code

The QR code should point to a stable page such as:

• /reserve

• /menu

• /estimate

• /quote

Do not point printed cards to a temporary preview address.

Point the QR code at your own domain rather than at the platform address. Printed cards last for years. Platform URLs do not.

Example prompt:

Create a standard 3.5 × 2-inch business-card design using the approved
brand system.

Front:
- Approved logo
- Business name
- Positioning line

Back:
- Owner name
- Role
- Telephone number
- Professional email
- Domain
- Address or service area
- QR-code placeholder

Keep all important content inside a safe print margin.
Make the telephone number and website easy to read.
Do not invent any contact details.
Prepare separate front and back designs.

Part 8: Build the website in ChatGPT

Open ChatGPT Sites from the ChatGPT menu or go to the Sites area. OpenAI recommends describing the audience, purpose, required behaviour and source information. You can then review, refine and manage the Site. Getting started with ChatGPT Sites.

In practice the workflow starts inside the chat itself. Include the word website in your prompt, or mention @Sites, and ChatGPT begins the Sites workflow. You then return to the Sites view to reopen, refine, configure and share what you built.

Do not begin by asking:

Make me a professional website.

That will usually produce a generic result because professional does not describe a specific customer experience.

Use a staged build.

Stage 1: Structure

Create a website using the attached Business Source of Truth.

Before designing it:
1. Identify the primary visitor
2. Identify the primary conversion
3. Propose the sitemap
4. Propose the homepage section order
5. Identify missing information
6. Identify claims that require verification

Do not build or publish the website yet.

Stage 2: Wireframe

Create a low-detail wireframe for each approved page.

For every section, show:
- Purpose
- Heading
- Information required
- Primary action
- Mobile behaviour

Do not add final colours, photographs or invented copy.

Stage 3: First working version

Build the first working version using the approved sitemap, wireframe,
brand direction and Business Source of Truth.

Requirements:
- Mobile-first layout
- Clear navigation
- Strong colour contrast
- Descriptive headings
- Visible keyboard focus
- Form labels
- Meaningful button text
- Descriptive alternative text for informative images
- No invented reviews, awards or claims
- No publication yet

Use OWNER MUST CONFIRM wherever information is missing.

Every Sites deployment URL is a production deployment. If you want to review a version before making it live, ask ChatGPT to save the version without deploying it. OpenAI’s Sites deployment guidance

A useful instruction is:

Save this as Version 1 for review. Do not deploy it.

A new Site is limited to its owner and workspace administrators until you change its access. Leave it that way while you review. Depending on your account, the sharing options can include selected people, everyone in your workspace, and anyone on the internet where public publishing is enabled.

Stage 4: Decide where form submissions go

This is the stage most tutorials skip, and it is the one that quietly loses customers.

A form that displays is not a form that delivers.

Sites offers two kinds of persistence:

At the time of writing, database storage is capped at 10 GB per Site and object storage has no fixed limit. OpenAI’s Sites limits.

The consequence matters more than the mechanism.

By default, a submission is stored. Stored is not the same as sent.

If nobody opens the stored list, the request is invisible. A roofer who checks a dashboard once a week has effectively lost every urgent lead in between, and will never know it happened.

So decide the delivery path deliberately:

For every form on this Site, state clearly:

1. Where each submission is stored
2. How the owner is notified that a submission arrived
3. What the customer sees after submitting
4. How long submissions are retained
5. Who can view stored submissions
6. What happens if the notification fails

If the Site cannot notify the owner reliably, tell me now and propose an
alternative, such as routing the form to an external form service or to
an email service the business already uses.

Do not describe a notification as working until it has been tested with
a real submission.

Then test it the boring way:

1. Submit the form exactly as a customer would.

2. Confirm the stored record exists.

3. Confirm the notification arrives.

4. Confirm it arrives on the owner’s phone, not only on a desktop.

5. Confirm an uploaded photograph can actually be retrieved.

6. Submit again from a different device and network.

Consent, spam and Canadian rules

If any form collects an email address that will later be used for marketing, Canadian businesses are subject to CASL.

In broad terms that means you need consent, clear identification of the sender, and a working unsubscribe mechanism in every commercial message.

Two practical rules:

• Leave the consent checkbox unticked. A pre-ticked box is not express consent.

• Keep a record of what the customer agreed to and when.

A customer submitting a quote request has consented to be contacted about that quote. That is not the same as consenting to a monthly newsletter.

Practical Build 1: Ember & Olive Kitchen

Ember & Olive is a fictional neighbourhood Mediterranean restaurant.

Business objective

The customer should be able to:

1. Understand the food and atmosphere

2. View the menu

3. Check hours and location

4. Reserve a table

Brand direction

Pages

• Home

• Menu

• About

• Reservations

• Location and hours

• Contact

Homepage layout

1. Navigation

2. Restaurant photograph

3. Value proposition

4. Reserve and View Menu buttons

5. Featured dishes

6. Short story

7. Dining experience

8. Verified reviews

9. Hours, address and map

10. Final reservation section

11. Footer

Restaurant master prompt

Create a polished, mobile-first website for Ember & Olive Kitchen, a
fictional neighbourhood Mediterranean restaurant.

Primary objective:
Encourage qualified visitors to reserve a table.

Secondary objectives:
Help visitors view the menu, find the restaurant and confirm opening
hours.

Audience:
Local professionals, couples and families looking for a warm,
high-quality dinner experience.

Visual system:
- Warm cream background
- Terracotta primary colour
- Olive secondary colour
- Charcoal body text
- Elegant serif headings
- Clean sans-serif body text
- Generous spacing
- Large authentic food photography

Build:
1. Home
2. Menu
3. About
4. Reservations
5. Location and Hours
6. Contact

Homepage sections:
- Hero
- Reserve a Table and View Menu actions
- Three featured dishes
- Restaurant story
- Dining experience
- Verified reviews section
- Location, hours and map
- Final reservation action

Rules:
- Do not invent menu items, prices, dietary labels, awards, reviews,
  opening hours or business history.
- Mark missing information OWNER MUST CONFIRM.
- Do not collect payment-card information.
- Use an approved external reservation and ordering link.
- Keep Reserve a Table visible on mobile.
- Save a review version. Do not deploy.

Add menu items

Prepare the menu in a table:

The entries above are illustrative. A real restaurant must provide and approve its actual menu.

Then use:

Replace the sample menu with the attached approved menu.

Organize items by category.
Display:
- Item name
- Short description
- Price
- Confirmed dietary label

Do not rewrite ingredients or prices.
Do not infer allergens.
Add a notice directing customers to contact the restaurant about severe
allergies.
Do not change any other page.

Restaurant acceptance test

• Is Reserve a Table visible without scrolling?

• Does it open the correct reservation service?

• Are menu prices accurate?

• Are dietary labels verified?

• Does the map open the correct location?

• Are today’s hours accurate?

• Does the phone number work when tapped?

• Is the menu readable on a narrow phone?

• Are holiday hours easy to update?

• Does the ordering link open the correct restaurant?

• Where do contact-form messages land, and who checks them daily?

• Is the reservation account owned by the restaurant rather than by one staff member?

Read more

A ChatGPT answer can have citations and still leave you with the wrong conclusion.

That is the part most people miss.

They ask:

What is the best CRM for my company?

Should I buy this stock?

Which AI tool should I use?

Research this company for me.

ChatGPT searches, reads, summarizes, and returns something polished.

There are sources.

There are headings.

There may even be a comparison table.

It feels like research.

But there is a problem.

A polished answer is still an answer.

Research is a process.

And those are not the same thing.

OpenAI itself acknowledges that hallucinations remain a fundamental challenge for language models. Newer models have reduced them, but confidently wrong statements can still occur. OpenAI’s research argues that models can be pushed toward guessing rather than admitting uncertainty.

There is another problem that searching the web does not automatically solve.

Researchers studying retrieval-augmented language models have found that when retrieved sources conflict, models can struggle to decide which evidence deserves more weight. Source credibility, ambiguity, misinformation, and contradictory documents can all affect the final answer.

So the upgrade is not:

Find a better prompt.

It is:

Stop treating ChatGPT as the researcher and start giving it a research system.

That system is what this guide builds.

ChatGPT has become much better at finding information

The tools are now considerably more capable than the old question-and-answer chatbot.

ChatGPT Search can search the web when current information would help and can return inline citations and a Sources panel.

Deep Research goes further.

OpenAI describes it as a system for planning, researching, and synthesizing complex questions into documented reports. You can let it search the public web, restrict it to particular websites, include uploaded files, and in supported cases use connected apps. It proposes a research plan that you can review before the research begins, and the resulting report includes citations or source links for verification.

Projects add another layer. They can keep chats, files, instructions, and reusable context together for ongoing research instead of making you rebuild the setup in every conversation.

Those are powerful components.

But components are not a methodology.

You can give someone access to Google Scholar, a library, spreadsheets, and a team of assistants and still get poor research if the process is weak.

ChatGPT has the same problem.

The mistake happens before the first search

Imagine you ask:

Research whether my company should move from Microsoft Teams to Slack.

ChatGPT could immediately start collecting:

pricing,

features,

integrations,

security capabilities,

reviews,

market share,

AI functionality,

and customer opinions.

It may produce a beautiful report.

But what decision are you actually making?

Maybe your main problem is employee adoption.

Maybe it is external collaboration.

Maybe compliance matters more than features.

Maybe migration cost is the real constraint.

Maybe the correct answer is not Slack or Teams at all, but fixing the way your existing system is configured.

The quality of the answer depends heavily on what question gets researched.

So the first principle of serious ChatGPT research is:

Do not start by searching. Start by defining what would make the decision correct.

That one change prevents enormous amounts of useless research.

More sources do not automatically mean more confidence

There is another common mistake.

People assume:

One source = weak.

Ten sources = strong.

Not necessarily.

Suppose ten articles all repeat a statistic originally published by one company.

You do not have ten independent pieces of evidence.

You may have one claim echoed ten times.

Or imagine five recent sources contradict three older ones.

Should the majority win?

Not automatically.

The newer sources might be weaker.

The older source might be the original research.

Or the underlying situation may genuinely have changed.

Research is not link counting.

It is evidence evaluation.

That means a good ChatGPT workflow needs to track at least four things:

What is being claimed?

What evidence supports it?

How reliable is that evidence?

What evidence points in the opposite direction?

This is where most casual AI research stops too early.

The answer should be the last thing ChatGPT produces

Most people use this sequence:

Question → Search → Answer

For low-stakes questions, that is fine.

For serious research, use:

Frame → Map → Search → Extract → Challenge → Verify → Decide

Seven stages.

Each one has a different job.

And importantly, ChatGPT is not allowed to jump to the recommendation until the evidence has survived the earlier stages.

The first stages create the investigation.

The middle stages pressure-test what was found.

The final stage turns evidence into a decision.

That is the Research System.

Step 1: Frame the decision

Do not begin with:

Research electric cars.

Start with the decision.

Try:

I am considering buying an electric vehicle in British Columbia within six months.

My priorities are five-year ownership cost, winter range, reliability, charging availability, insurance cost and resale value.

Do not recommend a vehicle yet.

First define the exact questions we need to answer before making a defensible decision.

Identify important constraints, missing information, assumptions and factors that could materially change the recommendation.

The last instruction matters.

You are stopping ChatGPT from solving the problem before it has properly defined it.

Your output should look less like an answer and more like an investigation brief.

You want:

• the decision

• decision criteria

• constraints

• important unknowns

• assumptions

• evidence required

• conditions that could change the outcome

If this stage is weak, everything afterward inherits the weakness.

Step 2: Build the research map

Now convert the problem into research tasks.

Use:

Turn the decision into a research map.

For every major question, provide:

1. The claim or question we need to investigate

2. The evidence required

3. The strongest type of source

4. How recent the information needs to be

5. Whether independent confirmation is necessary

6. What could make the evidence misleading

Prioritize the questions that could most change the final decision.

This prevents random browsing.

For an EV decision, specifications might come directly from manufacturers.

Government incentives should come from government sources.

Insurance costs may require current local quotes or insurer data.

Reliability should not be judged only from manufacturer marketing.

Owner forums can reveal possible failure modes, but anecdotes should not automatically become population-level evidence.

Different claims require different sources.

Your research map should reflect that.

Step 3: Search in passes

Now research.

But do not request one enormous answer.

Use separate passes.

Pass A: Landscape

Search this topic and identify the important players, terminology, current state, major disagreements and strongest primary sources.

Do not make a recommendation.

The goal is orientation.

Pass B: Decision evidence

Now investigate each high-priority question from our research map.

Prefer primary and authoritative sources.

Record publication dates and separate direct evidence from interpretation.

The goal is evidence.

Pass C: Exceptions

Search specifically for credible evidence that challenges the emerging picture.

Look for failures, counterexamples, contradictory data, hidden costs and important minority views.

The goal is resistance.

Pass D: Freshness

Recheck the claims most sensitive to time.

Identify anything that may have changed and replace outdated evidence where stronger current evidence exists.

The goal is currency.

This is where Deep Research becomes especially useful because it is designed for multi-step research across multiple sources rather than quick factual lookup. OpenAI also allows users to review and modify the proposed research plan before it begins.

Do not waste that capability.

Edit the plan.

Tell it what matters.

Step 4: Build the Evidence Matrix

Do not ask for the final report yet.

Ask for this:

Create an Evidence Matrix.

Use these columns:

Claim | Evidence | Source | Source Type | Date | Confidence | Contradicting Evidence | Unknowns

Include only claims that matter to the final decision.

Do not hide disagreement between sources.

Your matrix might contain something like:

ClaimConfidenceWhat weakens itVehicle A has lower five-year ownership costMediumInsurance and depreciation varyVehicle B has better winter rangeHighBased on several cold-weather testsCharging access strongly favours Vehicle AMediumDepends on exact travel patternVehicle B has better reliabilityLowEvidence is limited and conflicting

This is where research becomes inspectable.

You can now see which conclusions are solid and which ones are standing on weak evidence.

That distinction is more useful than another paragraph of polished prose.

Step 5: Make ChatGPT attack its own research

This may be the most important step.

Use:

Assume our emerging conclusion is wrong.

Your job is now to challenge it.

Search for the strongest credible evidence against it.

Look specifically for:

• contradictory studies

• stronger competing explanations

• failed examples

• hidden costs

• outdated assumptions

• selection bias

• conflicts of interest

• misleading statistics

• evidence we may have ignored

Do not defend the existing conclusion.

Tell me which findings, if true, would most seriously weaken our recommendation.

Think of this as a red team for your research.

You are changing ChatGPT’s task from:

Build the case

to:

Break the case.

That matters because retrieved evidence is not always consistent. Research has shown that conflicting information can reduce reliability in retrieval-based LLM systems, especially when source credibility is not handled carefully.

The contradiction step makes disagreement visible before the final decision hides it.

Step 6: Run the verification audit

Now classify the research.

Prompt:

Audit every important conclusion.

Classify each as:

VERIFIED FACT

STRONG INFERENCE

WEAK INFERENCE

UNKNOWN

For every VERIFIED FACT, identify the source that directly supports it.

For every inference, explain the reasoning connecting the evidence to the conclusion.

For every UNKNOWN, explain what additional evidence would be needed.

If several articles repeat the same original source, do not count them as independent confirmation.

Downgrade confidence when evidence is contradictory, outdated or indirect.

This introduces something most AI answers lack:

permission to remain uncertain.

OpenAI’s hallucination research is useful here. It argues that systems should be able to abstain when uncertain rather than being pushed toward confident guessing.

Your research workflow should reward the same behaviour.

Sometimes:

We do not know yet

is the most valuable result.

Step 7: Produce the Decision Brief

Only now does ChatGPT get permission to recommend.

Use:

Create the final Decision Brief using only the research that survived our verification process.

Structure it as:

1. Recommended decision

2. Five findings that matter most

3. Strongest supporting evidence

4. Strongest counterargument

5. Important uncertainties

6. What could change the recommendation

7. Next action

8. Confidence score from 0–100

Explain the confidence score.

Keep verified facts separate from judgement.

Do not make the recommendation sound more certain than the evidence supports.

Now your answer has lineage.

You know where it came from.

You know what weakens it.

You know what is missing.

And you know what new evidence could change it.

That is a decision-ready answer.

Turn the workflow into a permanent ChatGPT Project

If you research repeatedly, build the system once.

Create a Project called something like:

Research OS

Projects can keep relevant chats, uploaded files and project-specific instructions together, making them useful for repeated research workflows.

Add these Project instructions:

This Project is my research workspace.

For substantive research:

1. Define the decision before searching.

2. Break the problem into research questions.

3. Prefer primary and authoritative sources.

4. Record dates for time-sensitive claims.

5. Separate direct evidence from interpretation.

6. Search deliberately for contradictory evidence.

7. Do not treat repeated reporting as independent confirmation.

8. Label uncertainty explicitly.

9. Build an Evidence Matrix before recommending.

10. Distinguish verified facts, strong inferences, weak inferences and unknowns.

11. State what evidence could change the recommendation.

12. Never manufacture certainty merely to give me a cleaner answer.

Then keep relevant reports, PDFs, notes and prior decisions inside the Project.

You are no longer starting from zero.

You are building a research environment.

The one master prompt to save

If you want to run the entire workflow from one starting prompt, use this:

I want you to research the following decision:

[INSERT DECISION OR QUESTION]

Do not immediately answer it.

Run this research process:

1. FRAME
Define the real decision, constraints, assumptions, unknowns and criteria that matter.

2. MAP
Break the decision into research questions. For each one, identify the evidence required, ideal source type and freshness requirement.

3. SEARCH
Research in separate passes: landscape, decision evidence, contradictory evidence and freshness.

Prefer primary and authoritative sources whenever available.

4. EXTRACT
Build an Evidence Matrix:

Claim | Evidence | Source | Source Type | Date | Confidence | Contradicting Evidence | Unknowns

5. CHALLENGE
Assume the emerging conclusion may be wrong. Find the strongest credible evidence against it.

6. VERIFY
Classify important conclusions as VERIFIED FACT, STRONG INFERENCE, WEAK INFERENCE or UNKNOWN.

Do not treat multiple articles repeating one original claim as independent verification.

7. DECIDE
Produce a Decision Brief containing:

• recommendation

• five most important findings

• strongest evidence

• strongest counterargument

• unresolved uncertainties

• what could change the recommendation

• next action

• confidence score

Do not optimize for giving me a confident answer.

Optimize for giving me the most defensible answer the evidence supports.

Save it somewhere you will actually reuse it.

But do not confuse the prompt with the system.

The value comes from the sequence.

Three mistakes that quietly destroy AI research

1. Asking for the recommendation too early

Once you ask ChatGPT to recommend something, the rest of the conversation can start orbiting that conclusion.

Do instead: delay the recommendation until verification is complete.

2. Treating citations as verification

Citations make checking easier.

They do not remove the need to check.

Do instead: inspect the sources behind the claims that could change your decision.

3. Researching everything equally

Not every question matters equally.

Twenty minutes verifying a trivial specification may be worth less than five minutes challenging the assumption driving the entire decision.

Do instead: prioritize evidence by decision impact.

The research habit worth keeping

ChatGPT has made finding information dramatically easier.

That changes the bottleneck.

The hard part is increasingly not:

Can I find information?

It is:

Can I tell which information deserves to influence my decision?

That is why the future of good AI research will not be built around giant prompt libraries.

It will be built around disciplined workflows.

Do not ask ChatGPT to sound certain.

Give it a process that makes uncertainty visible.

Because the goal of research was never to produce the fastest answer.

It was to earn the right to trust one.

Subscribe now

Share