There is a category of AI work I don’t think we talk about nearly enough. Not asking ChatGPT a question, generating an email, or summarizing a document. I mean the ugly, multi-step projects sitting on your to-do list that you know need to get done but could easily consume half a day or more of your valuable time.

I had one of those projects this week. I had 22 recordings from my Operating Founder Office Hours sessions that I wanted organized, summarized, researched, packaged, branded, linked, and ultimately turned into a learning library for my community. Doing all of that manually would have meant hours and hours of opening files, reading transcripts, finding links, researching topics, formatting documents, and organizing folders.

Instead, I gave the project to Claude Cowork and Opus 5. Across roughly four or five major prompts, my agents worked on it for about three and a half hours. I did not spend three and a half hours doing the work. Claude did the work, while I focused my enery on more important tasks and was able to spend time with my family.

While the agents worked, I kept working on other things. I responded to client emails, wrote the outline for a new article, got some reading done answer, thought through other projects, and spent time playing with my daughter. That distinction is becoming incredibly important: the value of AI isn’t simply that something that used to take three hours can now take 20 minutes, it is increasingly that a task can take an AI agent three hours while taking almost none of yours.

Watch Me Build It Here on Loom

I recorded the build as I was doing it so you could see what this kind of work actually looks like. I give Claude a set of instructions, let it work, come back to review the output, decide what the next layer should be, and then send it back to work again. The cycle was essentially direct, delegate, leave, review, redirect.

That is increasingly what agentic work looks like for me. I am still responsible for the outcome, the judgment, and the quality of the final product. But I no longer need to personally perform every action from beginning to end.

Become an Operating Founder and join our weekly live office hours this week!

Become an Operating Founder

The Project Started With 22 Zoom Calls

Every week, I hold Office Hours for the founding members of my Operating Founders community. We discuss how we are building our companies, Operating with AI, sales and go-to-market strategy, creator and content strategy, and whatever problems founders and operators are working through at the time. I record those sessions on Zoom, and over time, I have accumulated 22 recordings filled with useful conversations.

But I didn’t want a folder containing 22 recordings. I wanted someone to be able to arrive months later and immediately understand what happened in each session, watch the relevant recording, and then go deeper through additional research and resources. I wanted to turn an archive into a learning system.

1. How I Built a $200,000 Newsletter Business From a Substack Nobody Was Reading

2. The Five Moves That Turned My Substack Into a $200,000 Business

3. The Operating Project: Most Coaching Ends With Notes. This Ends With a Machine.

4. How to Build Your Own AI Employees: A Simple 2026 Guide

Subscribe now

Prompt One: Retrieve and Organize Everything

The first layer was operational. Using my Zoom connection, I asked Claude to retrieve the recordings and associated material, move everything into Google Drive, create the appropriate folder structure, and apply a consistent naming convention.

Think about the manual alternative: finding recordings, downloading them, creating folders, renaming files, uploading everything, checking dates, and making sure every file is where it belongs. Necessary work, yes. A particularly good use of my time, no.

Prompt Two: Understand 22 Conversations

Next, I asked Claude to work through all 22 transcripts. For every session, I wanted a title, a concise description of what we discussed, the correct date, and a link back to the corresponding recording in Google Drive.

Even if I could manually review, summarize, and link each session in ten minutes, that step alone would have taken more than three and a half hours. Instead, I defined the desired output once, let Claude work through the entire collection, and returned when it was ready for review.

Prompt Three: Make the Existing Work More Valuable

I didn’t simply want summaries. For every session, I wanted my community to have high-quality resources that extended the conversation beyond what we discussed during Office Hours.

So I had Claude use an Elite Research skill I built to analyze the topics in each transcript and identify strong articles, research, media, and reference materials. The output was no longer simply “Here is what John said.” It became, “Here is what we discussed, here are the important ideas, and here are additional experts and resources that can help you explore the subject further.”

Take a look at one of the session note documents here.

Prompt Four: Turn the Work Into Finished Assets

Then I applied another reusable skill I have built around my personal brand. Claude took the transcripts, summaries, and additional research and turned them into 22 professionally formatted research and notes packages—one for every Office Hours session.

This is where the scale of agentic execution becomes easier to appreciate. Formatting documents, structuring sections, inserting research links, applying branding, naming files, exporting PDFs, and organizing the outputs are not one task. They are dozens of small tasks hiding inside a larger project, and those small actions are exactly where so much of our working day disappears.

Prompt Five: Build the Final Experience

The last step was creating a single resource that my community could actually navigate. For each session, I wanted a title, a short description, a link to the recording, a branded visual, and a link to the associated research package, all assembled into a document I could move directly into Substack.

The result was an organized library of 22 sessions rather than a pile of recordings. The overall workflow was straightforward: Capture → Organize → Understand → Enrich → Package → Publish.

Here’s a look at what the final product looks like on Substack:

Become an Operating Founder and join our weekly live office hours this week!

Become an Operating Founder

Your Agent’s Time Does Not Have to Be Your Time

This is the biggest lesson I took from the project. If someone hears that the build took three and a half hours, they might reasonably think that doesn’t sound particularly fast. But those were primarily Claude’s three and a half hours, not mine.

My time was concentrated at the places where judgment mattered: What am I trying to build? Is the output good? What should happen next? What needs to change? Once those decisions were made, I could hand the execution back and continue doing something else.

That changes the productivity question. Instead of asking only, “How much faster can AI help me perform this task?” I increasingly want to ask, “How much of this task requires me at all?”

Start Looking for Three-Hour Agent Jobs

Look through your own week for these projects. Maybe it is reviewing 30 customer calls, researching 40 companies, cleaning up a CRM, turning old webinars into training materials, analyzing proposals for common patterns, organizing a document archive, or converting sales conversations into objection-handling material.

These projects are everywhere, and many of us still calculate whether they are worth doing based on how many hours we would need to spend completing them. That calculation is becoming outdated. The better question is: Can I spend 10 or 20 minutes defining this well enough that an agent can spend the next three or five hours doing it?

You may not be able to delegate 100 percent of the work. That’s fine. Delegate 60 percent, then 80 percent, while keeping yourself at the judgment points and letting the agents own more of the execution.

Find the Work You Should Stop Doing Yourself

The bigger opportunity here is not better prompting. It is developing the ability to look at complicated work, see the complete system, break it into layers, and understand which parts require human judgment and which parts can increasingly be delegated.

Historically, I might not have built this library at all—not because it wasn’t valuable, but because the opportunity cost of spending most of a day doing it would have been too high. AI didn’t simply make the project faster. It made a project that might previously have been uneconomic worth doing.

That is what excites me most about where we are headed. The productivity revolution isn’t going to come from humans simply learning to perform the same work slightly faster with AI sitting beside them. It is going to come from learning how to orchestrate meaningful work that continues while we are doing something else.

My agents worked for three and a half hours on this project.

The important part is that I didn’t have to.

- j -

Join the Operating Founder Office Hours each week to work through the real problems, decisions, and opportunities you are facing alongside other founders and operators. Every session is recorded and added to a growing learning library of summaries, research, and additional resources, so the value of the conversation continues long after the call ends. Come for the live discussion, and leave with a deeper body of knowledge you can keep using.

Become an Operating Founder

John Brewton documents the history and future of operating companies at Operating by John Brewton. He is a graduate of Harvard University and began his career as a PhD student in economics at the University of Chicago. After selling his family’s B2B industrial distribution company in 2021, he has been helping business owners, founders, and investors optimize their operations ever since. He is a career consultant and business operator, asking the question: What is the future of companies?

Subscribe now

Most companies are using AI to solve the least interesting problem with meetings.

They record the call. They generate a transcript. They ask AI for a summary. Someone skims it, closes the window, and the team moves on with their business.

That is useful. But it dramatically underutilizes the opportunity at hand. A meeting transcript is not merely a record of a conversation. It is a structured source of decisions, commitments, customer intelligence, objections, deadlines, requirements, ideas, risks, owners, and institutional knowledge.

In an AI-NAtive company, all of those things can become inputs into other systems.

The transcript can update the CRM. It can create tasks. It can draft the follow-up email. It can write the project brief. It can generate a proposal. It can prepare an executive summary. It can tell someone who missed the meeting exactly what they need to know. Months later, an AI agent can search thousands of previous conversations and answer a question nobody remembers discussing.

That changes the purpose of the transcript entirely.

The transcript should not simply tell us what happened in the meeting. It should reduce the probability that another meeting is required.

That may eventually become one of the most important operating advantages of AI.

A Practical Example: Turning One Transcript Into a Finished Asset

Watch the Loom video here.

I wanted to show one small example of what this looks like in practice.

Every Friday, I host Operating Founder Office Hours for members of my paid Substack community. We spend the session working through questions from operators, builders, solopreneurs, and small-business owners. Those conversations can move across dozens of ideas, resources, frameworks, and problems in a single hour.

Historically, the recording itself would have been the primary asset.

Now the transcript becomes the starting point.

In the Loom above, I show Claude pulling the recording and transcript directly from Zoom and creating a comprehensive, anonymized set of session notes. I then have it run my Elite Research skill across the conversation, identifying the concepts, books, academic research, articles, and thought leadership connected to what we discussed.

That produces an expanded research library members can use to keep learning long after the call ends.

Finally, I use my JB Personal Brand skill to turn everything into a polished PDF that can be distributed alongside the recording.

One conversation becomes a summary, a knowledge resource, a research library, a learning tool, and a finished community asset without requiring me to manually reconstruct the session.

Members can then feed the document back into Claude or ChatGPT, ask which resources are most relevant to them, explore an individual topic, or have an extended conversation about the research.

Record it once. Structure it once.

Reuse the intelligence many times.

The same principle becomes much more powerful when applied across an entire company. Here is a sample of the weekly session notes that all Operating Founders receive from our weekly calls:

Subscribe now

What You’ll Learn in This Article

• Why is the meeting transcript becoming an operating layer rather than just a record of what happened?

• How transcripts can automatically improve CRM quality by capturing decisions, objections, commitments, stakeholders, and next steps.

• How AI can turn meetings into execution by creating tasks, assigning owners, identifying deadlines, and surfacing blockers.

• How better transcript workflows can eliminate future meetings, especially recap, handoff, status, and information-sharing calls.

• How one conversation can become an asset factory, producing follow-ups, briefs, proposals, project plans, and internal documentation.

• Why thousands of transcripts can become proprietary institutional intelligence that teams and AI agents can search and learn from.

• The new KPI for meetings: not how well they were summarized, but how much work they created—and how many future meetings they made unnecessary.

What You’ll Get With This Article

• The Meeting Transcript Operating System: Paid subscribers get a complete interactive workbook designed to be uploaded directly into Claude or ChatGPT. Bring one real meeting transcript, and the workbook walks you through building a repeatable system for extracting decisions and commitments, routing intelligence into your CRM and other tools, creating high-value assets automatically, identifying meetings that can be eliminated, and turning the finished workflow into reusable AI instructions, skills, scheduled tasks, and agents. You leave with a working Transcript-to-System Map, CRM Extraction Schema, Asset Factory, Meeting Elimination Report, Transcript Agent Instruction Set, Transcript Value Scorecard, and a 30-day implementation plan. This is not simply another set of prompts.

Become a Paid Subscriber

Read more

The best AI model in one financial-modeling benchmark edits 89.69 percent of the correct cells.

It gets only 34 percent of the final answers right.

That gap explains the problem with most AI spreadsheet workflows.

The model can make a lot of plausible-looking changes. It can move quickly. It can even appear productive. But none of that guarantees that the final answer is correct, traceable, or safe to use.

That is why I spent this morning building a 17-page spreadsheet operating system instead of writing another “better” prompt.

I did not build one all-purpose spreadsheet agent.

I built a five-role team.

Here is how I got there.

Watch my live loom recording of the process here.

First, Understand What the Workbook Is

Before I explain the research or introduce the five roles, it is worth explaining what I mean by a workbook.

This is not a workbook in the traditional sense.

It is not a document you print, complete by hand, and place in a binder. It is also not simply a collection of prompts.

The workbook is a system of instructions that the AI reads and then uses to interview you. You upload it into a new working session and ask the model to walk you through it. The model follows the rules inside the workbook, asks you a structured series of questions, and collects the information it needs to build a working AI team around the way you actually operate.

That distinction matters.

Most generic AI workflows begin with a finished prompt. Someone else decides how the work should be done, packages those assumptions into a template, and hands it to you.

The workbook begins one step earlier.

It helps the model learn how you need the work to be done before it starts building anything. For spreadsheet work, that means collecting details such as:

• The types of analysis you perform

• The workbooks and files you regularly use

• The outputs you need to produce

• The formulas and modeling standards you expect

• The software and data sources involved

• The assumptions that must always be stated

• The controls required before work can be published

• The recurring tasks that may eventually be scheduled

The result is not one enormous prompt that tries to anticipate every possible spreadsheet request.

The result is the infrastructure needed to handle many different requests consistently.

Why the Workbook Is Built Around Questions

The quality of an AI system depends heavily on the quality of the context behind it.

But most people do not naturally provide that context in one clean, complete prompt.

They leave out details because those details feel obvious to them. They forget to explain how a file is organized, which source should be treated as authoritative, what level of precision is acceptable, or what must happen before an answer can be shared.

Those omissions become failures later.

The workbook is designed to prevent that by turning system design into a guided conversation.

Instead of asking you to sit down and write a complete operating manual from scratch, the model asks one question at a time. Your answers become the raw material for the skills, role instructions, rules, checklists, and scheduled tasks that the team will use.

This makes the process more accessible, but it also makes it more rigorous.

The questions force decisions that a loose prompt allows you to avoid.

• What is the team responsible for?

• What is outside its authority?

• Which files must it inspect?

• Which software must it use?

• What must it produce before another role can begin?

• What conditions should stop the workflow?

• How should the final work be checked?

By the end, the model is no longer guessing what “good spreadsheet work” means to you. You have defined it.

Why the Instructions Live in a Workbook

I use a workbook because the system needs to be portable, repeatable, and visible.

A good conversation with an AI can produce excellent work once. But conversations are easy to lose, difficult to audit, and hard to reproduce with another person or in another session.

The workbook gives the process a durable structure.

It contains the questions to ask, the order in which to ask them, the rules that must be followed, and the outputs that should be created from the answers.

That means I can upload the same workbook into a clean session and have the model run the process again without relying on the memory of a previous conversation.

It also means I can give the workbook to someone else.

They do not need to understand every decision I made while designing it. They can upload it, ask the model to follow it, answer the questions, and build a version of the system that reflects their own work.

The workbook is therefore doing two jobs.

First, it captures the operating logic behind the system.

Second, it teaches the user how to build that system for themselves.

Why the Workbook Comes Before the Team

The five roles I built are important, but the roles are not the starting point.

The workbook is.

Before the AI can build an Intake Clerk, Analyst, Modeler, Auditor, or Publisher, it needs to understand what those roles mean inside your environment.

Your Analyst may need to prioritize cash flow, while someone else’s Analyst focuses on inventory or customer acquisition.

Your Modeler may work inside Excel with files stored in SharePoint. Another company may use Google Sheets, a finance platform, and data exported from its ERP.

Your Auditor may need to reconcile every result to the general ledger. Someone else may need to compare it with a prior board report or an external data source.

The role names can remain the same while the actual instructions change substantially.

The workbook collects the information needed to make those roles specific.

It turns a general team design into your team design.

That is why I do not begin by telling the model, “Build me five spreadsheet agents.”

I begin by giving it a structured process for learning what those agents must know, what they are allowed to do, what they must never do, and how their work should move from one role to the next.

The Workbook Is the Control Layer

A useful way to think about the workbook is as the control layer between you and the AI team.

The roles perform the work.

The skills give the roles specialized methods.

The connected software, files, and data give them access to the materials they need.

The scheduled tasks determine when recurring work happens.

But the workbook defines how all of those pieces should be assembled.

It creates the instructions that govern the system before the system begins operating.

That is why I have started building workbooks for more and more types of work. They give people a repeatable way to create new AI capacity without needing me to personally design every team, skill, or workflow for them.

Once you understand how to build one, you can use the same approach for inbox management, research, financial analysis, reporting, project management, or almost any other structured knowledge-work process.

You are no longer collecting isolated prompts.

You are learning how to design the team that will use them.

Now for the process I documented for you today…

Phase 1: Research the failures

Time: 40 minutes

I began with one question:

What actually goes wrong when AI works inside a spreadsheet?

I limited the research to benchmarks and spreadsheet audits. No opinions. No productivity claims. Just documented failure modes.

Three findings shaped the entire system.

1. AI often fails because it does not inspect the workbook properly

SpreadsheetBench 2 evaluates AI agents on multi-sheet workbooks with real formulas, dependencies, and linked calculations.

The best model achieved 34.89 percent overall accuracy.

On debugging tasks, it scored just 12 percent.

The researchers identified two recurring problems:

• Insufficient inspection

• Incorrect target selection

In plain English, the models often failed before they started calculating.

They did not fully understand the workbook, or they changed the wrong cells.

That is not primarily a math problem. It is a looking problem.

2. AI can produce the right number in the wrong way

WorkstreamBench evaluates agents on end-to-end financial tasks.

The best agent scored 69.1 out of 100. Once the task required several chained calculations, the score fell to 53.4.

One of the most important documented failures was deceptively simple: agents sometimes pasted a correct number into a cell where a formula should have been.

The answer may be right today.

But the workbook is now broken.

Change an assumption next week, and the pasted number will not update. A reviewer cannot follow the calculation. The spreadsheet becomes a static answer rather than a working model.

Would you like to work with John?

Click Here

3. Review cannot be treated as an afterthought

Raymond Panko’s field audits found errors in 91 percent of operational spreadsheets.

A single reviewer typically catches about half of the errors.

A second reviewer, working independently and using a checklist, can raise detection to roughly 80 percent.

That matters because most AI spreadsheet workflows ask one agent to inspect the data, choose the method, build the model, calculate the result, and verify its own work.

That is the spreadsheet equivalent of asking an analyst to approve their own analysis.

1. How I Built a $200,000 Newsletter Business From a Substack Nobody Was Reading

2. The Five Moves That Turned My Substack Into a $200,000 Business

3. The Operating Project: Most Coaching Ends With Notes. This Ends With a Machine.

4. How to Produce World-Class Research with Perplexity Computer & Claude -> Fable 5

Subscribe now

Phase 2: Design the team

Time: 20 minutes

The research killed my original idea.

I had planned to build one spreadsheet employee, similar to the inbox employee I taught last month.

But the evidence pointed in another direction.

Inspection, analysis, modeling, and verification are not one job. They are separate jobs with different failure modes.

Combining them inside one large prompt does not create a capable employee. It creates an analyst checking their own work.

So I divided the workflow into five roles.

Each role has:

• One primary responsibility

• One artifact it must hand to the next role

• One action it is not allowed to take

The Intake Clerk

The Intake Clerk profiles the workbook before any analysis begins.

It identifies the sheets, tables, formulas, inputs, outputs, date ranges, missing fields, and potential data-quality problems.

It is not allowed to calculate the answer.

Its job is to understand the environment before anyone starts changing it.

The Analyst

The Analyst decides how the question should be answered.

It defines the method, assumptions, required calculations, and level of precision before producing a result.

The rule is simple:

State the method before stating the number.

The Modeler

The Modeler turns the approved method into a live workbook.

It builds formulas, links assumptions to outputs, and creates a model that recalculates when the inputs change.

It is not allowed to hardcode a value into a computed column.

If a number is calculated, there must be a formula behind it.

The Auditor

The Auditor independently re-derives the headline number using a second method.

It checks the formulas, assumptions, units, ranges, and reconciliation gaps.

It also has the authority to stop the workflow.

If the analysis does not hold, the Auditor issues a hold.

The Publisher

The Publisher prepares the finished workbook and communicates the result.

But it cannot publish over an audit hold.

That rule is what turns the system into a team rather than five prompts pasted one after another.

Every role can reject the work it receives.

Phase 3: Build the operating system

Time: 45 minutes

The finished workbook is 17 pages.

It includes:

• Nine questions that define your analysis standards

• Eight prohibited spreadsheet habits

• Four prompts to run in sequence

• Six reusable prompts for future analyses

The prohibited habits may be the most important part.

They include:

• Pasting values into cells that should contain formulas

• Reporting numbers with no supporting cell reference

• Applying assumptions without stating them

• Presenting more precision than the underlying data supports

• Editing a workbook before profiling its structure

• Publishing an answer that has not been independently reconciled

Each prohibition is paired with a replacement rule.

That distinction matters.

Telling an AI what not to do is rarely enough. You also have to define what it should do instead.

And every rule in the workbook traces back to a documented failure mode. These are not formatting preferences. They are controls.

What this system actually gives you

Most people use AI in spreadsheets because they want speed.

But speed is not the main constraint.

A fast answer with no formula trail still has to be checked by hand. That manual review is the work you were trying to eliminate in the first place.

The real opportunity is not faster analysis.

It is additional analytical capacity.

This system gives an operator access to three functions that many small teams have never been able to staff properly.

A data profiler

Someone examines the workbook before the analysis begins.

They identify what is present, what is missing, what is linked, and what may be unreliable.

A modeler

Someone builds a workbook that continues working when the assumptions change.

The goal is not a spreadsheet that was correct once. It is a model that remains useful.

An auditor

Someone independently checks whether the headline number holds.

This is the role most small finance teams do not have. It is also the role the research suggests matters most.

A private company with a four-person finance team may never have had a second analyst independently re-derive a number before it reached the board.

Historically, that verification could cost as much as the original analysis.

It does not have to anymore.

The final output is not simply a number.

It is:

The number.
The method above it.
The formulas beneath it.
And an independent check beside it.

What to try this week

Choose one analysis you completed manually last month.

Pick something where you already know the correct answer.

Then run it through the five roles in order.

Have the Intake Clerk map the workbook before making any changes.

Have the Analyst explain the method before calculating the result.

Have the Modeler prove that the headline cell contains a formula rather than a pasted value.

Have the Auditor calculate the answer a second way and report the difference between the two results.

Every place where the AI’s work differs from your manual process should become a new rule in your operating profile.

Repeat that process for a week, and the system becomes more consistent.

Then use the Auditor on the spreadsheets your team produced last quarter.

Panko found errors in 91 percent of operational spreadsheets.

There is no reason to assume ours are the exception.

Become a paid subscriber to receive the working and build your Excel team this week!

Subscribe now

And if building this kind of AI fluency is especially appealing to you, and you want to learn everything you can in the year ahead, become an Operating Founder and join my 52-week-per-year Friday live Q&A sessions with The Operating Founder Community. We are a +200-strong group of Operators and Builders choosing a new vision for our futures.

Become An Operating Founder

Download the workbook to build your Excel team below.

Read more

Historically, every time a tool made ideas cheaper to share, the people who had mastered the older, slower way treated it as a threat to defend against rather than a shift to move with. The scribes who spent their lives copying manuscripts did not celebrate the printing press. They saw a craft and a livelihood going to zero. Plato warned that writing itself would ruin memory and hollow out real understanding.

That was around 2,400 years ago.

Civilization grew anyway.

We are running the same fear-based reflex again. Our social feeds and inboxes are full of people litigating whether AI writing counts, whether it is cheating, whether the craft is dying. The scribes and the press are at it again. And while this argument burns, the enabling machines that create distribution for our ideas, work, products, and services are moving faster than they have in 500 years.

Here’s our current reality: The cost of producing words just fell to zero, and the only things left worth owning are the idea you choose and the distance it travels.

1. How I Built a $200,000 Newsletter Business From a Substack Nobody Was Reading

2. The Five Moves That Turned My Substack Into a $200,000 Business

3. The Operating Project: Most Coaching Ends With Notes. This Ends With a Machine.

4. How to Produce World-Class Research with Perplexity Computer & Claude -> Fable 5

Subscribe now

Five thousand years of the same job

For 5,000 years, every writing tool did one thing successfully.

It made ideas cheaper to share.

Clay tablets held temple receipts, not poetry.

A medieval book took weeks of skilled labor and got chained to the desk because it was worth more than the room it sat in.

Then the printing press cut the price of a book by two-thirds within 50 years.

Each drop did the same work.

More people could read, and more people could write.

This cycle of payoffs compounded for centuries.

In 1820, 12 percent of the world could read. Today it is roughly 90 percent. Cheaper copies did not make people read less. This innovation turned reading from an elite privilege into a basic human skill.

Every one of those tools shared a blind spot. The personal computer and the laptop made producing and editing words faster, and none of it touched the oldest cost of all, the effort of composing the thought itself. You still had to sit down and produce the words. That is the cost ChatGPT and Claude just took to zero. These are the first tools that compose and reason back. They draft, they read, they research across sources in minutes instead of weeks.

Subscribe now

What actually got scarce

When a cost falls to zero, value moves to whatever sits next to it and is still expensive.

Composition is now cheap. So look at what it sits between. Upstream is the choice of what is worth saying. Downstream is getting it to the right people and being believed once it lands.

Judgment and distribution.

Those are the survivors.

Everything the AI-writing debate is fighting over is the one output that now fails to be scarce. The question was never how fast you can write. Speed is solved. The new question we should be centering our collective, ambitious, entrepreneurial gaze on is what you choose to think about, and how far you can move it.

The gate is gone

Here is my take on the opportunity we should all be aligned behind, but very few actually are.

For most of the internet’s life, reach was a function of accumulation. You built a list, a following, a set of contacts, and your ideas reached roughly the people you had already collected. That is the social graph. If you were small or new, you stayed small. Distribution was gated by who you already knew.

That gate is gone. A recommendation model now reads the content itself, scores it against what each person actually engages with, and serves it to strangers who have never heard your name. Follower count stopped being the gate and the metric to chase. The quality of your idea and what you can use the new stack to produce, distribute and build with it is the new hurdle, the new gate.

This is not a forecast I’m making as some Substack author you’ve never heard of befroe. Meta told a federal court, in its own antitrust filing, that just 7 percent of the time people spend on Instagram, and 17 percent on Facebook, is spent on content from friends. Zuckerberg has said more than 50 percent of what people see on Instagram is now chosen by AI, not by who they follow. The man who built the social graph testified that Facebook is no longer the culture, and floated wiping everyone’s friend connections to reset the platform around interests. The social network is barely social. The algorithm runs the room.

Subscribe now

So yes, the interest graph removed the gate, but it did not remove the competition. Cheap distribution does not spread attention evenly. Anita Elberse at Harvard studied this directly. In one set of 8 million songs that each sold at least 1 copy, roughly a third sold exactly 1 copy. Attention still stacks on whatever earns it. A model serving your idea to strangers is a momentary shot, not a guarantee, by far.

So the constraint did not disappear. It has just moved to a new place in the supply chain. It used to sit on permission, on access, on the size of your list. Now it sits entirely on the quality and clarity of the idea and how well you shape it. That is a harder game to fake and a fairer one to win. For the first time, a small company with a sharper idea can out-reach an incumbent with a giant list, because the list no longer decides.

Voice is the accelerant

If the interest graph decides who gets heard, voice decides how fast you can feed it.

Voice is the most information-dense form of human communication, and for the first time it is programmable. Andreessen Horowitz calls it the first and primary way people will now interact with AI, and 22 percent of the most recent Y Combinator class were voice-agent companies. You speak, the machine drafts and researches, and the loop from raw thought to finished asset closes to near real time. The cost of running these models fell about 280 times in 18 months, so the constraint is no longer money or minutes. It is what you point them at.

The reading side moved the same way. Most Google searches now end with no click at all. People take the answer, not the source. The idea that the machine surfaces matters more than owning the page it came from.

What to do with this

The uncomfortable part is that none of this rewards the skill most knowledge workers spent 20 years building. Fast, clean production is now table stakes a machine clears for free. What it rewards instead is judgment, and judgment is harder to fake.

So we need to focus our operating energies on the part that pays.

• First, spend your scarce hours choosing ideas, not producing words. Keep a running list of the few ideas you want to be known for. That choice is now the work that pays most.

• Second, move at conversation speed. Talk your thinking out loud, let the model draft and research, then you cut and judge. Do not type what you can speak.

• Third, publish into the interest graph on purpose. Turn each idea into something consumable and teachable, ship it on a schedule, watch what the algorithm carries, and make more of that.

• Fourth, own a few ideas in public until your name is attached to them. Repetition around a tight set of ideas is how you get known for them. Do not scatter.

• Fifth, convert reach into revenue. Distribution is not a vanity metric. The point is to become known, sell more product, and sell more service off the back of the ideas.

Let other people spend the decade arguing about whether how the idea was written or produced matters. Don’t want to give away the ending, but none of that is going to matter one bit. Pick 1 idea worth being known for and publish it every week for the next 52. The writing was never the asset. What you choose to think about, create from those ideas, and how far you can push those concepts into the world of products, serrvices, subscriptions, VALUE FOR YOUR AUDIENCE, is the only thing that matters.

Stop worrying about debates that don’t matter.

Stop letting irrelevant online voices shame you.

Get to work, Substack.

- j -

The Interest-Graph Operating System

I built the system that is the logical, operating approach for this set of ideas. And my paid subscribers get it. The Interest-Graph Operating System is a workbook you paste into your own model (anyone you like or prefer), which then turns into your facilitator and walks you through the four moves: choosing the ideas worth owning, composing them at conversation speed, publishing them into the interest graph, and turning reach into revenue.

You answer its questions in your own words. It compiles the answers into a system you can run this week: your idea ledger, your prompt pack, your publishing plan, and the four key numbers that tell you if it is working.

If you’re just starting to understand these ideas, this workbook is the place to begin building the AI-fluent machine to architect your new professional future and business ambitions.

Become a Paid Operating Subscriber

DOWNLOAD BELOW AFTER THE PAYWALL, FOLLOWING THE APPENDIX AND FAQ SECTION.

Appendix: Sources

Opening and the history of writing tools

• Plato, Phaedrus (Socrates on writing and the loss of memory), Perseus Digital Library, Tufts University. https://www.perseus.tufts.edu/hopper/text?doc=Perseus:text:1999.01.0174:text%3DPhaedrus:page%3D275

• Jeremiah Dittmar, “Information Technology and Economic Change: The Impact of the Printing Press,” Quarterly Journal of Economics, 2011 (book prices fell by roughly two-thirds; cities that adopted the press grew about 60 percent faster). https://academic.oup.com/qje/article-abstract/126/3/1133/1855353

• Jeremiah Dittmar, plain-language summary, CEPR / VoxEU. https://cepr.org/voxeu/columns/information-technology-and-economic-change-impact-printing-press

• “Literacy,” Our World in Data (global literacy 12 percent in 1820 to roughly 90 percent today). https://ourworldindata.org/literacy

The distribution shift: social graph to interest graph

• “Facebook is Dead; Long Live Meta,” Stratechery by Ben Thompson, 2025, citing Meta’s FTC antitrust filing (7 percent of Instagram time and 17 percent of Facebook time spent on content from friends). https://stratechery.com/2025/meta-earnings-meta-turns-the-dial-social-network-r-i-p/

• Mark Zuckerberg, Meta Q1 2024 earnings call, reported by Benzinga (more than 50 percent of Instagram content is now AI-recommended). https://www.benzinga.com/news/24/04/38420928/meta-ceo-mark-zuckerberg-more-than-50-of-the-content-people-see-on-instagram-is-now-ai-recommended

• “The AI behind unconnected content recommendations on Facebook and Instagram,” Meta AI, 2023 (recommendation of content from accounts users do not follow). https://ai.meta.com/blog/ai-unconnected-content-recommendations-facebook-instagram/

• “Spinning off Instagram, the decline of friending, and other takeaways from Mark Zuckerberg at the FTC monopoly trial,” CNN Business, April 16, 2025. https://www.cnn.com/2025/04/16/tech/mark-zuckerberg-testimony-meta-ftc-trial/index.html

• Anita Elberse, Blockbusters, Harvard Business School, via “The Way of the Blockbuster,” Harvard Magazine, 2013 (market concentration; a third of 8 million songs sold exactly 1 copy). https://www.harvardmagazine.com/2013/12/the-way-of-the-blockbuster

Voice and the reading side

• “AI Voice Agents: 2025 Update,” Andreessen Horowitz (voice as the first and primary interface; 22 percent of the latest Y Combinator cohort). https://a16z.com/ai-voice-agents-2025-update/

• “The 2025 AI Index Report,” Stanford Institute for Human-Centered AI (inference cost fell roughly 280x in about 18 months). https://hai.stanford.edu/ai-index/2025-ai-index-report

• “Google zero-click searches study,” Search Engine Land, 2026 (most Google searches end without a click). https://searchengineland.com/google-zero-click-searches-2026-study-479717

Frequently Asked Questions

On “Stop Arguing About AI Writing” and the interest-graph shift

Isn’t most AI writing just slop? Some of it is. Slop is a production problem, and production is the part that got cheap. The point of the piece is that arguing about whether a machine drafts well is the wrong fight, because the value moved to the two things a machine does not do for you: choosing the idea and judging what comes back. Quality still matters. It just stopped being about who typed the words.

Does this mean writing skill no longer matters? Writing skill matters less as production and more as judgment. Fast, clean production is now table stakes a model clears for free. What pays is taste, the ability to choose what is worth saying and to cut what is not. That is a harder skill than typing, and it is the one worth building.

What is the interest graph, in plain terms? The social graph reaches the people you already follow or already know. The interest graph reaches strangers, because a recommendation model reads your content and serves it to whoever is likely to care. Follower count stops being the gate and the idea becomes the gate. Meta told a federal court that only 7 percent of time on Instagram, and 17 percent on Facebook, is now spent on content from friends.

I have a small audience. Can this actually work for me? Yes, and that is the opportunity. The interest graph does not ask how big you are. It asks how good the idea is. A sharper idea from an unknown account can out-reach an incumbent with a large list, because the list no longer decides who gets seen.

If everyone can produce, isn’t the competition infinite? The competition is real, and attention still concentrates on a few winners. Anita Elberse at Harvard showed that digital markets are more concentrated, not less. The gate moved. It did not vanish. Removing the permission barrier does not hand you the crowd. It gives you a shot you did not have before, and the idea is what earns it.

Isn’t building on someone else’s algorithm risky? It is a real risk, and it is the cost of the reach. Manage it two ways. Convert algorithmic reach into assets you own, an email list and paying customers, so a platform change cannot erase you. And do not depend on one platform. Publish where the idea travels and move with it.

Do I need to use voice and AI to do this? No, but they are the accelerant. Voice is the most information-dense way to get an idea out of your head, and the model turns it into a draft you judge and cut. The loop from raw thought to finished asset closes to minutes. You can do the work by typing. You will just be slower.

How do I make money from this, not just collect reach? Attach an offer to every idea you choose to own. Publish the argument free so it travels, and gate the system, the template, or the service that acts on it. Track four numbers every week: reach, email subscribers, paid subscribers, and revenue. If a week moves none of them, change the idea, not the effort.

How is this different from telling me to post more? Volume without a chosen idea trains the algorithm on nothing and trains the reader on nothing. This is the opposite. Own 3 to 5 ideas, repeat them in new forms until your name attaches to them, and let the schedule do the compounding. One flagship a week for 52 weeks beats 12 brilliant pieces shipped on mood.

How much time does the weekly system take? The cadence is one flagship and five short cuts a week. Monday you choose the idea, Tuesday you compose with the prompt pack, Wednesday you ship, and the rest of the week you distribute and log the numbers. The initial build takes about 40 minutes. The upkeep is a few focused hours a week.

What is the Interest-Graph Operating System, and how do I use it? It is a workbook you paste into your own model. The model becomes your facilitator, asks you one question at a time, and compiles your answers into a finished system: your idea ledger, your prompt pack, your publishing plan, and your scoreboard. You build it once, then you run it every week. Subscribers get it with this piece.

Read more

The full methodology and workflow behind a 22-page, 12-chart research report: Perplexity Computer for data collection and essential research, then Claude on Opus 4.8 and Fable 5 for verification and production, along with my elite research skill file that holds the whole thing to an institutional standard. The skill’s architecture is broken down below, and the workbooks for building your own are attached.

Too Long →Didn’t Read

• Perplexity Computer produced a 22-page research report and a full data workbook in 33 minutes across roughly 140 agent tasks. It still contained 5 substantive errors.

• The core doctrine: never publish research from inside the machine that produced it. Move the output to a second, different model and audit it against a written standard.

• The audit runs on a skill file with 6 sections: a closed 3-tier source architecture, 4 research modes, search and paywall discipline, 3 output templates, voice rules, and 8 quality checks.

• Charts get rebuilt from the raw data in your own visual identity. The final report is written in your own voice and read by you in full before anything ships.

• Attached to this post: the full report this article walks through, plus the Your Research Employee workbooks in 4 editions (Claude, ChatGPT, Gemini, Perplexity) so you can build your own analyst in about an hour.

The report runs 22 pages.

It carries 12 original charts, 4 parts, 2 appendices, and a complete source register with a live URL behind every figure.

The machine research behind it took 33 minutes across roughly 140 tasks.

The verification pass that followed caught 5 errors before a single chart was drawn.

The complete process took my team of agents 65 minutes. The work would have taken me two weeks.

Anyone can get an AI research report in 2026. The tools are extraordinary and improving quarter over quarter. Which means the differentiation has moved. World-class research is no longer about access to information. It is about the discipline you wrap around the machines and agentic teams you deploy to make it happen.

This piece walks through the exact pipeline I used to produce “The Last Bubble Went Public. This One Is Staying Private,” my quantitative history comparing the dot-com bubble of 1995 to 2003 against the AI boom of 2022 to 2026. I recorded the whole build on Loom. This is the written version, plus the part the video could not show: the full anatomy of the elite research skill that runs quality control, and the workbooks for building a research employee of your own on Claude, ChatGPT, Gemini, or Perplexity.

Hope this helps you this week, friends!

- j -

If you want to become AI-fluent this summer and learn how to build a team of agents to do the work discussed in this article and so much more, you should seriously consider joining my Oper(AI)tors Summer of AI Fluency cohort.

Grab Your Seat Now!

Use code TAKE500 to receive a $500 discount on your seat.

What is the best AI stack for producing research?

Perplexity Computer collects. Claude verifies and produces the output in my primary stack. Skills encode the standards so neither step depends on my memory or my mood.

Three tools, three jobs, no overlap.

The research ran in Perplexity’s Comet browser (download today!), which is my daily browser on every machine I own. The verification, chart production, and final write-up ran in Claude Cowork sessions on Opus 4.8 and the Fable 5 model. Dictation ran through Wispr Flow, which is why I can commission a research project at the speed of speech.

Step 1: How do you brief an AI research agent like Perplexity Computer?

I opened Perplexity Computer and dictated a brief. Not a query. A brief. The distinction decides the quality of everything downstream.

Watch the Loom here.

The brief had 4 components:

1. The history: Deep research on the dot-com bubble and crash, from the late 1990s through the early 2000s. I want to understand the period, not skim it.

2. The comparison: Set that period against the current AI cycle, quantitatively and qualitatively. Startups then against startups now. Trajectories then against trajectories now.

3. The data demand: Come back with financial data and historical data sufficient to build charts, graphs, and visualizations. This clause is the one most people skip. If you do not ask for chartable series, you get prose, and prose cannot be re-plotted in your own brand.

4. The standard: Exhaustive, attributable, accurate. Every claim is sourced.

Then I left my computer for a bit, and Perplexity deployed a team of agents that worked for 33 minutes across roughly 140 discrete tasks. Watching the full run is instructive, partly because Perplexity shows which models it selects for which activities. But you do not need to watch. You need the 2 deliverables it returns: a complete quantitative and qualitative report, and an Excel workbook holding every underlying data series as chartable tables.

View the Perplexity Excel data file here.

That workbook is the asset. The report is a draft. Hold that hierarchy in your head, because the next step depends on it.

1. How I Built a $200,000 Newsletter Business From a Substack Nobody Was Reading

2. The Five Moves That Turned My Substack Into a $200,000 Business

3. The Operating Project: Most Coaching Ends With Notes. This Ends With a Machine.

4. How to Build Your Own AI Employees: A Simple 2026 Guide

Subscribe now

Step 2: How do you verify AI research for accuracy?

Here is the doctrine at the center of this whole system. Never publish research from inside the machine that produced it.

Perplexity Computer is a serious research engine. The output is attributable, the writing is clean, the coverage is broad. And it still shipped me 5 errors. Not typos. Substantive, number-level errors of the kind that end up in someone’s investor memo and then in someone’s screenshot.

So the first thing I do with any Perplexity research package is take it out of Perplexity. I downloaded both files, opened a Claude Cowork session, attached them, and ran my elite research skill against the full package with one instruction: check that everything here is accurate, and show me what is not.

The audit came back (read the full audit MD file here) with a list.

The 5 that mattered most

1. Cisco’s crash was understated by 22 percentage points: The figure that circulates for Cisco’s dot-com decline is 66 percent. That number measures only through December 2000, which is 22 months before the actual bottom. The true peak-to-trough is 88 percent, from March 2000 to October 2002, from $79 to $9.50. The distinction changes the lesson. Being right about the technology protected Cisco’s business. It did not protect the price.

2. Anthropic’s valuation ladder was mislabeled: The verified sequence reads $61.5 billion, $183 billion, roughly $350 billion, then $965 billion, with the $965 billion Series H dated May 28, 2026. Versions dating that figure to November 2025 conflate it with the Microsoft-Nvidia deal, valued at roughly $350 billion. A 15x repricing in 14 months deserves correct dates.

3. The AI venture funding series stopped a year early: The draft ended the series at 2024, at $114 billion. The full record runs through 2025 at $202.3 billion, close to half of all venture capital deployed worldwide that year. Comparisons ending at 2024 understate the cycle. Flattering the past is the most common failure mode in bubble analysis, and it usually arrives disguised as a data cutoff.

4. The marquee AI IPO count was stale: The correct count is 2. Cerebras was listed on the Nasdaq on May 14, 2026, joining CoreWeave’s March 2025 listing.

5. The NASDAQ year-end close needed pinning: The verified 2025 monthly series puts it at 23,241.99.

I read the list, approved all 5 corrections, and instructed that all work going forward be updated against them. Every downstream chart and paragraph automatically inherited the corrected numbers. The final report states each correction openly in its methodology appendix, which is what a reader deserves and almost no AI-assisted research bothers to do.

The general principle: collect everything in one place, then verify with a different model, ideally a stronger one than the one that did the collection. Two models will not share the same blind spots. One model grading its own homework will.

This is the first layer of checking, not the last. When the full report is assembled, I read the entire thing myself. The machine audit removes the errors a human skims past. The human reader catches what no rule can encode. You need both, in that order.

Step 3: How do you rebrand AI-generated charts in your own visual identity?

The research you publish should look like yours. Perplexity’s charts are competent and generic. So the next instruction to Claude: take my personal brand skill, take every data series from the workbook, and re-create every chart plus any additional visualizations the data supports.

Claude returned 12 charts. Each one is rendered in my visual identity, the Forest Ink field register, as an individual PNG, plus a single PDF holding all 12, plus an HTML index page for reviewing the full set in one scroll. NASDAQ’s 6.2x five-year multiple. The casualty table shows that the survivors still fell to 88 percent. The private-marks ladder no public index can show you. The 280-versus-2 IPO chart encapsulates the entire thesis in a single image.

View all twelve charts here.

The proportion and consistency across the set is the point. A reader cannot articulate why a report feels institutional. Uniform chart grammar is most of the answer.

Two practical notes. First, this only works because the data came back as chartable tables in Step 1. Second, keep the individual PNGs. You will reuse single charts in newsletters, decks, and posts for months. The report is one container for the charts, not the only one.

Step 4: How do you turn AI research into a publishable report?

Final instruction: produce the complete research report, including the charts, written in the Operating voice. Clearly written, professionally written, minimal emotion. Tell the story, walk the reader through the numbers, let the two periods of history do the arguing. Return it as a branded PDF.

It came back as a 22-page document: executive summary, 4 parts, a methodology appendix listing the corrections, and a full source register. Some charts embedded in the first PDF pass imperfectly. This does not bother me. I had Claude produce a Markdown version I could format and embed cleanly myself. Expect this. The last 5 percent of production is still manual, and pretending otherwise is how polished-looking errors ship.

Then the human read. The whole report, my eyes, before anything publishes. The machines compress 2 weeks of analyst work into an afternoon. They do not remove the afternoon.

The finished report. 22 pages, 12 charts, 4 parts, every correction stated in Appendix A, the complete source register in Appendix B.

What goes inside an elite research skill for Claude?

Everything above depends on one file. The elite research skill is a document I have built and tweaked over roughly 6 months, and it is the difference between asking Claude to “check this research” and running an institutional-grade audit with repeatable standards. A skill, mechanically, is just a structured instruction file the model loads when the task calls for it. The leverage is in what the file encodes.

Mine has 6 sections. Here is what each one does and why it exists.

1. What sources should an AI research skill be allowed to cite?

The skill’s foundation is a closed list of approved sources, organized in 3 tiers. Tier 1 is elite publications: the New York Times, the Financial Times, the Wall Street Journal, the Economist, Harvard Business Review, MIT Technology Review, and MIT Sloan Management Review, each listed with its domain and its specific strength. Tier 2 comprises academic research centers: Harvard, Stanford, Princeton, Yale, Chicago Booth, and LSE, each with named search targets, such as specific working paper series. Tier 3 is institutional investment research: Goldman Sachs, Morgan Stanley, Bridgewater, Blackstone, AQR, BlackRock, JP Morgan, plus the top VC firms, a16z, Sequoia, Benchmark, Bessemer, and their peers, with a rule to pull published essays and market maps rather than portfolio announcements.

The operative word is exclusively. The skill instructs the model to cite nothing outside the architecture unless I explicitly ask. This single constraint kills the SEO-farm citation problem before it starts. Most AI research quality problems are source-selection problems wearing a disguise.

Every finding then carries 1 of 4 credibility tags: [PEER-REVIEWED] for Tier 2 academic work, [INSTITUTIONAL] for Tier 3 research, [EDITORIAL] for Tier 1 reporting, and [OPINION] for op-eds and commentary, which must still come from approved sources. I can scan a brief and weigh every claim by the tag beside it.

2. How does the skill match research depth to the task?

The skill defines 4 modes, each with trigger phrases so the model selects the right one from how I naturally talk, each with its own numbered method.

Thesis Research activates on phrases like “I’m writing about” or “I want to argue that.” Its method: restate the thesis in 1 sentence, find supporting evidence, then deliberately hunt complicating evidence, then historical parallels, then 2 to 3 quotable data points, then any disagreements between sources. The output emphasis is narrative ammunition for an article.

Market Intelligence activates on client names and phrases like “competitive landscape.” It works through recent Tier 1 coverage, institutional research, relevant academic work, then risks and opportunities not yet in mainstream coverage. Output emphasis: decision-ready intelligence for advisory work.

Trend Scanning activates on “what are the smartest people saying about” and similar. It restricts to the last 30 to 90 days, then maps consensus, divergence, and the narrative that is forming but has not crystallized. Output emphasis: compressed, newsletter-ready signal.

Deep Dive activates on “comprehensive research” or “everything on.” Broad search, findings organized by sub-theme, a timeline where the topic has one, the key debates mapped, the gaps named, and a meta-narrative synthesized at the end. This is the 10-plus-source mode, and it is the one the dot-com project ran in.

The modes matter because “research this” is not one task. It is at least 4 tasks with different depths, different recency requirements, and different outputs. Naming them lets the file route the work.

3. How should AI research handle search queries and paywalled sources?

The skill specifies exact search patterns, site: operators per source, working-paper queries for the academic tier, and firm-name patterns for institutional research. It sets a floor of 5 searches per request, more for Deep Dive. It mandates recency within 12 months unless the topic is historical.

It also contains the section I consider non-negotiable: paywall discipline. Most Tier 1 sources are paywalled. The skill instructs the model to attempt the full fetch, and when blocked, to extract only what is actually visible, provide the direct URL for me to open through my subscriptions, and tag the finding [PAYWALLED — link provided]. The controlling rule is on one line: never assume content is behind a paywall. Only report what you can actually see. Fabricated paraphrases of paywalled articles are among the most common and least detectable AI research failures. This section is the immune response.

4. What output formats should AI research produce?

Three fully specified templates, chosen by destination. Format A, the Research Brief, feeds articles: thesis, key findings with tags and full citations, complicating evidence, historical parallels, quotable data points, source disagreements, narrative hooks, and a list of paywalled sources to review. Format B, the Advisory Intelligence Memo, feeds client work: executive summary, market context, a key-intelligence table where every row pairs a finding with its source, credibility tag, and a “so what” implication, then competitive landscape, risks and blind spots, and recommended discussion points for the client call. Format C, the Signal Pack, feeds the newsletter: the signal in 1 paragraph, top insights at 2 sentences each, where the smart money disagrees, the emerging narrative, and a prioritized reading list.

Templates sound bureaucratic until you run without them. Unstructured research output is a wall of prose you excavate. Structured output is a component you install.

5. Why should AI research output never hedge?

The skill enforces the Operating voice on the research output itself. Declarative sentences, no hedging, active voice only, findings stated as facts. “Goldman’s research shows X,” never “Goldman’s research seems to suggest X.” A banned-word list that includes “perhaps,” “arguably,” and “it seems.” The reasoning: a research brief is not a neutral academic document. It is intelligence, written with an operator’s eye for what matters. Hedged findings defer the judgment work to future me, and future me is busy.

6. How do you quality-check AI research before delivering it?

The skill closes with 8 verification items the model must clear before delivering anything: every finding cites a specific approved source, every finding carries a tag, nothing off-architecture slipped in, paywalled items are flagged with URLs, the format matches the destination, the voice complies, source disagreements are surfaced rather than smoothed over, and the output is actionable rather than merely informational.

That 7th check deserves a sentence. Models are consensus machines. Left alone, they average conflicting sources into a smooth, confident, wrong middle. Forcing disagreements to the surface is where the most interesting angles live. Goldman versus Bridgewater is an article. Their average is nothing.

The skill also declares its integrations, it feeds my meeting-intelligence pipeline, hands off to my brand system for visuals, and supplies the source architecture when my contrarian filter needs counter-evidence. Skills compound when they reference each other. A lone skill is a tool. A connected set is a staff.

How do you build your own AI research employee?

You do not need 6 months to get a working version of this. With my Oper(AI)te co-instructor, Wessal Khader, I have built the whole construction process into a set of workbooks called Your Research Employee, attached to this post in 4 editions:

Each workbook is a file that an AI can run. You upload it, paste 1 kickoff prompt, and the AI walks you through the build as a guided interview, 1 question at a time. No fields to fill in by hand. The build takes about an hour across 4 steps:

1. The Research Standard, about 20 minutes: Nine questions that force specificity: the decisions your research actually feeds, the output shape you need, when a fast answer suffices versus a deep dig, which sources you trust on sight, which are noise to you, your domain, and whether you want a recommendation or evidence laid out for your own call. The test of a good standard is that it names a real decision. A brief with no decision behind it is a Wikipedia page.

2. The Source Rules, about 15 minutes: The side nobody teaches. The workbook ships with a measured floor built from published research on how AI research fails, including the Columbia Tow Center finding that 8 AI search engines got the source wrong more than 60 percent of the time across 1,600 queries. On top of that floor sit 9 banned research habits, uncited claims, single-sourcing, stale data stated as current, marketing repeated as fact, and 5 more, each paired with the positive rule that replaces it.

3. Sources and Assembly, about 15 minutes: Your trusted-source list, your internal documents, and your question workflow get tested and assembled into 1 paste-ready analyst profile: standard, rules, sources, mode, output format.

4. The Source Audit, about 10 minutes: A repeatable protocol for scoring any brief against a fixed table. 0 to 1 violations are decision-grade. 2 to 4 means fix the claims and update your rules tonight. 5 or more is a guess, wearing a citation. Each workbook closes with a swipe file of 6 prompts for daily operation, including my favorite, the test drive: make the analyst research a question you already know the answer to, then grade it against reality.

The 4 editions exist because the 4 platforms do this job differently. Claude is the only one that wires your private data in as sources through connectors, and it thinks the longest before answering. ChatGPT produces the strongest long-form synthesis and is the most confident when wrong. Gemini brings the biggest context window and the tightest Google Docs handoff. Perplexity tested most accurately on citations and holds a standing persona most loosely. No edition is best at everything. Pick the edition for the machine you live in, and if you run several, the steps are the same on each.

These are the files we will be using in Oper(AI)te this Summer, starting this Tuesday, the 21st of July

The Claude Edition: The only edition that connects your private data as sources through connectors.

The ChatGPT Edition: Built around Deep Research and Memory.

The Gemini Edition: Built around Deep Research with Google grounding and Workspace.

The Perplexity Edition: Built around Research mode, Labs, and Spaces.

Grab Your Seat Before They're Gone!

What should you do with this workflow this week?

The pipeline, compressed: commission research with a real brief that demands chartable data. Move the output to a second, different model and audit it against a written standard before you trust a number. Rebuild the visuals in your own identity. Produce the final document in your own voice, then read every word yourself before it ships.

This week, run 1 piece of it. Take any research you are currently relying on, drop it into a different model from the one that produced it, and ask for an audit against the workbook’s source rules. Count the violations. The number will tell you whether you have research or a guess with a citation.

And if you want to build the full system live, this is exactly what we teach in Oper(AI)te, the Summer of AI Fluency cohort I am running with Wessal Khader. Six live builds: a research analyst, a voice employee, a meetings employee, a visual brand, wired into the tools you already use. We start Tuesday, July 21. Through the deadline, code TAKE500 takes $500 off, $1,999 instead of $2,499. Details at unfazedfounder.com/operaite.

Grab Your Seat Before They're Gone!

The tools will keep improving without your permission. The standards will not.

Frequently Asked Questions

Can you trust AI research from Perplexity without verifying it?

No. Perplexity Computer is the strongest collection engine I use; every claim arrives attributed, and it still shipped 5 substantive errors in this single report, including a 22-point understatement of Cisco’s crash and a mislabeled valuation ladder for Anthropic. Attribution is not accuracy. Audit everything that feeds a decision or a publication.

Why should you verify AI research with a different model than the one that produced it?

Two models do not share the same blind spots. A model auditing its own output tends to re-approve its own errors, because the same weights that produced the mistake evaluate it. Collect in one place, then verify in another, ideally with a stronger model than did the collection.

Which AI is best for research: Claude, ChatGPT, Gemini, or Perplexity?

They do different jobs. Claude is the only one that reads your private data as sources through connectors, and it thinks the longest before answering. ChatGPT produces the strongest long-form synthesis and is the most confident when wrong. Gemini brings the largest context window and the cleanest Google Workspace handoff. Perplexity tested most accurately on citations. The staff you want runs each job on the machine that does it best.

What is a Claude skill?

A skill is a structured instruction file that the model loads when a task calls for it. Mine encodes an approved source list, 4 research modes, search and paywall rules, 3 output templates, voice rules, and 8 quality checks. The file replaces my memory and my mood as the quality-control layer.

What are the most common errors in AI research?

Wrong citations delivered confidently. The Columbia Tow Center found 8 AI search engines got the source wrong more than 60 percent of the time across 1,600 queries. After that: fabricated links, stale data stated in present tense, single-sourced claims, and vendor marketing repeated as fact. Every one of these is testable, which is why the workbooks ship with a source audit.

How long does this workflow take?

The machine research took 33 minutes. The full pipeline, audit, charts, and write-up included, fit in an afternoon, plus the time it takes you to read the final report in full, which is not optional. Building your own research employee from the attached workbooks takes about 1 hour.

Do you need to know how to code to build this?

No. Every piece of this system is a document and a set of prompts. You upload a workbook, paste 1 prompt, and answer questions in plain language. The most technical act in the entire pipeline is downloading a file from one tool and attaching it to another.

About John

John Brewton documents the history and future of operating companies at Operating by John Brewton. He is a graduate of Harvard University and began his career as a PhD student in economics at the University of Chicago. After selling his family’s B2B industrial distribution company in 2021, he has been helping business owners, founders, and investors optimize their operations ever since. He is a career consultant and business operator, asking the question: What is the future of companies?

Connect with John on LinkedIn

TL → DR

• Every AI model writes in a default style. Researchers can identify which model wrote a text with 99.88 percent precision, even when the model was told to write differently.

• The proof has been very public. The word “delve” rose 14-fold in science papers. “Meticulous” ran at 34.7 times its normal rate in AI-edited text. At least 10 percent of 2024 medical abstracts show AI processing.

• Describing your voice to the AI is not enough. Models drop instructions as you stack them, and every dropped instruction gets replaced with the default style.

• The fix has 2 parts and a compendium of samples. Write a positive profile of what you do. Write a prohibition list that bans the AI’s defaults by name. Attach 3 samples of your real writing. Load it all before every task.

• Each platform needs its own setup. Claude takes the full profile as a Project or saved Skill instructions. ChatGPT caps custom instructions at 1,500 characters per field, so the full profile goes in a Project. Gemini needs the samples placed right before the task.

• The stakes are growing. The AI’s favorite words are now rising 25 to 50 percent a year in human speech, and 80 percent of people who wrote essays with AI help could not quote their own essays.

Want to learn how to make AI actually write in your true voice?

Register for my free masterclass, happening tomorrow.

Grab Your Free Masterclass Seat Here!

In the first quarter of 2024, the word “delve” appeared in scientific papers at roughly 14 times its historical rate. For 2 decades it had idled at about half a mention per 1,000 papers. Then it hit 7.9. Nobody called a meeting about this. No style guide changed, and no editor sent a memo asking the world’s researchers to start “delving.” A machine developed a preference, and several hundred million people carried it into the permanent scientific record, where it sits now like a watermark.

That watermark is the subject of this essay. The machine you write with has a voice of its own, and it is overwriting yours. Until you name that voice and explicitly turn it off, your instructions are negotiating with a fingerprint that was trained long before you arrived.

Most advice about AI writing never touches this. It teaches you to describe what you want, in ever more detail, and assumes the model is a blank instrument waiting for direction. The evidence says otherwise. The model arrives with commitments. You do not have a prompting problem. You have a fingerprint problem, and the fix has 2 jobs where almost everyone does 1.

The live masterclass is tomorrow. Everything in this article gets built in the room, in 1 hour: the 2-sided voice profile, the prohibition list, and the 3-machine setup for Claude, ChatGPT, and Gemini. You watch me build mine, then you build yours. The session is free.

Grab Your Free Masterclass Seat Here!

Why does all AI writing sound the same?

A team led by Dmitry Kobak analyzed more than 14 million biomedical abstracts and found that at least 10 percent of 2024 abstracts had been processed with a language model. In some subfields, the estimate reached 30 percent. The giveaway sat in 329 “excess words” that spiked past their projected baselines, and the composition of that list is the telling part. 66 percent were verbs, 18 percent adjectives. Vocabulary of style rather than substance.

A separate study of conference peer reviews measured the same signature from a different angle. In the AI-modified text, the adjective “commendable” appeared at 9.8 times its expected rate. “Intricate” ran at 11.2 times. “Meticulous” ran at 34.7 times. At EMNLP 2023, an estimated 16.9 percent of review sentences had been substantially modified by a model. The reviewers judging the field’s science were, in 1 of 6 sentences, letting the machine speak for them.

“Meticulous” at 34.7 times baseline, “delve” at 14, “intricate” at 11.2, “commendable” at 9.8. At least 10 percent of 2024 biomedical abstracts show LLM processing.

Subscribe now

The issue reaches well past academia. Stanford researchers tracked it across the wider economy through late 2024 and found it in 17.7 percent of consumer complaints, 23.8 percent of corporate press releases, 13.7 percent of United Nations press releases, and roughly 1 in 10 job postings.

A quarter of corporate communication now arrives pre-inflected. Somewhere in a training pipeline, a preference formed the way an accent forms, absorbed rather than chosen, and it is now the most widely spoken accent on earth.

How do researchers detect AI writing?

Catching an author by counting words predates the AI era by 60 years. In 1964, the statisticians Frederick Mosteller and David Wallace settled a 175-year dispute over who wrote 12 unclaimed Federalist Papers. They ignored the arguments and counted the filler. Madison wrote “whilst” where Hamilton wrote “while.” Hamilton reached for “upon” at rates Madison never approached. The invisible words convicted him. All 12 papers went to Madison, and forensic stylometry became a discipline.

The same method now works on machines, and it has become far more effective. A 2025 study trained classifiers on 200,000 texts from 4 model families and identified which model wrote a given passage with 0.9988 precision. The authors’ summary sentence deserves to be quoted in full: large language models “have distinct and consistent stylistic fingerprints, even when prompted to write in different writing styles.”

Models are trained toward human approval, and the training compresses them. Researchers have documented that reinforcement learning from human feedback reduces output diversity. The text comes out more similar, regardless of what goes in. A 2025 paper traced the cause to what it called "typicality bias." Human raters systematically reward conventional, expected phrasing, independent of quality, so the model’s probability mass collapses toward its most typical sentences. Anthropic describes the endpoint of this process plainly in its own research. Post-training, in the lab’s words, selects “one particular character from this enormous cast” and places it center stage. The character is stable. It generalizes across models. It is a house voice.

Why does AI ignore my writing instructions?

The standard response is to write better instructions. Describe the voice. Add detail. Be firm. The instruction-following literature explains why this keeps disappointing you.

On IFEval, the benchmark for objectively checkable instructions, a GPT-4 class model failed roughly 1 multi-instruction prompt in 4; these were instructions a script could verify, including word counts and required keywords. A 2025 study stacked constraints and watched compliance sag, with the best frontier models honoring about 68 percent at a density of 500 instructions. The Allen Institute for AI then ran the sharpest version of the test. Models scoring high on familiar constraints collapsed on unfamiliar ones. Claude 4 Sonnet went from 91.3 percent to 42.3. Gemini 2.5 Pro went from 65.4 to 52.3. OpenAI’s o3 went from 95.0 to 69.3.

The same 3 models on seen constraints (IFEval) and unseen constraints (IFBench). Claude 4 Sonnet drops from 91.3 to 42.3. The best frontier accuracy at 500 stacked instructions is 68 percent.

Subscribe now

Your tone rules are exactly the kind of unfamiliar, unverifiable constraint that decays first. When a constraint decays, the model reverts to the trained default. So the more elaborate your description of your voice, the more constraints you have queued for decay, and every decayed constraint is refilled with house style.

There is a second finding that matters even more for the fix we are all seeking. Models exhibit what MIT researchers call affirmation bias. They under-process negation and instead attend to the named items, with performance in the underlying study dropping by nearly 25 percent when statements were negated. The study was conducted on vision-language models. But the direction of the finding aligns with every operator’s experience with text models. A default you have not named stays on. A default you have named weakly also stays on. Suppression has to be explicit. It has to be specific. And the machine needs to hear it more than once, because its reflex is to register the noun and miss the “never.”

1. How I Built a $200,000 Newsletter Business From a Substack Nobody Was Reading

2. The Five Moves That Turned My Substack Into a $200,000 Business

3. The Operating Project: Most Coaching Ends With Notes. This Ends With a Machine.

4. How to Build Your Own AI Employees: A Simple 2026 Guide

Subscribe now

What is a 2-sided voice profile?

The first side of a working voice profile is the positive profile, the description of what you actually do. Your sentence rhythm. Your openings and your closings. Whether you lead with the conclusion or build to it. Your punctuation habits. How your register shifts between a cold email and an internal memo. Most people who take AI writing seriously have built some version of this. It is necessary, but by itself it does almost nothing about the fingerprint because it never mentions the fingerprint.

The second side is the prohibition list, the explicit inventory of what the model must never default to. The negation reflex, “it’s not X, it’s Y,” which the machine deploys with the regularity of a metronome. The rule of three, triplets in every sentence, makes it feel complete. The vocabulary you now know by heart: “delve,” “robust,” “seamless,” “testament,” “pivotal,” “underscore.” The grand contextual opener that clears its throat for a paragraph before saying anything. Signposting. Paragraphs of eerily uniform length. Each of these is a measured, documented default, and each one stays on until it is named and switched off. The complete inventory, with the measured effect sizes, sits in the appendix at the end of this piece.

The 2-sided architecture. Positive profile, prohibition list, and 3 few-shot anchors, loaded together before the task.

You can spend an hour describing your voice in exquisite detail, and the model will still hand you “delve” and a triplet, because you described what you want and never prohibited what it does. Both sides need to load together, or the calibration fails.

Do writing samples work better than instructions?

The third component outperforms both descriptions.

Show the model your actual writing.

The finding is as old as the modern field. The 2020 GPT-3 paper demonstrated that models can perform new tasks from a handful of examples in the prompt, with no retraining required, and quantified the gap. On 1 benchmark, accuracy went from 64.3 percent with instructions alone to 71.2 percent with a few examples. Later work sharpened the reason in a way that matters for voice. When researchers deliberately corrupted the content of in-context examples, performance barely moved. What the model was absorbing was the shape. The distribution of the text, the format, and the rhythm of how input becomes output. Which is to say, examples convey exactly what you cannot fully put into words about your own prose.

Anthropic’s documentation calls examples “one of the most reliable ways to steer Claude’s output format, tone, and structure.” OpenAI’s guidance compresses it to 3 words: “show and tell.”

This should only take 10 to 15 minutes. Pick 3 of your best pieces of writing in different registers: 1 email, 1 memo or post, and 1 longer piece. These are your few-shot anchors, your voice in its native format.

How do I set up a voice profile on ChatGPT, Claude, and Gemini?

Here is where nearly all AI writing advice goes soft. It claims to work on any model, which is precisely why it works well on none of them. The machines differ, and the differences are documented. A classifier separates their prose at 0.9988 precision. The same voice profile, loaded identically, produces 3 different drifts.

Claude honors explicit prohibitions most faithfully, and its failure mode is over-compliance. Strip out enough defaults and it can go flat. The fix is to pair each major prohibition with 1 positive rule. Avoid signposting, but open with strong declaratives. Load the full profile, positive side first, prohibition list second, anchors third, as per Project instructions, so every conversation inherits it.

ChatGPT’s default is agreeable polish. It will sand your fragments and finish your abrupt endings unless you name them as features. Its custom-instructions fields cap at 1,500 characters each, so the play is a condensed rule set there and the full profile with anchors as the first message in a Project, where it persists.

Gemini over-formalizes, an elevation reflex more pronounced than the other 2, and it weights recent context heavily. Put the profile in a Gem’s instructions, place the examples immediately before the task rather than burying them early, and give the prohibition list an explicit register rule. Match the examples, not what reads as professional. The position advice has empirical teeth. Language models reliably privilege the beginning and end of their context window over the middle, and Anthropic’s own docs report that placing the query after long context improves response quality by up to 30 percent.

1 profile, 3 platforms. Claude over-complies, ChatGPT polishes, Gemini formalizes. Each drift has a 1-adjustment fix and a placement rule.

The differences are proof that voice calibration is a technical discipline rather than a mindset, and that platform-agnostic advice stays shallow for that reason.

Is AI changing how people write and speak?

The strangest finding in this literature concerns your speech rather than your drafts.

Researchers at the Max Planck Institute for Human Development analyzed 360,445 academic talks on YouTube and 771,591 podcast episodes, roughly 7.35 billion spoken words. In the 18 months after ChatGPT launched, the words the model favors- “delve,” “realm,” “meticulous,” “underscore,” rose in unscripted human conversation at annual rates of 25 to 50 percent. The patterns stored in the technology “seem to be transmitting back to the human mind.”

Machine-favored words are rising 25 to 50 percent yearly in 7.35 billion words of human speech. 80 percent of AI-assisted writers could not quote their own essays.

The written evidence points the same way, and it is peer-reviewed. A Science Advances study gave 300 writers AI-generated story ideas and had roughly 600 evaluators judge the results. Individual stories got better, especially from the weaker writers. The collection got worse. Stories written with AI help were measurably more similar to one another than stories written without it. Each writer improved into the same writer. And the pull has a direction. A Cornell study of American and Indian writers using AI autocomplete found the suggestions steadily moved Indian participants toward Western styles, which means the default is 1 register, exported, rather than an average of everyone.

The New Yorker’s Kyle Chayka compressed the phenomenon into 5 words. AI, in his phrase, is “a technology of averages.” The MIT study he reported carried the detail that should bother you most. Of the subjects who wrote essays with ChatGPT’s help, 80 percent could not quote from what they had putatively written.

The professional class spent the last 2 years paying monthly subscriptions to sound like each other. The tool was sold as amplification and adopted as ventriloquism, and most users never noticed the transfer because the prose kept arriving fluent and finished. The fluency we were purchasing was cheap. What belongs to the machine is “finished,” and increasingly, the memory of having written at all.

Can AI detectors prove a text was written by AI?

The word-level tells are aggregate signals, not individual proof. AI detectors built on them misclassified 61.22 percent of essays by non-native English speakers as machine-written in a 1 study at Stanford. A single “delve” convicts nobody. The mechanism behind the vocabulary is also unproven. The popular theory blaming annotation workers’ dialects was tested directly and found unsupported. And the fingerprints move. One 2025 analysis found that DeepSeek’s prose was classified as OpenAI’s 74.2 percent of the time, which suggests the house styles can converge as labs train on similar data. The prohibition list you write this year will need an audit next year.

How do I make AI write in my voice this week?

What to do this week…

Pull your last 5 AI-assisted drafts and audit them against the tells in Figure 1.

Count the negation structures, triplets, marker words, and signposts.

This count is your baseline.

Write the positive profile in 1 sitting. Rhythm, openings, escalation, punctuation, argument order, register shifts. Keep it to one page max.

Write the prohibition list from your audit. Name every default you found, specifically. Not “avoid AI-sounding language” but “never use ‘delve,’ never construct ‘it’s not X, it’s Y,’ never signpost.”

Choose your 3 anchors and load all of them together before the task, on whatever machine you use, configured the way that machine requires.

Then track your edits for a week. Every word you swap and every sentence you reorder is data on a rule the profile still lacks. The gap between what the model gave you and what you shipped is your voice, captured in real time.

In the live session, we build both sides in the room, run the same profile through Claude, ChatGPT, and Gemini side by side, and fix each machine’s drift as it appears. Bring the 3 anchors.

Register For Oper(AI)te Live For Free!

What words and patterns give away AI writing? The full prohibition list.

The difference between a writer’s reframe and the machine’s reflex is about choice, and frequency. Use the pattern once, on purpose. Without clear directoin, the machine uses it 6 times a page. The same goes for every entry below. Possession of a single marker creates little impact, and detectors built on these markers falsely flagged 61 percent of essays by non-native English speakers. The tell most always lives in the presence of the same move, arriving again and again.

The markers fall into 2 evidentiary classes:

• The vocabulary is measured, with effect sizes from studies covering more than 14 million documents.

• The structures are documented by detection researchers and linguists, who observe rather than quantify them.

• Load all 4 sections into your AI as explicit prohibitions, then pair each structural prohibition with 1 positive rule from your own writing.

Subscribe now

Which words does AI overuse?

The 15 measured word families. “Delves” grew 6,697 percent, 2020 to 2024.

The core cluster, with effect sizes. Growth figures measure frequency change in scientific writing, from 2020 to 2024 (Juzek and Ward). Fold-increase measures overrepresentation in AI-modified peer reviews (Liang et al.). Frequency ratios measure excess above baseline in 14 million PubMed abstracts (Kobak et al.).

1. delve / delves / delved / delving. “delves” grew 6,697 percent. The flagship marker.

2. underscore/underscores/underscoring. “underscores” grew 904 percent.

3. showcase/showcasing/showcases. “showcasing” grew 1,396 percent.

4. meticulous / meticulously. 34.7-fold increase in AI-modified reviews.

5. intricate / intricacies. 11.2-fold increase. “Intricacies” grew 773 percent.

6. commendable. 9.8-fold increase.

7. boasts. Grew 918 percent.

8. comprehending. Grew 899 percent.

9. surpass/surpasses/surpassing. Grew 667 percent.

10. garnered. Grew 437 percent.

11. emphasizing. Grew 397 percent.

12. realm. Grew 381 percent.

13. groundbreaking. Grew 330 percent.

14. advancements. Grew 278 percent.

15. aligns. Grew 267 percent.

The documented adjective set, from the same studies’ extended lists: notable, versatile, noteworthy, invaluable, pivotal, potent, ingenious, laudable, lucid, cogent, methodical, profound, refreshing, intriguing, admirable, exceptional, remarkable, seamless, holistic, insightful, comprehensive, compelling, unprecedented, cohesive, tangible, multifaceted.

The documented adverb set: notably, aptly, methodically, compellingly, impressively, undoubtedly, undeniably, seamlessly, effortlessly, robustly, thereby, hitherto, herein, additionally, subsequently, crucially, markedly, judiciously.

A composition note: Of the 329 excess words of 2024, 66 percent were verbs and 18 percent were adjectives. The fingerprint lives in how the machine says things. Content words like “potential,” “findings,” and “crucial” also rose and remain weak signals on their own since any writer covering research reaches for them.

Which phrases give away AI writing?

1. “It’s not just X, it’s Y” and every variant.

2. “In today’s fast-paced world.”

3. “In the ever-evolving landscape of.”

4. “Plays a pivotal role in.”

5. “Serves as a testament to.”

6. “Rich tapestry.” “

7. Navigating the complexities of.”

8. “Unlock the potential of.”

9. “At the intersection of.”

10. “In the realm of.”

11. “A delicate balance between.”

12. “It’s important to note that.”

13. “It’s worth noting that.”

14. “Let’s dive in.”

15. “Let’s delve into.”

16. “In conclusion” and “Ultimately” are closing moves.

17. “Whether you’re a beginner or an expert.”

18. “Underscores the importance of.”

19. “Sheds light on.”

20. “The key takeaway.”

21. Stacked hedges like “could potentially.”

22. And every trace of the eager assistant: “Great question.”

Register For Oper(AI)te Live For Free!

Which sentence structures does AI default to?

1. The antithesis construction: “This isn’t X. It’s Y.” Deployed several times per page.

2. The rule of three: Nouns, adjectives, and clauses arriving in threes, in every sentence that wants to feel complete.

3. The participial tail: A clause bolted on with an “-ing” verb: “...showcasing its versatility,” “...underscoring the need for further research.” 5 of the 21 fastest-growing marker words are exactly these participles.

4. The em dash cascade: Several per paragraph, doing work commas and periods should do.

5. Uniform sentence length: Human writing mixes 4-word sentences with 40-word sentences. Machine writing hums at a steady cadence.

6. The balanced concessive: “While challenges remain, the opportunities are significant.” Every tension is neutralized in a single sentence.

7. Elevated formality throughout: No contractions, no slang, no fragments, even where a human would relax.

8. The signposted transition: “Additionally,” “Furthermore,” “Moreover,” “Notably” at the head of paragraphs after paragraphs. Several of these adverbs appear on the lists of measured overuse.

Register For Oper(AI)te Live For Free!

Which document habits give AI away?

1. The throat-clearing open: A paragraph of context before anything is said.

2. Uniform paragraph length: Every paragraph has 3 to 4 sentences; the whole page is a solid gray.

3. The scaffolded essay: First, second, third, finally, whether or not the material has that shape.

4. Bolded lead-in bullets: Lists where every item opens with a bolded phrase and a colon, much like this section, but it’s how I like to construct sections like this, so I do it anyway.

5. The both-sides shuffle: Every claim balanced by its counterclaim, no position taken.

6. Over-explanation: Restating the question, summarizing what was just said, announcing what comes next.

7. The tidy moral: A closing paragraph that resolves everything and lands on an uplifting generality.

8. Completion compulsion: Nothing left open, every thread tied, even when the honest answer is unresolved.

Register For Oper(AI)te Live For Free!

Frequently asked questions

Why does my AI writing sound like AI?

Every model is trained toward the answers human raters prefer, and raters reward conventional phrasing. That training compresses the model into 1 default style, the same vocabulary and sentence shapes for every user. Your instructions sit on top of that default. They soften it. They do not remove it until you ban its parts by name.

What words does AI overuse?

The measured core: delve, underscore, showcase, meticulous, intricate, commendable, boasts, comprehending, surpass, garnered, emphasizing, realm, groundbreaking, advancements, and aligns, plus their inflections. “Delves” grew 6,697 percent in scientific writing between 2020 and 2024. The full list, with phrases and structures, is in the appendix above.

How do I stop ChatGPT, Claude, or Gemini from sounding like AI?

Build a voice profile with 2 sides. Side 1 describes what you do: your rhythm, your openings, your punctuation. Side 2 bans the machine’s defaults by name: the words, the “it’s not X, it’s Y” reflex, the rule of three, the signposting. Attach 3 samples of your real writing. Load all of it before the task, every time.

Do writing samples work better than written instructions?

Yes. The foundational GPT-3 research showed accuracy rising from 64.3 to 71.2 percent on the same task when examples were added. Later work found that examples transfer format and rhythm, the parts of a voice you cannot fully describe. Anthropic calls examples “one of the most reliable ways to steer” tone. OpenAI’s version is “show, and tell.”

Can AI detectors prove a text was written by AI?

No. The word-level tells are reliable across millions of documents and unreliable on any single one. In 1 Stanford study, detectors falsely flagged 61.22 percent of essays by non-native English speakers as AI-written. A single “delve” convicts nobody. The tell is the pattern repeating across a whole body of text.

Is the same voice profile enough for every AI platform?

The words are the same. The setup is different. Claude takes the full profile as Project instructions and needs each ban paired with a positive rule. ChatGPT caps custom instructions at 1,500 characters per field, so the condensed rules go there and the full profile goes in a Project. Gemini weights recent context heavily, so the samples go right before the task.

Is AI changing how people talk?

The early evidence says yes. Max Planck researchers analyzed 7.35 billion words of podcasts and talks and found the machine’s favorite words rising 25 to 50 percent a year in unscripted speech after ChatGPT launched. The study is a preprint, so treat it as a strong early signal rather than settled fact.

How long should a voice profile be?

Short enough to survive. Instruction-following decays as rules stack, so a bloated profile fails faster. Target 1 page for the positive side, a named prohibition list, and 3 writing samples. Fewer, harder rules beat many soft ones.

Build it live, tomorrow.

This article gives you the doctrine. The masterclass gives you the working setup.

Tomorrow, live, we will build every piece from this article in 1 hour. Your positive profile. Your prohibition list, drawn from the measured word lists above. Your 3 few-shot anchors. Then we load the same profile into Claude, ChatGPT, and Gemini side by side and fix each machine’s drift as it appears, so you leave with your voice running on whatever AI you already use.

You do not need a technical background.

You need 3 pieces of your own writing and 1 hour.

The session is free.

Register For Oper(AI)te Live For Free!

Research appendix: every source in this article.

The measured tells

• Kobak, González-Márquez, Horvát and Lause, “Delving into LLM-assisted writing in biomedical publications through excess vocabulary,” Science Advances 2025. 14M+ PubMed abstracts, the 329 excess words, the 10 percent figure. arxiv.org/abs/2406.07016

• Liang et al., “Monitoring AI-Modified Content at Scale,” ICML 2024. The peer-review fold increases: meticulous 34.7, intricate 11.2, commendable 9.8. arxiv.org/abs/2403.07183

• Liang et al., “The Widespread Adoption of Large Language Model-Assisted Writing Across Society,” Stanford HAI / Patterns 2025. The economy-wide adoption rates. arxiv.org/abs/2502.09747

• Juzek and Ward, “Why Does ChatGPT ‘Delve’ So Much?,” COLING 2025. The 21 focal words with growth rates, and the tested-and-rejected annotator theory. arxiv.org/abs/2412.11385

• Shapira, “Delving into ‘delve,’” 2024. The 14-fold rise of the flagship word. pshapira.net/2024/03/31/delving-into-delve

• Gray, “ChatGPT ‘contamination’: estimating the prevalence of LLMs in the scholarly literature,” 2024. The scholarly keyword growth figures. arxiv.org/abs/2403.16887

The fingerprint and its mechanism

• Bitton, Bitton and Nisan, “Detecting Stylistic Fingerprints of Large Language Models,” 2025. The 0.9988 precision result and the DeepSeek convergence finding. arxiv.org/abs/2503.01659

• Kirk et al., “Understanding the Effects of RLHF on LLM Generalisation and Diversity,” ICLR 2024. Post-training reduces output diversity. arxiv.org/abs/2310.06452

• Zhang, Yu, Chong et al., “Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity,” 2025. Typicality bias as the cause. arxiv.org/abs/2510.01171

• Anthropic, “The Assistant Axis: Situating and Stabilizing the Character of Large Language Models,” 2026. The default character, from the lab itself. anthropic.com/research/assistant-axis

• Mosteller and Wallace, “Inference in an Authorship Problem,” Journal of the American Statistical Association, 1963. The Federalist Papers attribution. jstor.org/stable/2283270

Why instructions decay

• Zhou et al., “Instruction-Following Evaluation for Large Language Models” (IFEval), 2023. The 1-in-4 prompt failure figure. arxiv.org/abs/2311.07911

• Jaroslawicz et al., “How Many Instructions Can LLMs Follow at Once?,” 2025. 68 percent at 500 stacked instructions. arxiv.org/abs/2507.11538

• Pyatkin et al., “Generalizing Verifiable Instruction Following” (IFBench), Allen Institute for AI, 2025. The seen-versus-unseen collapse: 91.3 to 42.3. arxiv.org/abs/2507.02833

• Alhamoud et al., MIT, on affirmation bias and negation, 2025. Why prohibitions must be explicit. news.mit.edu

Why examples beat descriptions

• Brown et al., “Language Models are Few-Shot Learners,” NeurIPS 2020. 64.3 to 71.2 percent with examples added. arxiv.org/abs/2005.14165

• Min et al., “Rethinking the Role of Demonstrations,” EMNLP 2022. Examples transfer format and structure, corrupted labels barely matter. arxiv.org/abs/2202.12837

• Liu et al., “Lost in the Middle: How Language Models Use Long Contexts,” TACL 2024. Position in the context window changes what the model uses. arxiv.org/abs/2307.03172

Platform documentation

• Anthropic, multishot prompting guidance. Examples steer tone and structure. platform.claude.com

• Anthropic, prompting best practices. The query-at-the-end 30 percent figure. platform.claude.com

• OpenAI, prompt engineering best practices. “Show, and tell.” help.openai.com

• OpenAI, custom instructions for ChatGPT. The 1,500-character limit. help.openai.com

• Google, tips for creating custom Gems. support.google.com

The spread into human writing and speech

• Yakura et al., Max Planck Institute for Human Development, on LLM-favored words in spoken language, 2025, preprint. 7.35B words, 25 to 50 percent yearly growth. arxiv.org/abs/2409.01754

• Doshi and Hauser, “Generative AI enhances individual creativity but reduces the collective diversity of novel content,” Science Advances 2024. science.org

• Agarwal, Naaman and Vashistha, on AI suggestions homogenizing writing toward Western styles, CHI 2025. arxiv.org/abs/2409.11360

• Chayka, “A.I. Is Homogenizing Our Thoughts,” The New Yorker, 2025. “A technology of averages” and the MIT 80 percent figure. newyorker.com

The limits of detection

• Liang et al., “GPT detectors are biased against non-native English writers,” Patterns 2023. The 61.22 percent false-positive figure. cell.com

About John

John Brewton documents the history and future of operating companies at Operating by John Brewton. He is a graduate of Harvard University and began his career as a PhD student in economics at the University of Chicago. After selling his family’s B2B industrial distribution company in 2021, he has been helping business owners, founders, and investors optimize their operations ever since. He is the founder of 6A East Partners, a research and advisory firm asking the question: What is the future of companies

TL → DR

• Most people run 2 jobs with AI. They ask a question and close the tab.

• I run 25. Each one is a skill you can own, not a prompt you retype.

• The list below is the whole staff, sorted into 5 groups.

• The gap between 2 and 25 is the difference between using AI and running it.

• 11 of the 25 run as autonomous loops, several agents working together with no hand on the wheel.

• You can build your first one this summer, live, in about 90 minutes.

What does an AI-fluent founder or operator actually do all day?

Most people are still using AI like a Google search box. They ask one thing, read the answer, and close the tab. And if the results aren’t strong enough, or seem not to make sense, they walk away and choose to do the work themselves.

That is their first mistake

That is 1 job, done by hand, every time, from scratch.

I run a team of +25 Agentic Employees every week. Not 25 prompts I retype. 25 agents that know my voice, my standards, my thinking, and my work. 25 team members who produce the same high-quality output every time and who adapt to the changing needs of my business each and every day

The gap is no longer between those who use AI and those who do not. The gap is who runs 25 jobs with it, while everyone else runs 2.

Here is the whole entry-level list of the work my team produces each week and the 25 most common first builds I engineer for my clients each month

Each one carries a tag.

• Run means you operate it, one job at a time.

• Scheduled means it runs on a trigger without you.

• Loop means several agents run it end to end, in a loop, checking their own work, with no hand on the wheel.

What are the 25 things operators run with AI right now?

Writing in your voice

1. Draft a client email in 3 versions, one for the client, one for the team, one for the board, each in your tone, off one line. Run

2. Turn a 2-minute voice memo into a finished post in your own written voice. Run

3. Self-check every draft against your banned-phrase list before you read it. Run

4. Research a named prospect, write the cold outreach, then run the whole follow-up sequence on its own. Loop

5. Rewrite a stiff paragraph into your voice in under a minute. Run

Research that comes back sourced

6. Watch your calendar and deliver a sourced brief before every meeting, every claim cited, with no ask. Loop

7. Scan 3 rivals on a cadence, one agent each, and return the positioning gaps you can take. Loop

8. Turn 10 customer interviews into the 3 jobs they hire you for, one reader agent per transcript into one synthesizer. Loop

9. Fan out across a new market and return a sourced read in an afternoon, not a week. Loop

10. Pull a one-page memo on a person before every first call, triggered off the invite. Loop

Meetings that write themselves up

11. Turn rough bullets into a deck outline you would actually send. Run

12. End every call with the notes, the decisions, and the next steps already written, every time, automatically. Scheduled

13. Run the meeting pipeline, from transcript to notes to counterparty research to a relationship memo to a drafted email, on its own. Loop

14. Get the agenda and the pre-read from a 2-line goal. Run

15. Turn a long email thread into a summary and a recommended reply. Scheduled

1. How I Built a $200,000 Newsletter Business From a Substack Nobody Was Reading

2. The Five Moves That Turned My Substack Into a $200,000 Business

3. The Operating Project: Most Coaching Ends With Notes. This Ends With a Machine.

4. How to Build Your Own AI Employees: A Simple 2026 Guide

Subscribe now

Money and deals

16. Normalize EBITDA on a founder’s spreadsheet and flag what will not survive diligence. Run

17. Read a deal room with one agent per section into one screen on whether the business is real. Loop

18. Turn a discovery call into a full proposal and a scope on its own. Loop

19. Build a 13-week cash view from raw numbers each week and name the weeks that get tight. Scheduled

20. Reconcile 2 months of statements and surface the duplicates and the gaps on a schedule. Scheduled

The work that is yours alone

21. Build the one weekly report that eats your Friday so it runs from your inputs each week. Scheduled

22. Stand up a brand voice profile and write everything through it. Run

23. Turn a research report into a finished carousel, research agents to writer to designer, charts and all. Loop

24. Triage your inbox every morning into needs-you, needs-a-reply, needs-nothing before you sit down. Loop

25. Build a custom skill for the one task no template fits. Run

Which of these run without me, completely?

Run, schedule, loop.

11 of the 25 reach the top rung.

Most teams stop at Run. They operate one skill at a time, by hand, and call that AI fluency. They are mistaken.

The ladder has 3 rungs. A skill you run becomes a job that runs on a schedule. A scheduled job becomes a loop that several agents run together, each one doing a piece, checking the work, and handing it to the next, until the result lands on your desk finished.

11 of the 25 above are at the top rung. A research brief that triggers off your calendar and writes itself before the meeting. A meeting pipeline that takes a transcript and returns a relationship memo and a drafted email with no prompt from you. A market read that fans 6 agents across the sources and comes back in the afternoon.

You are not prompting these. You build and schedule them once, and they run in perpetuity.

Caption: Four agents on one loop. The work moves without you.

That is the move most people have not seen yet. Not a faster search box. A staff of agents that runs the loop while you are in the room doing the work only you can do.

Why is this the summer to become AI-fluent?

Count how many of the 25 you run today.

If the number is 2, that is normal. It is also the problem.

Get Your AI-Fluency Score Today!

Waiting until the gap is loud is the costly move. By the time it is obvious, it has already cost you the deal, the raise, or the year.

Oper(AI)te is where you build the whole staff, not one skill: 8 sessions, 4 employees, and the one workflow that defines your job, built start to finish with Wessal Khader and me.

The summer cohort starts July 21, and for the next 48 hours, you can use code EARLY500 to receive a $500 discount on the regular price of $2500. Seats are selling fast; make sure to grab yours today!

Secure Your Seat Today!

How do I build my first one?

You do not need to build all 25 to start. You need to build 1.

Become an Operating Founder. $99, one time.

What comes with it:

• A live build session. We take one skill from this list and build it on screen, start to finish, yours to keep. You leave with a working employee and the method to build the next 24.

• One year in my Skool community, with my Friday noon Q&A every week.

• One 50-minute working session with me, one to one.

Become An Operating Founder

What the $99 unlocks next:

• Your seat in Oper(AI)te at $1,500 for the July 21 cohort, $1,000 off the $2,500 list.

• A place in the affiliate program, 20% net on every seat you sell.

The $99 is the front door. The live build proves the method. Oper(AI)te is where you build the other 24, and the $1,500 rate closes with the July 21 cohort.

The founder rate is open for 14 days. When the 14 days end, the founder door closes. The list above is the menu. The first build is the key to the first item.

Become An Operating Founder

Frequently asked questions

What is the difference between using AI and running it? Using AI is one prompt at a time, from scratch, every time. Running it means you built a setup that already knows your voice and your work and does the same job the same way every time.

Can these really run on their own? 11 of the 25 can. Past the single skill, a job runs on a schedule, then several agents run it as a loop, each doing a piece and checking the work. You set it once and it runs without you.

Do I need to know how to code? No. You answer questions and paste in samples. The build does the rest.

Which AI app should I use? Any of the big ones works. Claude is strong for writing and voice. Perplexity is built for research. ChatGPT and Gemini both work. Pick one and start.

How long does the first one take? About 90 minutes in the live build session. Then you run drafts to make it sharp.

What does $99 get me? Lifetime access to the Operating Founder community, the weekly Friday office hours, a new weekly AI fluency session through the summer, one live build session, the right to join Oper(AI)te at $1,500 for the July 21 cohort, and a place in the affiliate program at $300 a seat.

What happens after the 14 days? The founder door closes. The $99 rate is sold only inside the 14-day window.

Is this safe with my private writing? Your profile lives in your own account. You read every file before you turn it on.

About John

John Brewton documents the history and future of operating companies at Operating by John Brewton. He is a graduate of Harvard University and began his career as a PhD student in economics at the University of Chicago. After selling his family’s B2B industrial distribution company in 2021, he has been helping business owners, founders, and investors optimize their operations ever since. He is the founder of 6A East Partners, a research and advisory firm asking the question: What is the future of companies

TL → DR

• AI fluency is the capacity to see the whole workflow and build it yourself. Not knowing a tool. Operating one.

• Prompt lists are a phrasebook. They run out. Fluency is a grammar. It does not.

• The discipline is 5 moves: Frame, Assign, Brief, Inspect, and Iterate then persist.

• You drill them on small prompts until they are automatic. Then the same 5 moves run multi-agent systems.

• Paid subscribers get the tool: the Rudiments skill for Claude and ChatGPT, plus a one-page chart.

Subscribe now

I live in a bilingual household. My wife is from Mexico; Spanish is our first language at home, and our two-and-a-half-year-old daughter is becoming bilingual day one. She moves between Spanish and English as if it’s all one language. She produces sentences no one taught her. It’s one of the most beautiful experiences of my life.

I studied Spanish for 4 years in high school and Chinese for 4 years in college. 8 years of classes, and I could not have held a conversation in either. I never became fluent. I was memorizing lists for the test and forgetting them by summer. Since meeting my wife, I have worked hard to become bilingual. I’m not quite there yet, but getting very close. What is making me fluent now is my household, my wife, our daughter, our friends, and my extended family. I speak more Spanish than English in my personal life each day.

That is how most people should learn AI, and almost no one does. They collect prompts instead. “10 prompts for marketers.” “The 50 best ChatGPT prompts.” They are building a phrasebook, and they will struggle to produce high quality outputs the first time the work is not on the list.

Nobody has defined AI fluency, so the word floats. People feel behind without knowing what they are behind on. Companies buy licenses and wait for a result that never arrives.

The gap is not access to the tools.

The gap is a definition nobody wrote down or thought to calrify.

What does it actually mean to be fluent?

Linguists test it with 3 markers:

1. A fluent speaker stops translating in their head.

2. They route around the words they do not know instead of freezing.

3. They create instead of reciting.

My daughter is 2 and already does the third one. Most adults with a language app never clear the first.

The same 3 markers tell you a person is fluent with AI. They stop reaching for a saved prompt. They route around what the model cannot do instead of giving up. They compose a workflow for a problem no one has templated.

New to Operating? You Might Appreciate These Popular Articles

1. How I Built a $200,000 Newsletter Business From a Substack Nobody Was Reading

2. The Five Moves That Turned My Substack Into a $200,000 Business

3. The Operating Project: Most Coaching Ends With Notes. This Ends With a Machine.

4. How to Build Your Own AI Employees: A Simple 2026 Guide

Why is working with AI different from using software?

For 40 years, enterprise software has made this skill unnecessary. The ERP told you which screen to open. The CRM told you what to enter and when. The workflow was built into the software, written by a vendor who had already anticipated your situation. You pointed at the phrase.

That arrangement ended in 1 step. The model does not hand you a directional form on a screen; it hands you a monumentally flexible canvas through which you can communicate to build anything, literally.

The cost of building a workflow fell to a sentence you can type. That changed what a tool is. Which changed what a worker does. This is now changing what it means to be a worker. Most people are still asking which prompt to use.

We have spent our careers doing work that computers could not do, yet. Now the computer does the work and asks us to compose and direct its intended output. That is not a smaller job. It is the work of an Operator, and very few companies, much less individuals, have been trained for this task.

Can you become AI fluent without a technical background?

I trained as a drummer at the Berklee College of Music for a year right after high school. I obsessively drilled rudiments, single strokes, double strokes, paradiddles, slowly and cleanly, thousands of times, until my hands moved without thought.

When the rock-and-roll dream didn’t work out, I became an equally obsessive student, ultimately graduating from Harvard with a degree in Economics. The disciplines and habits that helped me master an instrument were the same tools that helped me succeed in academia.

I did it again in business. I spent a decade at a B2B industrial distributor, from the sourcing desk to the chief executive role, and grew it from $27.5M to $75M before selling in 2021. The work computers could not do yet was the work I did by hand for 10 years.

I am doing it a third time now. I am building AI-first, enterprise-grade software using tools Attio, Claude, Supabase, and Resend. I am not an engineer. I am an Operator, a buildeer, a designer, and a founder.

The discipline that produced all 3 outcomes across my life transfers and compounds with each successive iteration.

For AI Fluency, this skill-building repetitive loop has a clear shape, I call it the Five Moves. It is rudimentary, and meant to be practiced over and over again

You do not become AI Fluent by reading about reps. You achieve this new reality by building this capability across your real work, alongside colleagues who are doing the same, and by learning from leaders who have already achieved the outcome.

That is what I built with Wessal Khader. It is called Oper(AI)te, and it runs this summer. 8 live sessions, no theory, and no replays to watch alone. You build a working staff of AI employees that does your real work, across Claude, ChatGPT, Gemini, and Perplexity. Beginner or advanced. No coding ability required. You leave with a team of agentic AI-employees running 24/7.

The summer cohort starts July 21, and seats are limited.

Take Your Seat @ Oper(AI)te!

Build it with us this summer to become AI-Fluent by the time Q3 ends and Q4 starts!

What are the 5 moves of AI fluency?

The five never change. Only the size of the task changes. You drill them on a one-line prompt the way a drummer drills a single-stroke roll, then the same 5 moves run a multi-agent system. Start small on purpose. A weak prompt costs nothing, and you can run the loop 40 times in a morning on real work.

1. Frame: One question: what is the actual outcome, and what does done look like? Small: “summarize this” becomes “pull the 3 decisions and who owns each.” Large: map where the workflow begins, ends, and every touchpoint between. Skip it and you accept the first plausible thing the model returns, because you never defined better.

2. Assign: One question: who does each part, you, the model, a tool, or an agent? Small: decide whether the model is thinking or just formatting a decision you already made. Large: the model’s job, a tool’s job, a sub-agent’s job, your judgment. Skip it and you hand the model work it cannot do, then blame it for guessing.

3. Brief: One question: what do you know that the model does not, and what must it not do? Small: paste the constraint, the audience, and one example of good. Large: persistent context, data sources, guardrails. Skip it and the answer comes back fluent, confident, and generic, because you described a generic situation.

4. Inspect: One question: where is this wrong, thin, or bluffing, measured against move one? Small: read the output as an operator checking work, not a customer accepting a delivery. Large: test points, spot checks, the places a workflow breaks quietly. Skip it and fluent-sounding wrongness ships. This is the most expensive failure in the discipline.

5. Iterate, then persist: One question: what one change improves this, and how do you keep what works? Small: change one variable, run again, notice what changed. Large: save what worked as a prompt, a project, a skill, or an agent, so it runs again without you. Skip it and every task starts from zero, and nothing compounds.

How do you practice AI fluency?

The cost of a mistake on a 2-line prompt is limited. The rep count is everything. Run the loop 40 times a day on real work, and the moves leave your conscious attention. By the time the task is a system with sub-agents across 4 tools, the moves are already in your hands, and your attention is free for the one thing that does not automate: judgment.

The rudiments are rudimentary. That is the point. Mastered, they become the foundation for building things that run without you.

Is there a tool that enables the 5 moves?

I built a skill that runs you through the 5 moves while you work. It is called Rudiments. When you sit down and say “help me draft the board update” or “build me a lead-routing flow,” it does not lurch into the task. It asks you the moves, calibrated to the size of the job, and it shows you the difference between your loose ask and the sharp one, so you learn the move and not just the answer. It is built to make itself unnecessary.

Here is what it does with every task you bring it:

• Sizes the job. A quick rep for a one-line prompt, a full rep for real work, a build rep for a multi-agent system.

• Walks you through the 5 moves as questions: Frame, Assign, Brief, Inspect, and Iterate then persist.

• Shows you your loose ask next to the sharp one, so the rep teaches you the move.

• Tracks the moves you skip and tells you which one to drill next.

• Saves what works, so the rep compounds instead of resetting to zero.

That is the loop and the rep that will help you get one step closer to becoming AI Fluent.

You do not need to write code. You do not need a developer. You need Claude Pro at $20 a month or a Premium ChatGPT account, and 60 minutes of real work to drill against.

Paid subscribers get 3 things, attached below:

• The Rudiments skill for Claude, installed in under a minute.

• The same discipline written as instructions for a ChatGPT, so it runs wherever you work.

• The one-page chart, print-ready.

Become a paid subscriber to download the Claude skill and the directions for how to build the same capability in ChatGPT after the frequently asked questions section below.

Subscribe now

Frequently Asked Questions

What is AI fluency? AI fluency is the capacity to see the whole workflow and build it yourself. It is not knowing a tool. It is operating one. The test is the same as for language. You stop reaching for a saved prompt, you route around what the model cannot do, and you compose a workflow for a problem no one templated.

What is the difference between AI fluency and prompt engineering? Prompt engineering collects lines that work. AI fluency builds the workflow that those lines sit inside. A prompt list is a phrasebook, and it runs out. The 5 moves are grammatical, and they do not.

Do you need to know how to code to become AI fluent? No. I am building an AI-native CRM for a public company, and no one trained me to write software. You compose the workflow in plain language and let the model and tools handle their parts. You need judgment, not a computer science degree.

What are the 5 moves? Frame, Assign, Brief, Inspect, and Iterate then persist. Frame the outcome. Assign each part to you, the model, a tool, or an agent. Brief the context and the guardrails. Inspect the output against the outcome. Iterate one change, then save what works so it runs again without you.

How long does it take to become AI fluent? Faster than you expect, because you drill on small tasks with a high rep count. Run the 5 moves 40 times in a morning on real work, and the moves start to leave your conscious attention in weeks, not years.

Should I use Claude or ChatGPT to practice? Either. The discipline is identical on both. Claude Pro is $20 a month, and ChatGPT has a free tier. The Rudiments skill ships for Claude, and the same discipline is written as instructions for a ChatGPT Custom GPT.

What is the Rudiments skill? It is a tool that runs you through the 5 moves while you work, calibrated to the size of the task, and shows you the difference between your loose ask and the sharp one so you learn the move. It is built to make itself unnecessary. Paid subscribers get it for Claude and ChatGPT, plus the one-page chart.

Download the first skill you need to become AI Fluent below.

Read more

What is this article about? (TL → DR)

• Most people use AI like a search box. They ask one thing, then close the tab.

• A few people do more. They build AI “employees” that do the same work every day.

• An AI employee is a setup you create once. It learns your voice, your job, and your rules.

• Four are worth building first: one for writing, one for research, one for meetings, and one for your own job.

• You can build the first one this week. No code. The free kits and a free live class are at the bottom.

Want to build one live with me, for free?

Register for my free Masterclass on June 16th, where I will teach you how to build your own AI Employee that runs your calendar and inbox! Register Here

A free 60-minute live class with me and . You come in with a full calendar and a packed inbox. You leave with an AI employee that runs both, and the way to build the next 10. No code. No replay.

Reserve Your Free Seat

What is the difference between using AI and operating it?

Using AI is simple. You open it, ask a question, and close it. The answer sounds flat. The tool knows nothing about you.

Operating AI is different. You build something that stays. You teach it your voice, your job, and your rules. Then you run it like a teammate who already knows your standards.

Here is a typical Monday for a person who built their staff:

• A big slide deck is due on Wednesday. It gets drafted in an hour, not over two nights. Three calls run back-to-back.

• Each one ends with the notes, the choices, and the next steps already written.

• A question that used to eat up an entire afternoon comes back with a sourced answer before lunch.

That is not a dream. That is what the right setup does.

What is an AI employee?

An AI employee is not a prompt you copy from Reddit or copy and paste from a LinkedIn post. It is a setup built for one job inside your team and workflows. You engineer it one time. It runs the same way every time. You adjust and tweak, but do not ever need to rebuild.

These are the four most worth building first:

1. The Voice Employee writes like you. You train it on your real emails and posts. It can write to a client, to your team, or to your boss, each in the right tone. It checks its own work before you see it. Done right, no one can tell you used it.

2. The Research Employee gets you ready. You give it the questions you ask every week. It pulls real facts, sorts them, and hands you a short report with sources. It preps you before any meeting that matters.

3. The Meetings Employee kills the blank slide. You give it your rough notes. It hands back a clean deck you would actually send. It writes the agenda. It turns your call notes into a summary. Every meeting ends with the work already done.

4. The Role-Specific Employee is yours alone. You have one task that eats hours it should not. A report. A proposal. A weekly grind. No template fits it. You build this one around your real job.

Connect With Me On LinkedIn

Connect on LinkedIn

Why do most people get AI wrong?

Most people stay stuck for one reason. They think of AI as a tool. Tools are passive. You pick them up, use them, and put them down.

An employee is different. It carries what it knows. It knows your standards. It knows what you would say no to. You do not explain yourself again every time.

Change that one idea, from tool to employee, and everything changes. How you build it. How you talk to it. What you expect to receive back.

Consulting Perspective: This is one of the biggest mental hurdles I see inside my clients’ companies. Individuals struggle to provide enough direction, detail, and specificity when they’re prompting one of the models to perform a taks, or within the build process for the engineering of their agent.

When you make the jump to looking at the AI as a new hire, a talented, smart analyst, and begin providing said analyst with all of the context, reference materials, and clear direction that you would to any team member you were asking to produce high-quality work product, that is when the quality of the results, the outcomes, and the strength of the builds you are engineering, dramatically improve.

The primary apps are not hard to use. Claude is strong in long writing and voice. Perplexity is built for research with sources. ChatGPT does a little of everything very well and is currently the cat’s meow for image generation. Gemini sits within Google, where many people already work, and has a unique capability to help with management and team-dynamics queries. You do not need to master one. You need to know which one to grab, and why.

Are most people actually good at AI?

Here is the hard part. Most people in your meetings are faking it. They put “AI” in their slides. They say they are “trying the tools.” They paste prompts into free accounts and take screenshots of the results.

The gap is getting larger. You can see it in who writes the clearest report the fastest. Who shows up to the meeting with the work done? Who gets more done without longer hours? That person, most likely, quietly built their staff.

The most costly move right now is to wait until it is obvious. By the time the gap is loud, it has already cost you.

How do I start building my AI staff?

You can build your staff this summer. Do it before the calendar fills and the year resets. Do it before a skill gap turns into a career gap.

Start with the Voice Employee. Train it on 10 to 15 real things you have written. Tell it what you never say. Tell it how you sound when you are rushed. Tell it what “clean but not stiff” means for you. Run 30 drafts before passing judgment.

What are the AI staff kits?

I built the four employees as kits so you do not have to start from scratch. One kit per employee: Voice, Research, Meetings, and Role-Specific.

A kit is a skill file. It holds the steps that run the job, the notes that carry your standards, and a template for the output. That is the gap between a prompt you retype and an employee who already knows the work.

Each kit has 2 modes. BUILD sets up your profile the first time. Your profile is the file that holds what the kit learned about you. OPERATE runs the job at your standard after that, and checks its own work before you see it.

How do I install and run the kits?

First, install it. If your app takes skill files, load the kit. If your app uses projects instead, paste the steps into a project and add the notes as files.

Second, run BUILD one time. Answer the questions. For Voice, that means 10 to 15 writing samples. For Research, your weekly questions. For Meetings, your formats. For Role-Specific, the one task worth it. This makes your profile. Keep it. It is the part that matters most.

Third, run OPERATE from then on. Give it the input. Take the output. Fix what missed. Save the fix back into the profile. It gets faster each time.

Start with the Voice Employee. Build it this week. Run 30 drafts before you judge it. Then add the next one. Each one makes the next build faster, because the work you did once carries over.

That is the whole setup.

Why does my AI output still sound generic?

The build misses in 2 ways. Both are quick to fix.

The skill does not run. The app only uses a skill when the description tells it when to. If nothing happens, just say it: “Use my Voice Employee to draft this.” Then check two switches. The skill and code tools are turned on in settings. One of those is almost always the cause.

The output still sounds robotic. That is a thin profile, not a broken tool. Add 5 more writing samples. Use the ones you wrote in a hurry. Make the “never say” list longer. Name 3 phrases you would never use. Run 10 more drafts. It gets sharp from your fixes, not from the first try.

When the output misses, fix the profile, not the prompt. The prompt stays the same. The profile is the part that learns.

Want me to build it with you?

We take one problem in your business and finish it across four sessions. You pick the problem: your pricing, your brand, your content plan, or one issue inside a company you already run. We name it, we plan it, we build the machine inside Claude or ChatGPT, and we put it to work on real tasks.

It normally sells at $999. The first 10 seats are $499, half off. Each seat includes a full year in the Operating Founders community in Skool and weekly Friday office hours. You can run your four sessions any time through December 31, 2026.

Claim Your Seat

Frequently asked questions

Do I need to know how to code?

No. You do not write code. You answer questions and paste in samples. The kit does the rest.

Which AI app should I use?

Any of the big ones works. Claude is great for writing and voice. Perplexity is great for research. ChatGPT and Gemini both work too. Pick one and start.

How long does it take to build the first one?

About an hour to set up the Voice Employee. Then run 30 drafts to make it sharp. You can do it in a week of short sessions.

Is this safe with my private writing?

Your profile lives in your own account. Only load kits from people you trust. Read the files before you turn them on.

What if the output still sounds robotic?

Add more writing samples and a longer “never say” list. Fix the profile, not the prompt. It learns from your fixes.

How much does it cost?

The kits are free. A paid AI plan runs about $20 a month. A free plan is enough to start.

Do I have to build all four at once?

No. Build the Voice Employee first. Add the others when you are ready. Each one makes the next faster.

What is the difference between a prompt and a kit?

A prompt is one message you retype. A kit is a saved setup that already knows your job and runs the same way every time.

Where do I get the kits?

Have a look at the Field Guide here and then become a paid subscriber to access all of the skill kits and files below.

Become a paid subscriber for as little as $7.99 per month and access them below…

Subscribe now

Read more

TL · DR (The 30-second version)

• AI fluency is not knowing what AI can do. It is changing how your work actually gets done. Knowing the capability exists changes nothing on its own.

• The research splits in two. Stanford finds 14 to 34 percent task-level gains. NBER finds 89 percent of executives see no enterprise boost. McKinsey finds only 6 percent capture…

Read more

Read more

Read more

Read more

Subscribe now

TL · DR (The 30-second version)

AI fluency is not a skill. It is three builds an operator installs once and runs on for the rest of their career.

• Build One: The Research Architecture - Constrain Claude to a curated universe of Tier 1 publications, Tier 2 academic centers, and Tier 3 institutional research shops. Name the mode and output format before the …

Read more