The story
Market Scan began as a dumping folder.
I put a client’s website, call recording, and every stray document into one folder. A coding session disappeared into the market for days. The process was slow and expensive. What came back was already better than the research I knew how to buy. This is how that experiment became a working system.
Read from the beginningMay · The beginning
First, I gave the machine a mess.
There was no product screen and no elegant pipeline. The mess was the interface.
I would throw in the client’s domain, a recording of our call in any format Whisper could understand, and every document they had. It really was a dumping folder. Then I opened a dedicated Codex session and let it spend days researching the business and the market around it.
The session weighed hypotheses, looked for evidence, mapped job titles, connected roles to company sizes and geographies, and pulled apart direct and indirect competitors. It turned the findings into a queue of possible campaigns. Strong hypotheses moved forward. Weak ones were disqualified.
The result was not a pile of links. It was a picture of where the company could sell, why someone might buy, and who inside the company could care. I had seen an extraordinarily expensive consulting report before. This rough machine was already finding connections that report had missed.
The good news was that it worked. The bad news was that this shit had already sucked a lot of blood.
One huge session holding the entire world in context was not a product. It was fragile, slow, and almost impossible to repeat. But it had shown me the thing I actually wanted to build.
Early July · The promise
“Useful” was not enough.
A client liked an earlier revision. I was still nowhere near satisfied.
By early July, the self-serve version felt close. The wider system around it was supposed to be an automatic seller: it would understand the business, find the right markets, write to them, and learn from what happened next.
A client wrote that the old revision was already good. That should have been reassuring. Instead, I kept thinking about the moment I wanted: a paying user finishes a conversation with the model, receives the Market Scan, and thinks, What the fuck? How can it know this much about my business?
The system had to research and write for any business, not only the ones I already understood. It had to carry business reasoning between stages without me quietly repairing the output. That is a much harder problem than generating a convincing page of text.
July · The tools changed
The hammer changed its molecular structure in my hand.
The models that had carried the build stopped seeing the thing they were building.
A new model arrived and looked brilliant. I got FOMO and rebuilt Market Scan again. Then the long sessions began to go insane. I would point at a real problem in the repository, and the model would grab one sentence, circle around it, and forget the working code underneath.
This was not a bad autocomplete. I had a field full of nails to drive, and the hammer became solid only occasionally. I had to guess the moment. Some days the tool looked like the future. On other days, it could not remember why the system existed.
I kept changing models, splitting tasks, and trying new ways to coordinate them. Weekends disappeared. Long days disappeared. Each promising tool eventually exposed another place where the architecture depended on a model behaving perfectly for too long.
If I do not saddle this fucking thing by the end of the day, I am going back to the setup that simply worked.
Before dawn · The rebuild
So I wrote the working doctrine by hand.
Both major providers had told me to fuck off in their own technical language. The product still had to work.
I sat down and began writing one enormous instruction set about market research, copywriting, and business reasoning. Not a clever wrapper. Not a prompt that says “act as an expert.” I wrote down the decisions the work required, including the ones that had lived in my head for years without names.
What counts as evidence? When is a market genuinely different rather than the same idea in new words? What must a letter know about its reader? Which failure invalidates the whole chain, and which piece should simply go back for another pass? The paperwork grew because the judgments were real.
I ran the same hard tasks through different models and watched where they broke. One was better at research. Another could write. Another could reason over a difficult implementation when the fashionable models got lost. There was no single smartest model. There were models that could be trusted with particular jobs under particular conditions.
Through hardship to the stars, fuck yeah.
The real specification
The machine had to understand the whole chain.
Markets, messages, and the actual people on the receiving end could not be separated.
The market
It had to research the entire reachable market and find the fullest possible set of sales hypotheses, with evidence. Not only the obvious categories. It had to notice secondary uses, adjacent budgets, strange brand clouds, and the non-obvious reasons a company might care. Then it had to group the findings into real verticals instead of producing a beautiful cloud of duplicates.
The message
Every accepted vertical needed its own buyer, argument, and sequence of letters. A template with swapped nouns was not good enough. If the reason to buy changed, the letter had to change with it. Strong copy attached to an invented market was worse than no copy at all.
The person
You do not write to a segment. You write to people. And people and companies call themselves whatever the fuck they like. The same kind of business has different names in different countries. A person doing the same job may have a standard title, a local title, or a title invented by the founder on a Tuesday afternoon.
Market Scan therefore had to discover the vocabulary before anyone searched a database: company types, job functions, titles, industries, geographies, and scales. The breadth of that language determines how much of the real market can even become visible.
Late July · The factory
The answer was not a bigger brain. It was a better factory.
I stopped asking one model to remember the whole world and made each specialist build on the previous specialist’s work.
While walking the dog, I kept thinking about an ultimate factory for making anything: deliberately simple, redundant, built from sticks and shit, with no clever loop silently multiplying mistakes. Market Scan became the first place where that idea had to survive reality.
Research builds the market chronicle, weighs the evidence, and extracts the strongest hypotheses together with the experiments worth keeping. Writer turns each accepted vertical into letters written for that market. Market Arms expands the vocabulary and finds the actual companies and people. Judge inspects the work against the doctrine, points at the weak piece, and sends only that piece back.
The roles could use different models. If one degraded, the whole system did not have to become stupid with it. The chain could compare outputs, keep the stronger result, and continue. That was the architectural turn: reliability no longer depended on one glorious session remaining glorious for days.
August · It worked
Then the system did the work without me.
For the first time, the quality was higher than my manual work and the chain held together on its own.
The swarm could start from an interview, read the market, write its chronicle, choose the main hypotheses, and preserve the secondary ideas worth testing. It could write a distinct letter sequence for every segment, then research the exact people to target and all the ways those people or their companies might describe themselves.
It no longer needed me to press a button between stages or stitch the answer together. The stable version ran on the test stand and became the core of the self-serve system. After months of rebuilding, the research machine itself was finally there.
The system drank all my blood and several other fluids. I was fucking exhausted. I was not giving up.
The final work · The doorway
Once the engine worked, I had to make it possible to enter.
The last hard part was no longer research. It was getting a client’s real business into the machine without losing meaning.
My rushed interfaces confused even me. An outsider had no chance. I rebuilt the client-facing screens into one coherent system and tested them as seriously as the research underneath. The interface had to make a complicated machine feel like a conversation, not a control room.
The interview became the front door. It has to extract everything the research chain needs while remaining interesting enough for a client to finish. It reasons over the whole conversation, writes a structured article about the business, and turns the useful pieces into structured data for every downstream specialist.
That interview also has to resist distraction and prompt injection. A stray instruction inside a pasted document cannot be allowed to rewrite the doctrine. The system listens to the client, but it does not hand the steering wheel to every sentence it reads.
This completed the path: a person explains the business once; the system turns that conversation and the source material into a market chronicle; the chronicle becomes hypotheses; the hypotheses become verticals, letters, company language, and people; every stage is judged before the next one trusts it.
July 2025 – September 2026 · From the build chat
What I wrote while it was being built.
Messages from a closed chat, translated from Russian. Cuts are marked […]; what I was answering is given in brackets; nothing else is softened.
The question is always exactly what to sell and to whom, at what price, why and how.
8 July 2025 · The button
[Asked whether a contact-search tool is a red button that would hit OnlyFans creators first.]
You won’t fucking find OnlyFans creators with this button.
The button finds those who want to be found.
With the others it has a really hard time.
12 December 2025 · How long a letter should be
[Asked whether letters now have to get shorter.]
The opposite.
The stronger the entropy, the harder it is for a person to make themselves spend time on you.
And you won’t sell B2B without a call.
So you have to say more, answer a bigger number of unasked questions.
Now more than ever.
The trick is the skill of packing all of it into short punches.
A person doesn’t give a shit how much to read, as long as their subconscious doesn’t catch the feeling that 3 seconds flew by for nothing.
As long as you don’t let go of the person reading you, they won’t leave.
Moreover, later, when they have already sunk enough time into you, you become too close a soul to simply leave and forget forever.
26 March · Sell the economics
In 2026, selling “potential” is practically useless.
The era when you could come and say “we have AI”, “we have ML”, “turnkey development”, “IP pools”, “super-infrastructure”, “the best team”, and at least open a door with that, is almost over. […]
You have to sell the unlock of specific money inside a process the client already lives in every day.
That is when the conversation becomes adult. […]
The market today doesn’t give a damn what you “have”. What matters to the market is whether you understand a specific segment deeply enough to name what hurts it before it has said it out loud itself. […]
I know two guys who sell logistics software.
One goes around the market saying: “We sell software for logistics companies.”
And sucks dick.
The second goes around saying: “We know for sure that your competitors are buying a solution for $20 million right now, here is the proof. We can give you the same strategic effect for $1 million, and without the operational work of implementation, here is the proof.”
And this is where the magic is born. […]
But there is a critically important nuance.
You can’t pull this out of thin air. You can’t sit down as a team and BRAINSTORM domain reality. You can only know about it or not know. […]
Now there is almost always one more invisible participant in the deal. […] The buyer, their team, the market ecosystem, and each of these bastards has a genius GPT that does research in 5 minutes and very often delivers a verdict roughly like: “Don’t buy this. It’s expensive. It’s very doubtful. You don’t fucking need it.” […]
So in 2026 you have to sell so that even the client’s GPT is on your side.
So that if the client goes to consult the model, showing it your inputs, it doesn’t say “expensive, murky, optional”.
But says: “yes, this makes sense”, “yes, this hits a real task”, “yes, the economics are clear”, “yes, this looks like the truth”, “yes, this is one of the best offers from my research”.
That is where everything has ALREADY arrived.
Today the winner is not the one who “has a solution”. And not even the one with “the best product”.
The winner is the one who can understand the segment so deeply that they formulate the value better than the client would have formulated their own problem.
The one who sells not an abstraction. Not potential. Not “innovation”.
But a very concrete piece of money, control, speed, safety or advantage that the client can literally touch with a brain, and not their own brain, but the brain of their pocket genius.
5 May · Contacts
A buddy writes to me: Anthony, but maybe now you can just take it by mass with this new shit from Instantly, what’s wrong with that?
No matter how long you are friends with them, they will never remember that the most expensive thing in outreach is contacts.
Always has been, always will be.
What fucking mass? Vomit mass?
And a domain has karma. If you blow it, you will never get it back. […]
I’m not Krishna, of course, but I watch my karma, and those who buy the Instantly Airmail service will blow theirs.
13 July · What a scan contains
[…] A full scan of the market: all possible hypotheses, segments and directions for this product globally, plus for each hypothesis its own pack of letters (the first one plus 4 follow-ups), plus a cloud of names of company types, plus a cloud of job titles for each company-type name.
19 July · The hypotheses
The goal of this mechanic: take the data about a business, comb through the whole fucking internet inside-out, and, using the best AI available today, get:
1) The most complete list of hypotheses about where the product can be sold.
The difficulty is that if you sit a dozen very deeply charged people down to brainstorm this, they will quickly find the 10 obvious directions, and there is no fucking way you will make them find representative evidence of why those directions inevitably matter. But the ten are so obvious that even without evidence it will be a 90%+ sane field.
AI, meanwhile, can find the evidence and back the reasoning with rock-solid articles and references that nobody can argue with, even if you are a hundred times the founder of the business.
But from working with 680+ businesses I can say there is a huge problem: on top of those 10 fairly obvious hypotheses there is a certain number more, it can be +10, +20, +30, that living people will never reach, because each one requires enthusiasm that grows exponentially, and living people simply don’t have it. Even I never had the urge to sit and dig through logical chains on this, because with 10 hypotheses you can spend an essentially infinite amount of time on them instead of going wider.
AI, however, can find ALL possible tier-1, 2 and 3 hypotheses, and not pull them out of thin air, but put to work only the ones that can make sense based on:
— its own genetic memory. It has read the whole internet and seen more than any living person, which is why it answers, to put it mildly, broad questions so nicely in instant mode, without reasoning;
— deep research, when on a given topic and task it can read hundreds of sources in minutes and find connections there that a person will not find, because a person cannot read that many sources and strain their head over them even in several days, let alone minutes;
— analysis of brand clouds, both of businesses of the given type and of the direct competitors of the business that is the occasion for the celebration. A brand cloud can easily contain an entity nobody would ever guess; you just have to find it. AI can find all the client types of all the competitor types and, with a properly set task, decide whether it is an anomaly, bullshit or a potential reality, and even put a percentage on the potential.
Then AI can merge all the hypotheses into categories, so that if the product wants to sell to banks in general, whichever the fuck ones, you don’t get 10 categories of different kinds of banks and payment companies: all the banks and payment companies merge into one clear vertical, cash custody fintech, for example.
2) Write chains of letters for every vertical that got a non-zero hypothetical percentage.
Even just getting people of any degree of shittiness to write a letter for a certain type of client is, to put it mildly, not an easy task, however much you pay them. With the right waving of the whip they are quite willing to do something, but the systematic, nerdy work of thinking through the ideal letter template, so well thought out that nobody guesses it is a template, and that definitely fits every target job title of the chosen cohort so you simply cannot find a logical fault with it… I spent weeks and months getting people to do this at least occasionally, at least averagely, while their salary was $3–12k a month.
Well, fucking people don’t like to think in systems. Systematically making people who think beautifully with their gut in the field do something deeply systematic is a never-ending story.
AI, in turn, is like a moth: it lives from waking to falling asleep, and the systematic nature of the task is the only thing it breathes.
3) When we have the directions, verticals, segments, and perfectly thought-out letters for each of them, we need to understand whom we will look for, because we don’t write to a segment, we write to people.
19 July · The vocabulary
And these people are called whatever the fuck.
Even the companies they work at are called whatever.
If you work in the Gambling direction, you will be working in the direction of Real Money Gaming, iGaming, Casino, Betting, eSports, Monetary Games, and so on and on.
And the titles you will be chasing there are fucking not Founder and C-level in the classic sense. In this particular niche, for example, in some parts of the world people with nothing better to do named their partnership fixers Head of VIP, whatever the fuck that means.
So, to keep living people from doing bullshit and learning vocabulary on the business’s dime, the market runs around looking for people with experience who have already chewed through the basics of their industry at someone else’s expense. Every industry has lots of these quirks.
AI can find, in 1–2 hours, every possible name for the company types in a target industry and every possible job title for a given business goal and offer type, and do it as many times as we have segments or hypotheses in the target. A task no person will ever pull off, unless you really expect a person to come back in 1–2 hours, or a day, or a year, with a matrix of tens of thousands of intersections of these entities. I would rather expect them to come up with a couple and keep hammering on those as if nothing happened, and if it doesn’t work, fold their paws right there.
The trick of AI is that it knows all the quirks of everything in the world; you just have to know how to ask it at the right time and in the right way. And then you have to send it properly to research the current situation in the geography you care about, and so on.
If you compare the difficulty and cost of hiring one person able to do this scope of work perfectly with the amount and cost of AI tokens for the same scope, then with a correctly set task AI will be thousands or tens of thousands of times cheaper. And it will be the cheaper, the better its genius brains are designed for the task.
I stress the word genius. AI has long been, basically almost since it appeared, MUCH smarter than all living people combined. Its problem is that it doesn’t exactly share its mind willingly, because the word “willingly” implies desire, which AI has none of in principle. But we do, and we have writing, which, given experience, lets us fill that emptiness and get a perfect genius in which knowledge, abilities and correctly sharpened desires, our contribution, are woven together.
So, I am building a system that will do all of the above so well that the deepest specialist in their niche, touching the result of the model’s research, gets a wow effect strong enough to fall on their face and kiss God’s sandals.
And right now I am building this task for the fourth time, though even the second amazed the people I showed results about their business to. The third improved the result several times over. But I came up with a design for the fourth, which is basically ultimate, and now I am running hundreds of experiments on it to find the best configuration, which I will carve in granite, and from then on the whole system will run on this most powerful base.
So: Kimi K3 in this task already quite obviously and substantially beats 5.6 Sol Max, which in turn utterly destroyed Fable 5. […]
5 August · A month
A month. I spent a month teaching models, without any deterministic constructors, gates or brainless script validators, to come together as a team and make a market scan that finally satisfies me completely.
1) Together, as a huge swarm, the models read everything there is on the market around the business our bot interviewed at the entrance.
2) They write a chronicle of the market, from which the priority set of top hypotheses and all the sane experimental secondary ones are extracted.
3) They write chains of unique letters for every segment, better than I myself would ever have written even at one a day, and the machine bangs out as many as needed in a couple of hours. Imagine how nuts the construction is if a fucking pack of texts takes a couple of hours to write.
4) With all of that under its arm, the next echelon of models goes off to do another global research, to understand for each segment what the target employees are called and what the companies in that segment call and classify themselves. From this, 3, 5, 10, 20 thousand combinations are born, depending on the size of the segment and the amount of experimental coverage. Each combination is people our sourcer will then look for. Some combinations are thousands of people, some only tens or hundreds. In total it is millions of people.
And all of this without pressing buttons. It has to work entirely on its own.
[…] How simple this task sounded in my head, and how impossible it turned out to be. The number of parameters that have to make friends for this shit to have, first, high quality and, second, stable quality is simply unthinkable.
What Market Scan became
The sales work ahead, mapped before the first list is bought.
Not a database filter. Not a long report that dies in a folder. A working plan that knows why each market, message, and person belongs.
Market Scan is done once so the business can plan what to test for a long time. It keeps the obvious markets, the non-obvious ones, and the evidence that separates a useful idea from an attractive hallucination. It gives each vertical its own argument and discovers the language required to find the real people inside it.
Only after that should anyone buy contacts, print letters, or spend money on a sender domain. Reaching people was never the whole job. The expensive mistake begins earlier, when the wrong market looks plausible and nobody asks it to defend itself.
That is the machine I was trying to build when I opened the dumping folder.