The build story · Market Scan

I spent a month turning 500 sessions into four.

Market Scan started as a folder full of a client’s documents and one coding session that disappeared into the market for two days. Four versions, billions of tokens and hundreds of experiments later, the hard part turned out to be the text.

4versions
290–500sessions at the peak
2.5 daysrecord run
4 × 50 minstable method

01 · Where it began

It started with a dumping folder.

No product screen. No neat four-step framework. I dropped in whatever the client had and waited to see whether one machine could hold the whole market in its head.

The first version worked like this. I dropped the client’s domain into a folder, added a recording of the call in whatever format Whisper could understand, and put every existing document beside it. Then a dedicated coding session went away for two days and dug through the market.

Every hour it deepened the weighting of hypotheses, found evidence, built maps of job titles, connected company size to role and market, and took apart every competitor around the business. What came back was already useful. It was also slow, expensive and frighteningly dependent on one session keeping the whole world in context.

Even then I knew what kind of animal this was. I had held a $2,000,000 consulting report in my hands before, and this rough machine could already see connections that report had missed. That was the good news. The bad news was that I could not make a business out of waiting two days for one enormous session to survive its own reasoning.

Version two stunned people. Version three made that look small.

The people I showed version two to were looking at their own businesses and seeing markets they had never named. Version three was several times better. By then the release looked close enough to touch. The research was wider, the letters were more specific, and the market vocabulary went far beyond anything a normal database filter could express.

Then GPT 5.6 Sol Max appeared, and I could not leave the possibility of another jump alone. Several more days disappeared. Every weekend. Sixteen hours a day. I kept running the same difficult work through different models because “probably good enough” is a ridiculous standard for something I was going to sell and answer for.

02 · 13 July, 18:50

The hammer changed its molecular structure in my hand.

In the middle of the work, both models I depended on stopped seeing the actual repository. I would say one thing, and they would grab that sentence and walk circles around it while the working code sat untouched. At 18:50 on 13 July, GPT 5.5 and Opus were doing both the design and the implementation noticeably better than Sol Max and Fable 5.

I had watched models degrade before. I had never watched them swing tens of times in either direction inside a week. With a small task, this looks like a bad answer. With a genuinely large repository, it feels like the tool in your hand has changed material and forgotten what a nail is.

I had 1,000,000 nails to drive. The hammer became solid occasionally, and I had to guess the moment.

03 · Version four

At five in the morning, I started writing the whole thing by hand.

Both major providers had told me to fuck off in their own technical language. The system still had to ship.

I sat there and wrote a ten-thousand-line prompt: market research, copywriting and business knowledge in one document. It was not a clever wrapper around a few API calls. It was an attempt to write down every judgment the work required, including the things that had lived in my head for years and had never needed names.

That became version four. At its peak, one scan for one product used between 290 and 500 sequential sessions. The record run took two and a half days. I burned tens of billions of tokens on experiments, and for eight straight days I ran almost nothing else.

290–500sequential sessions
2.5 daysfor the record run
8 daysrunning nothing else

A month went into making the models produce work that held together without deterministic constructors, brainless gates or a person assembling the answer between steps. I was not chasing a prettier paragraph. I was trying to make research, business reasoning and writing stay good at the same time, on any business, without me standing between the stages and fixing the joins.

The number of parameters that had to get along was absurd. A broad market map with weak letters was useless. Strong letters built on invented markets were worse. Thousands of search combinations meant nothing if the names collapsed into duplicates or if the person at the end had no authority. Every part had to survive the next part.

04 · The breakthrough

Everything hard turned out to be in the text.

I kept refining the doctrine and the paperwork until one machine could inspect another machine’s work, point to the weak part and send only that part back. I had to define “good” precisely enough for the system to use it, without reducing meaning to a dumb validator.

That is where the four sessions came from: Research, Writer, Market Arms and Judge. Each gets one job. Each can take up to fifty minutes. Each stands on the result before it. The stable version no longer asks one model to keep the whole world in its head.

Before500 sessionsup to 2.5 days
After4 sessionsup to 50 minutes each

The quality went up. The system drank all my blood and a few other fluids. I am sick of it. I am not stopping.

05 · Fifteen full runs

Belief was no longer enough. I ran the whole system.

Five models received the same work across three real client inputs, with no hints along the way. The data was exported read-only from production into an offline sidecar. Production stayed untouched. All 15 cells completed with valid JSON. The run took a full day.

Totals across three business inputs, August 2026. Minutes are the average per complete run.
ModelAvg minSegmentsArmsArms / segmentDone
Qwen3.8 max41.15057511.53/3
GPT 5.5 xhigh16.44866113.83/3
GPT 5.6 Sol Max41.8734,40760.43/3
Kimi k3 max29.0391,27532.73/3
GLM5.2 max via Qwen CLI31.1392366.13/3

The useful column is arms per segment. Sol produced 60.4. Kimi, 32.7. GPT 5.5, 13.8. Qwen, 11.5. GLM, 6.1. The same work and the same inputs created a tenfold spread in the combinations a sourcer could actually pursue.

On one client, GPT 5.5 found 27 segments and 463 arms. Sol found 29 segments and 2,301 arms. The segment count looked tied. The usable combinations were five times apart. Another twenty to thirty minutes of grinding through vocabulary changed the width of the market for the whole year.

Why I built it

The buyer can already be in the database and still remain invisible.

I opened Apollo or Clay, selected an industry, company size, country and job title, and got a file with 10,000 rows. It looked like the market. It was the answer to the filters I knew how to type.

Any company that described itself differently stayed invisible. So did every buyer carrying a title I had never heard.

Over 17 years I gave LinkedIn $20,196, mostly through Sales Navigator at $99 a month. I cancelled it in January.

I had also watched a channel deliver 300 leads and produce 0 deals. Three hundred and zero. Reaching people was never the whole job.

Every wrong row becomes a paid envelope.

Physical mail puts the error on an invoice: paper, printing, the object inside, assembly, postage and a week of travel. The market, the person, the argument and the building all have to agree before that spend begins.

The method, briefly

Four sessions. One job each.

  1. 01

    Research

    Builds markets, weighs hypotheses and marks where the evidence stops.

  2. 02

    Writer

    Gives every accepted vertical its own buyer, argument and chain of letters.

  3. 03

    Market Arms

    Finds the company names, functions, titles, geographies and scales no single filter can express.

  4. 04

    Judge

    Rejects attractive answers that cannot defend themselves and returns weak pieces for rework.

After the scan, company sourcing and physical-route verification begin. The system connects the current company, current person, current role, correct site, mailing address, sources and conflicts. Only verified routes enter production.

The next step

Send the business. I will start with the market.

A name, work email and website are enough to open the file. Add call recordings, a presentation or other materials if you have them.

Send the business file