A market that lives in the founder's head cannot be checked. Written as a file with edges, it becomes the search the next stage runs on its own.
A spreadsheet of who you sold to. Scored on what they share. A column for size, one for sector, one for the deals that never closed. For a forty-person agency the market lives in the head of the founder, and nowhere a machine can read it.
By hand today
“I’d track who you try to sell to in a spreadsheet (including those that just drag on and never close) and I’d score them across the behaviors/things you would expect your target audience would have in common”
james_impliu, on Hacker News
That is the honest version of this stage. The founder lists every company they have tried to sell to and scores each row on what the good ones shared: size, sector, whether the buyer was the owner, whether the deal closed or dragged.
What comes out is a feel for the market, and the feel stays with the founder. When a hire is brought in to build the list, the advert hands them “defined ICP criteria”, already written. Nobody is paid to write the definition. It is assumed to exist, usually as a paragraph in a deck and a set of instincts in one person. A list built from instinct cannot be checked.
What it costs you each week
$800 to $1,200 a month, the advertised pay for the specialist who turns your market definition into a list. The definition itself is never in the advert. You hand it over, already written.
Source: a remote-hire job advert for a Clay and Apollo specialist on OnlineJobs.ph, read 7 September 2026
There is no sourced figure for the hours a founder spends defining a market. The time figures that circulate come from vendors of list tools. What can be sourced is the hire, at that monthly rate, plus the seats on the tools. The cost nobody invoices is the loose definition: it lets in companies that were never going to buy, and every later stage spends time on them.
The build
Open Claude Code. Write the sentence. One line of who you sell to, then the edges: size as a band, sector from what the company says it does, place. It comes back as a file the next stage reads. A dry run over twenty companies shows the edges before anything is saved.
Open Claude Code in the folder that will hold the whole engine, and say what you want in plain words:
Help me write down my market so a script can list it. Ask me who I sell to,
in one sentence. Then ask for the edges: the smallest and largest company
that fits, as a headcount band; the sectors, named by what a company says it
does rather than by the label a directory gives it; the countries or cities;
and who must never be contacted even when they fit. Save the answers as
market.yml, one field per edge with a plain line saying why it is there.
Then take twenty companies I paste in and tell me, for each, whether the
file keeps it or strikes it and which edge decided.
Claude Code asks one question at a time and pushes back where an answer is vague, so “mid-sized” becomes a band and “not the big ones” becomes a ceiling with a reason. Then it writes market.yml at the root of the engine folder.
Paste in twenty companies you know, some you sold to and some you wish you had not chased. It returns a table: company, keep or strike, and the edge that decided. Where you disagree, change the file, not the row. The file is the definition now, and the next chapter reads it as the search.
The pattern is the one ArkOne runs on its own address book. plans/prospect-research/classify-industries.mjs sorts every company into one of twelve main industries from the company’s own description, never from the label a directory imported, and runs as a dry run first (line 18). The definition itself lives as a stored profile: docs/mcp-team-guide.md lines 89 and 213 describe the ceo profile holding the headcount band, the geography and the hard gates, and a separate !has-ai-capability profile for who is never contacted however well they score. Two stored rules every later stage can read is the shape to copy.
What it feeds
Listing every company. The file is the search. The next stage runs it over free sources and returns a list with its total. The number to read: how many of a twenty-company sample you strike by eye. More than a handful means a loose sentence.
The next stage lists every company the definition returns, without buying a database, and it can do that only because the definition is now a file with edges.
Two numbers prove this stage worked. The first is the count of companies the definition returns, shown with its total. The second is the strike rate: take twenty companies from that list, read them by eye, and count how many you would throw out. More than a handful, and the file has a loose edge to tighten before the next run.
What stays with a person
Where the edge sits. A company on the line between two sectors. A headcount that is a band for one row and a count for another. A person rules once, writes the ruling into the file, and the machine applies it every time after.
Every real market has a fuzzy edge, and the tool finds it on the first run: a firm that does half consulting and half software; a company whose imported label says one sector and whose own website says another; a headcount given as a band for one company and an exact count for the next.
The tool cannot rule on these. A person reads the case, decides, and writes the decision into market.yml with the reason. What stays with the person is the ruling, and the sentence itself, which nobody else can write for you.
The tools you already pay for
Nothing new to buy.
CRM gains one field per edge, so a company can be filtered by what it is, and counted.
Spreadsheet holds the old list of who you sold to. Read once, to score what the good ones shared, then retired.
Notes hold the sentence, pinned where everyone can see it.
One sentence of who you sell to, with its edges written down. Then the list can build itself.
Ask: Could a stranger read your market definition and strike the wrong companies from a list without asking you anything?
This definition is chapter 1 of twelve; the contents list every stage, and chapter 2 builds the list from it. The book, Building a real outbound engine with Claude Code and Codex, end to end, runs the whole engine in order, one chapter per stage. A free starter edition for firms under fifty people is coming through the 100k programme.