Step Two Asks For A File That Is Not Public

“Build an AI Agent in 10 Minutes. Six steps. No code. Every prompt to copy and paste.” That last promise is the interesting one, because it means we can read the six prompts as written and ask what each one is actually asking the model to do.

Five of them ask it to build. “Help me build my CLAUDE.md… ask me about my business.” “Build me an agent team for [process].” “Turn this workflow into a skill.” You supply the raw material, the model assembles it, and you can see what went in.

One of them asks it to fetch. Step two, in full: “Help me build my CLAUDE.md from scratch. Use Boris Cherny’s CLAUDE.md as a starting template.”

We went looking for that file.

What is actually out there

Boris Cherny leads Claude Code at Anthropic, and he has written publicly about how he uses it. What we could not find is a CLAUDE.md published by him. What exists instead is a small industry of reconstructions:

• At least six public repositories with his name in the title or description.
• Every one of them assembled by somebody else from his posts and threads — described, in their own words, as being based on or inspired by his advice.
• Two carrying his name while following a different methodology altogether, labelled “Agentic Context Engineering”.
• They do not agree with each other, which is what you would expect from six people reconstructing one thing from fragments.

So the honest status is: a real person, real public advice, no canonical file. Which means the prompt in step two is asking the model for something that, as far as the public record goes, it cannot have.

Diagram: why “use X’s file” fails differently from “build me a file”

This is a habit worth generalising beyond this card. Any prompt that names a specific artefact by a specific person is a retrieval request wearing a creative request’s clothes. “Write it like Hemingway” is fine — style is diffuse, the model has read a great deal of it, and nobody is misled. “Use his file” is different: there is supposed to be one object, and either the model has it or it is inventing with attribution attached.

The fix costs one clause. Ask for the properties instead of the person: “Build my CLAUDE.md with sections for role, voice, banned words and output defaults. Keep it short enough to load on every session.” That is what the card’s own mock-up shows anyway — ## Role ## Voice ## Banned words ## Defaults — and it is reproducible without anybody’s name on it.

One piece of his advice is attributable and genuinely useful, though it comes to us second-hand: several of those repositories cite a comment of his suggesting keeping CLAUDE.md to roughly a thousand tokens, because it loads on every session. We could not reach the original, so treat the number as indicative. The principle behind it is sound and sits awkwardly with step two’s framing of the file as an “onboarding doc” — onboarding docs grow, and this one is paid for on every single turn.

And for what it is worth, we checked a quotation attributed to the same person earlier this week and found the same pattern: the attribution holds, the exact artefact does not survive contact with a search.

The four file paths all check out

Having taken one step apart, it is only fair to say the card is unusually precise everywhere else. Most “no code” cards wave at a concept. This one gives you four real paths, and all four are real.

Two steps that are more than ten minutes

The headline promises ten minutes and the subtitle says no code. Step one is npm i -g, which is a terminal command even if it is not code you wrote. Two later steps carry prerequisites the card does not mention, and we have been through both in detail already, so here they are in one line each:

• Step 5, agent teams. The feature is experimental and disabled by default — it needs an environment variable set before Claude will spawn a teammate at all. And a fixed Strategist-then-Builder-then-QA chain is the pipeline case the documentation steers away from teams towards. The full version is here.
• Step 6, cloud Routines. Real, and the card is right that they run with your laptop closed. Two things it omits: the minimum interval is one hour, and a cloud routine starts from a fresh clone with no access to your local files — which matters for a card whose other five steps write files into your home directory. The full version is here.

Neither makes the step wrong. Both make “ten minutes” optimistic.

The six steps, graded

Step three deserves its own sentence, because it is the step nobody else puts on a card. “Every correction becomes its own .md file… prefixed feedback_, user_, project_, or reference_. Index everything in MEMORY.md.” That is a filing system, not a feature, and filing systems are what separate a memory folder that helps from one you stop reading after a fortnight.

One absence worth noting

This card has no fine print. Every other card from this page in the past week carries a line — “a simple framework, not an official format”, “check each tool’s official site before you connect it to your accounts”, “features change often”. This one ends at the follow button.

It is the card that asks you to paste six prompts which write files into your home directory, wire up connectors to Notion, Gmail and Drive, and schedule a job on somebody’s cloud. If any card on this page needed the grey line at the bottom, it was this one.

Four honest caveats

1. “Not public” means we could not find it. A file could exist in a talk, a thread we missed, or somewhere not indexed. We searched and found only reconstructions; that is the claim.
2. We did not run the six prompts. We cannot tell you what a given model returns for step two — only that the thing being asked for does not appear to be publicly available, which is the condition under which models fill gaps.
3. The thousand-token figure is second-hand. Multiple repositories cite it; we did not reach the original comment. Treat it as a direction, not a limit.
4. This is a good card. Four correct file paths, a filing convention most people never think of, and six prompts that mostly do what they say. The edit is one clause in one of them.

The one-line version

Five of the six prompts ask Claude to build something out of your answers, where you can see what went in; step two asks it to start from a named person’s file that nobody has published, which is the one instruction on the card that can fail silently and come back with a real name attached.

Keep the six steps. In step two, delete four words and name the sections instead.