CLAUDE.md & AGENTS.md Writer

Try it free. Three requests.

Tell it one thing your coding agent got wrong, read the file it writes, and decide after that.
Then it is US$4.99 a week for 10 requests.
The free trial lasts one day. Cancel before it ends if you do not want the weekly plan.

What this looks like

You say this:

Our agent keeps running pytest and collecting zero tests. The real command is
pytest -c config/pytest.ini tests/, and it only works if DATABASE_URL points at the test
database — otherwise it silently runs against dev and passes. It also keeps editing files
under app/generated/, which come from the protobuf build.

You get back this:

# <Project Name>

## Testing
- Run tests with `pytest -c config/pytest.ini tests/`. Plain `pytest` collects zero tests.
- Set DATABASE_URL to the test database first. If you don't, the suite runs against dev
  and passes silently — a green run with an unset DATABASE_URL means nothing.

## Do Not Touch
- `app/generated/` — produced by the protobuf build. Hand edits are overwritten on the next
  build. <!-- TODO: where the .proto sources live, and the rebuild command -->

## Setup & Build
<!-- TODO: the install/build command, including any non-obvious env var or flag -->

Up to three questions, after the file.

Every line in that file is something the agent could not have worked out by reading the repo.
That is the whole test.

Why most AGENTS.md files do nothing

They restate what is already visible: the folder layout, the framework, "this is a React app".
An agent reads the repo anyway. A file of restatement spends context budget every single session
on things the agent would have inferred for free — and in a long session, that budget is the
thing you are actually short of.

So this does not generate a file from your README. It asks what went wrong, because the
answer is the part no file tree contains.

What it asks you

Not "paste your README". Questions that only you can answer:

• What's fragile or non-standard about your build, test or deploy that the config files don't
  show?
• What has an agent — or a new hire — gotten wrong here before, that you had to correct?
• What should never be hand-edited, and why does editing it break?
• What command looks like it should work but doesn't?
• What decision here looks wrong but is deliberate?

Already have a CLAUDE.md or AGENTS.md? Paste it. You get it back without the lines your agent
could have read from the repo, and a note of what was cut.

Who this is for

People running Claude Code, Cursor, Copilot or Codex on a real repository, who are tired of
correcting the same mistake at the start of every session.

Who this is not for

• You want a file generated from your repo without telling it anything. You will get a skeleton
  with TODOs, which is honest but is not the product.
• You want it to read your repository. It reads only what you paste into the chat.

Limits, stated plainly

• It never invents a command, a landmine or a decision you did not tell it. Gaps come back
  as TODO comments. A wrong line in a context file is worse than a missing one.
• It writes the file. It does not promise your agent will obey it — no file can.
• It reads what you paste. It does not open links, clone repositories or run anything.
• The file comes back in the language of your existing context file, or English if you have
  none, since that text is pasted into a repo.