agentreadmelisted

# agentreadme Grades a repository on how well an AI coding agent can work in it, then fixes what it finds. Runs entirely on the local machine, so it works on private code. The five things it measures, worth 100 marks in total: | Category | Marks | The question | |---|---|---| | Instructions | 27 | Is there an AGENTS.md, and does it name real commands? | | Setup | 20 | Can an agent install this and run it unattended? | | Verification loop | 25 | Can an agent check its own work, or is it editing blind? | | Context economy | 20 | Does the repo fit in a context window, or fight it? | | Navigability | 10 | How fast can something new orient itself? | ## Step 1 — measure Run it against the repository root: ```bash npx --yes agentreadme@latest . --verbose ``` Nothing is uploaded and no network call is made. It grades what `git ls-files` reports, so untracked build output does not count against the score. Useful variants: ```bash npx --yes agentreadme@latest . --json # machine readable, for reasoning over npx --yes agentreadme@latest ../other # a different directory ``` If `npx` is unavailable, say so and stop rather than guessing at a score. ## Step 2 — report what matters, not everything Give the user the overall mark, the weakest category, and the top two or three fixes. Do not paste the whole output back at them; they can run it themselves. Lead with the single most costly thing. A missing test command matters more than a missing licence, and the tool alread