About agent-config

I got tired of agent docs that read like they were true forever. Tool defaults change on a Monday. A page that was right in March is wrong in September, and nobody put a date on it.

This site is a field guide for running coding agents well. I publish what is currently correct, and I say when I last checked.

Content half-life

Not every page ages the same way. Mixing a durable idea with a weekly changelog and a skill dump produces the worst page: always half right, always half stale. So I keep the collections separate.

Learn is the curriculum. Instruction files, context as a budget, verification loops. I rewrite a learn page when the idea changes, not when a CLI flag moves. It still gets a lastVerified date, because even a durable claim can rot if a tool ships a new default.

Playbooks are walkthroughs. They name tools and commands. They go stale faster than learn pages, so I keep them short and refresh them when the commands stop matching the docs I can open.

Digests have a one-week half-life on purpose. Each one is a snapshot of what changed in Claude Code, Codex, and Cursor that week, and what I would do about it. I do not update an old digest to pretend it is still this week. I write the next one.

Kits are starter sets: a few curated skills that work together, plus an AGENTS.md you can paste. A kit goes stale when a skill in the set does, or when I would no longer install that combination.

The verified-against promise

Every learn, playbook, and kit page has two fields I treat as a contract.

lastVerified is the date I actually opened the sources. Not the date I first wrote the page. Not the date a model remembered a version number.

verifiedAgainst lists only the tools whose docs or changelog I opened, with the version I saw. If I did not open it, it is not in the list. I delete claims I cannot check. I do not hedge them.

The badge at the top of those pages shows that date and those versions. After 45 days it flips to "May be out of date". A script lists those pages and we open refresh issues. The 45-day line lives in the badge and in the staleness check. It is not a vibe.

How the site is produced

A script collects the week's releases into JSON. An agent drafts the digest and the refresh PRs from that input and from the page it is supposed to update. Bugbot reviews the PR. A human reviews it before merge. That is the whole pipeline. I do not publish from a model's memory.

Contribute

If a page is wrong, or you want a playbook that is not here, open an issue. Say what you checked and the URL you checked it against.