SKILL.md files, CLAUDE.md, AGENTS.md. The instructions your agents follow are prose you wrote, or prose your agent wrote about itself. Either way, it steers everything that happens after.
skills/deploy/SKILL.md
The skill tells the agent how to ship.
Always push directly to main. Open a pull request. Never push to main.
An agent edit, caught in review.
Instruction files are some of the most important prose you own. A single line in a CLAUDE.md can change how every agent session in your project behaves. And they are exactly the prose nobody reviews.
Agents edit their own instructions constantly: adding a rule after a mistake, updating a convention, expanding a skill. Usually nobody reads the diff. The file that shapes every future session gets rewritten by the thing it is supposed to constrain.
The evidence that this goes wrong is not theoretical. Auto-generated context files measurably hurt model performance, and roughly half of AGENTS.md-style files never get a second commit, because nobody is tending them.
The documentation for the SKILL.md format itself.
github.comThe open-source skills repo.
morphllm.comTracking AGENTS.md adoption across 60,000-plus projects.
todatabeyond.substack.comOn whether AGENTS.md and CLAUDE.md files actually help, including the finding that auto-generated context files hurt performance and half of these files never get a second commit.
An instruction file degrades the same way any unattended document does: one contradictory rule added six months ago, one stale reference to a folder that moved, one vague sentence nobody ever tightened. Multiply that by every agent session that reads the file at the start of every task.
Nobody set out to let this happen. It happens because reviewing a one-line change to a markdown file feels like too much ceremony for a git commit, so it gets skipped, every time, until the file is quietly wrong in a dozen small ways.
Open your skills and instruction files in Stet and every edit, yours or the agent's, shows up as a proposal you accept or reject, small enough to actually read before it takes effect on the next session.
Use an AI note to tell the agent exactly what to fix in a skill, and check the file's history to see how a rule got added and whether it still holds up.
Download the starter kit. It has a template CLAUDE.md, a complete example SKILL.md, and a checklist for reviewing skill edits.
CLAUDE.md.