Below is a lightly redacted version of my AGENTS.md.

The accompanying reference files look like this:

references/
├── about-me.md
├── agent-clis.md
├── ai-models.md
├── browser-login.md
├── db-and-bulk-writes.md
├── dual-machine-setup.md
├── gog-cli.md
├── hermes.md
├── project-map.md
├── signal-bridge.md
├── signal-contacts.md
├── whatsapp-bridge.md
└── whatsapp-contacts.md

Notes:

• Consensus says that the July models (Fable, Sol, Opus 5) work better with less directive prompting. This is a good time to trim your AGENTS.md and move task-specific stuff to skills or reference files.
• Claude Code has a new /doctor command to help refresh your AGENTS.md and skills setup.
• Hard rules: I really care about the instructions in the “hard rules” section. Most of them are enforced by deterministic hooks as a backstop.
• Agents should write and iterate your AGENTS.md, but my sense is that bothering to read and hand-edit the file—and your most frequently used skills—seems worth doing roughly once a month.
  • It’s hard to know what’s best, but it’s easy to prune the obvious mess / ambiguity / inconsistency.
  • I find that reading the files often generates ideas for new instructions or skills.
  • I’ve added this task to my monthly review: “Read AGENTS.md and at least three of my 20 most-used skills. Also check the debug log″.