Automattic Update Agent Context
Updates any part of the agent instruction layer: AGENTS.md, CLAUDE.md, CLAUDE.local.md,
agent_docs/, agent_docs_local/, or agent_docs/MEMORY/.
Accepts an optional topic — e.g. /update-agent-context testing or /update-agent-context linting.
If no topic is given, infer from the current task context.
What to Update
| File / Folder | When |
|---|---|
AGENTS.md |
Always-run instructions changed |
CLAUDE.md / CLAUDE.local.md |
Discovery or setup instructions changed |
agent_docs/<topic>.md |
Stable pattern learned — would have saved time if documented earlier |
agent_docs_local/<topic>.md |
Machine-specific override or local workaround |
agent_docs/MEMORY/<TOPIC>.md |
Mistake made and correct approach found — record immediately |
Two-Layer Rule
AGENTS.mdandagent_docs/are universal — fully self-contained, no global Claude assumptionsCLAUDE.md,CLAUDE.local.md,agent_docs_local/are personal Claude layer — may reference developer machine config
Rules
- Update at the same time as the code change, not after
- Keep entries actionable — commands, not prose
agent_docs/is for stable patterns;MEMORY/is for corrections and gotchas- Prefer editing an existing file over creating a new one
- Run prettier after editing:
npx prettier --write AGENTS.md CLAUDE.md CLAUDE.local.md agent_docs/*.md agent_docs_local/*.md 2>/dev/null
Memory Entry Format
- `<wrong thing>` is wrong. Correct: <right thing>.
One bullet per correction. Specific, terse, immediately usable.