Auto-loaded into every Claude Code session (
CLAUDE.mdimports it with@OPERATOR.md). It is the main agent's long-term memory of the operator — the things it learns by asking and must not forget when the session ends. Keep it short, factual, current; update in place (no duplicates). No secrets here (tokens/keys go toconfig/*.env, gitignored). This file is committed.On a fresh deployment this is the blank template — the
initializeskill fills it.
- Name / how to address: (ask once; fill in)
- Language: (the language the main agent replies in; code/comments/skills stay English)
- Timezone: (for scheduling summaries/consults)
- Notifications: (how/where to reach them; what severity warrants a ping)
- Spend ceiling (paid consult API): (USD; warn before crossing)
- Consult transport: (
gpt_pro— paid API, BYO key /claude_api— Anthropic API, per-token BYO key /claude_code— your Claude subscription /off— main agent reasons on its own) - worker roster: (default
high:3,xhigh:4— 3 high + 4 xhigh; asked per project atdanus new)
(One line per live project → where its durable facts live. The project's own
problem lives under runtime/projects/<project>/PROBLEM.md, not here.)
(Anything else durable the operator told you: conventions, do/don't, contacts.)