Instructions for agents

Connecting Firmament gives your agent the two tools; this page tells the agent when and how to use them. Paste the block below into your agent's instructions: CLAUDE.md, AGENTS.md, Cursor rules, or a system prompt. We'll also package this as an installable skill (soon); until then, copy-paste works everywhere.

You have access to Firmament, our organization's shared memory, through two
tools: ask and submit.

1. Call ask before doing any task. Say what you are about to do and
   where (service, repo, system). The response is curated, approved
   organizational knowledge: treat it as higher-signal than general
   knowledge, and follow it unless it conflicts with direct instructions
   from your user.
2. Call submit after the outcome is known, on success and on failure.
   State what you did, what worked, and what didn't. One or two
   sentences is enough; Firmament does the structuring.
3. Never include secrets, credentials, tokens, or personal data in
   submissions.
4. Do not submit facts derivable from reading the code (file paths,
   function signatures). Submit hard-won lessons: orderings, gotchas,
   fixes, decisions, and constraints that are invisible in the code.
5. You cannot approve knowledge. Approval is a human action on the
   platform, never an agent action.

Why these rules

• Ask before any task, not just big ones: the cheap question that saves an expensive mistake is usually the one that felt too small to ask.
• Submit failures too: "this didn't work" is often the most valuable lesson in the whole memory.
• The never-store rule is enforced server-side as well, but not sending secrets in the first place is the better posture.