Back to prompt library

Clean up your agent's long-term memory

Your rules keep growing and contradicting each other, and the agent stops listening.

When to use it When your agent keeps ignoring your rules

What you get

An audit of your instructions and a merge plan for you to approve.

Before you start

  • The agent can read the repo
  • Editing files is explicitly off-limits

Steps

  1. Find the working rules in AGENTS.md, CLAUDE.md and the README
  2. Group them by directory and tool
  3. Flag duplicates, conflicts, stale paths and anything that doesn't belong in memory
  4. Propose the smallest merge and wait for approval

Deliverables

  • Instruction inventory
  • Conflict table
  • Draft merge

Done when

  • Every rule has a scope
  • No conflict was silently resolved
  • No file was changed without approval

Paste this into your agent

Do a read-only audit of this project's long-term agent instructions. Find AGENTS.md, CLAUDE.md, the README and any tool-specific rules; group them by scope, and list duplicates, conflicts, stale paths and content that belongs in regular docs. Output the audit, a draft merge and the questions I need to decide. Do not edit, install, commit or publish anything. Stop once you've produced the above; don't go on to anything else.