What problem does it solve? AI agent configurations accumulate token waste: bloated CLAUDE.md files, never-invoked skills, and context blocks re-pasted every session. The caveman learn command measures where tokens go, but acting on its findings requires careful, reversible edits — this Skill closes that loop by proposing and applying each fix only with explicit per-edit user consent. ## Core Features & Use Cases - Reducible sink trimming: Proposes concrete diffs for heavy CLAUDE.md files or unused skills, applies them only after a yes, and re-measures to enforce a net-token-negative gate (reverting if tokens do not drop). - Recurring context offload (cavemem_offload): Moves blocks re-established every session into cavemem storage with a cheap pointer left behind, verifying content integrity via sha256 and confirming recall works before trimming the source. - Longitudinal outcome tracking: Records applied fixes via caveman learn applied so later scans report improved, unchanged, regressed, or insufficient_data verdicts with exact revert paths. - Use Case: After running caveman learn, you discover a 200-line style guide in CLAUDE.md costs tokens every turn. The Skill proposes offloading it to cavemem, verifies recall returns the content, writes a pointer, and confirms the per-turn token count actually decreased. ## Quick Start Ask the agent to run the caveman learn report and walk you through fixing the top token sinks one approved edit at a time.
