What problem does it solve? AI agents accumulate token waste through heavy CLAUDE.md files, never-invoked skills, and context that users re-paste 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 each fix and applying it only with explicit per-edit user consent. ## Core Features & Use Cases - Consent-gated cost reduction: Reviews ranked token sinks from caveman learn report --json, proposes concrete trims for reducible sinks, and applies edits only after a yes, with a net-token-negative re-measure gate that reverts any edit failing to reduce tokens per turn. - Context offloading to cavemem: Moves recurring context blocks into cavemem storage with a cheap pointer left behind, verifying content integrity via SHA-256 and confirming recall works before trimming the source. - Honest savings reporting: Presents measured outcomes grouped by evidence strength (deterministic re-measure, controlled holdout, counterfactual replay, interrupted time series) without summing across rungs or projecting savings forward. - Use Case: Your agent's CLAUDE.md has grown to thousands of tokens and you re-paste the same project conventions every session. Run the Skill to review the Cave Score, trim dead weight, offload recurring context to cavemem, and track whether each fix actually improved token usage over time. ## 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.
