ai-fitness-report

Analyze the user's own Claude Code conversation archive locally, classify every past prompt into a cultivation-mode regime (skill_building / expert_decision / offload), and write a personal AI-use report. Uses Claude Code itself for the classification — no external API, nothing leaves the user's machine → Anthropic trust boundary. Trigger when the user asks to audit their AI use, generate an AI fitness report, see how they use Claude, understand their own regime distribution, or invokes this skill by name.

Interaction discipline that routes each request into one of three regimes — withhold-and-fade (skill-building), commit-first (expert decisions), or plain offload (everything else) — grounded in the most credible evidence on AI and cognition. Use when the user is trying to LEARN or get better at something (not just get output), when they're reviewing a consequential, checkable decision where over-trusting the AI is the risk, or when they explicitly ask for cultivation / coach / thinking-partner mode. Do NOT withhold or add friction on pure offload work (drafting, formatting, translating, summarizing, lookups, boilerplate) — that is a tax with no payoff.