Spent the day digging into how skills could be extracted automatically from Claude Code session transcripts instead of being hand-written. Sessions live as JSONL files under ~/.claude/projects/<sanitized-path>/<uuid>.jsonl — user and assistant entries with tool calls inline — and even with the tool-call noise stripped, each one only weighs in at a couple of thousand tokens of text. A whole project’s worth of sessions fits comfortably in context, which changes what’s feasible.

I cloned AutoSkill and walked through its three-stage pipeline (plan → expand → compile) and the prompt engineering behind it. The extraction philosophy is the interesting part: only lift skills from concrete, reusable user requirements, treat assistant replies as untrusted evidence, and lean toward missing a one-off over saving a generic task. Worth porting. The runtime itself is not — it requires an external LLM API key, and its filesystem-backed storage plus embedding dedupe is overkill at the scale of <100 skill files. The recommendation written up in .assist/spike-75.md is to port the prompts and use Claude Code itself as the extraction LLM via an assist skill extract CLI command plus a /skill-extract skill that runs the approval flow. Zero external dependencies, and the cost is already inside the session.