usage-audit
1Mines your past Claude Code sessions for what you should actually automate next — evidence, not guesses.
The story
Instead of guessing which workflows deserve a skill or a script, this reflects on the sessions already run: repeated manual work, recurring friction, anything with a stable enough shape to script. It doesn't implement anything — it hands back a ranked, evidence-backed list and you decide what's worth building. The one hard constraint baked into the design: mining a month of transcripts is a job for a cheap model, not a frontier one, so four narrow-scope subagents on Haiku/Sonnet-tier models do the raw digging, and only the final clustering-and-ranking pass touches the main session's model.
Highlights
- ✦Reflects on past sessions instead of guessing what to automate
- ✦Hands back a ranked, evidence-backed list — doesn't implement anything itself
- ✦Four narrow-scope Haiku/Sonnet subagents do the digging; frontier model only ranks
- ✦Install via git clone into ~/.claude/skills/usage-audit, or copy SKILL.md directly
- ✦Scriptable to run monthly as a standing self-audit
Want this level of depth on your own project?
The classes teach the same production techniques these repos are built with — scar tissue included.