An evidence-based subtraction audit for your Claude Code setup.
Most people evaluate their AI tooling by reading descriptions of it: the skill's README, the hook's comment, the memory of installing it six months ago. This skill evaluates it by measuring what ran. Your session transcripts already record every skill invocation, every MCP call, every hook fire, and every correction you typed. subtract mines them, verdicts every installed item against real numbers, and hands you a staged plan for removing what the evidence says is dead.
It is an audit, not a chainsaw. The skill is read-only, every proposed removal is an archive move with a named rollback path, and the deliverable opens by stating that its verdicts are defaults your judgment overrides.
Running the audit produces one document:
- Inventory. Everything that exists: skills, hooks, agents, MCP servers, rules, commands, memory systems. Including the broken symlinks and the things you forgot installing.
- Evidence. Per-item invocation counts, last-used dates, which hooks fire versus merely exist, and the subprocess tax every tool call pays. All mined from your transcripts with the included script.
- Verdicts. One of KEEP / FIX / TRIAL / ADOPT / CUT per item, each citing a number or an honest "unmeasured". You can reject any single verdict without invalidating the rest.
- Ablation waves. Snapshot first, then risk-ordered waves, each with an observation window, a watch-for list, a friction log, and an explicit rollback. Security deny-lists survive every wave untouched.
A gates file (GATES.md, CHECK/EXPECT style after unlazy) keeps the audit from stopping at 80% coverage.
As a Claude Code plugin:
/plugin marketplace add OpenScribbler/subtract
/plugin install subtract@openscribbler
Or manually: copy skills/subtract/ into ~/.claude/skills/, or into any repo's .claude/skills/.
Ask for it in plain words. "Audit my Claude setup", "what can I remove?", and "which of my skills do I actually use?" all trigger it. Or start from the evidence alone:
bash skills/subtract/scripts/mine-usage.sh --days 180The script targets the Claude Code default layout and fails loudly, with NOT FOUND lines and a nonzero exit, rather than reporting silent zeros. On other harnesses, references/evidence-mining.md explains how to rebuild each measurement from whatever session logs you have. The method itself (inventory, evidence, verdicts, waves) is harness-neutral even though the tooling is Claude Code-first.
The method was developed by running it for real: a six-month audit of one heavily layered setup with three generations of infrastructure, roughly 70 skills, and 11 hooks firing on every Bash call. Among the findings: the owner's single most-typed command had a symlink that had been silently dead for months, several MCP servers with zero invocations were riding along in every session, and a memory system had captured 1,100+ entries that were read back 7 times, ever. The skill encodes the process, not those results.
Design debts: the gates pattern is from unlazy; trigger-keyword descriptions, one skill per trigger, and bad/good example pairs are patterns from Theo's AGENTS.md breakdown, re-derived here.
MIT