← ClaudeAtlas

tuninglisted

Reads the pack's own evidence (fire log, saved test runs), finds skills that never fire or keep failing, and proposes the smallest fix, applied only after the user's ok. Run by hand: /foma-skills:tuning.
t8jwscc7qy-sketch/foma-skills · ★ 2 · AI & Automation · score 81
Install: claude install-skill t8jwscc7qy-sketch/foma-skills
Finds which pack skills are not earning their place, from files already on disk, and fixes the worst few with the least spend. Evidence first and free; model work only for flagged skills. ## Steps 1. *Always.* Run `python3 "${CLAUDE_SKILL_DIR}/scripts/evidence.py" $ARGUMENTS` and read nothing else yet. It prints the mode (`repo` = the pack's own repository, `project` = anything else), fire counts, skills that never fired, and failing or flaky saved task runs (runs older than the skill's last edit are ignored and listed: rerun its smoke instead of judging it). Nothing flagged, or a log under 20 sessions and no failing runs: say so and stop. 2. Take at most 3 flagged skills, failing runs first. Read each one's description only; read a body only for a skill with a failing run, together with that run's reason. 3. Diagnose one cause per skill. Never fired: the description misses how the user words the task, or the skill is not needed here; ask for one prompt that should have fired it, do not guess. Failing or flaky run: the body lacks or ignores a line, or the test is wrong; say which. Fires on the wrong task: fix its "Not for" clause. 4. Propose the smallest fix per skill: one description edit (at most 250 characters), one body line, switch off, or delete. Show evidence and the exact diff. Never rewrite a skill whole; never loosen a test to make it pass. 5. Mode `repo`: after the ok, apply and check: `python3 scripts/trigger_test.py <skill> --set heldout`, plus `python3 scripts/