muse-dev-patternslisted
Install: claude install-skill wlsdks/muse-agent
# muse-dev-patterns — the recurring slice discipline
Distilled from 200+ loop fires (docs/goals/loops/*), the backlog's dedup
ledger, and live sessions. `improve-muse` decides WHAT to build and drives
the end-to-end cycle; THIS skill is the HOW at each step — the moves that
kept working, and the failure classes each one exists to prevent.
## 1. When a gate/test fails: triage BEFORE fixing
A red gate has exactly three causes. Fixing before classifying produces the
worst class of bug — "fixing" the assertion when the product regressed, or
"fixing" the product when the assertion went stale.
1. **Reproduce and read the determinism.** Run the failing case k=3–5
times. `0/5 with the identical wrong output` is STRUCTURAL (something in
the path changed); `3/5` is stochastic (model variance, load, race).
Structural failures deserve root-cause; stochastic ones deserve
`MUSE_EVAL_REPEAT` and a look at thresholds/timeouts.
2. **Is it pre-existing? Prove it with stash.**
`git stash && <run the failing test> && git stash pop` — if it fails on
clean HEAD, it is main's rot, not your change. File it as its own T0 fix
(never bundle it silently into your slice's commit).
3. **Date the flip.** `git log --format='%ci %h %s' -S '<changed token>' --
<file>` finds WHEN the behavior changed and WHICH commit did it. Read
that commit's intent before touching anything.
4. **Classify: bug vs stale expectation.** If the flip commit was a
DELIBERATE product improvement, th