optimize-looplisted
Install: claude install-skill dustin-olenslager/claude-phalanx
# optimize-loop
Guessing is banned. Measure → find the real hot path → change ONLY it → prove
the gain or revert. This skill sets the pipeline `planned` flag.
## Precondition
Observability (§12) must exist — structured logs + spans/metrics on I/O
boundaries and use cases. If it doesn't, STOP: you cannot optimize what you can't
measure. Add instrumentation first (that's a build/maintain task), then return.
## Phases (state machine)
1. **baseline** (exit: numbers on record) — capture current metrics from
observability/benchmarks: p50/p95/p99 latency, throughput, alloc/GC, db time,
$ per op. Write them down. This is the bar.
2. **profile** (exit: bottleneck identified) — profiler/trace/flamegraph or span
breakdown to find the REAL hot path. Amdahl: optimizing a 2% cost is wasted.
Name the single dominant cost.
3. **hypothesize** (exit: ADR recorded) — state the expected gain AND the
trade-off (memory, complexity, cache staleness, consistency) in an ADR. A perf
change with no stated trade-off is suspicious.
4. **implement** (exit: change made) — ponytail + caveman; touch ONLY the hot
path. Don't rewrite cold code for elegance.
5. **benchmark** (exit: gain proven) — re-measure the SAME way vs baseline. Real,
repeatable gain beyond noise? keep. Marginal/none? **revert** — complexity
without payoff is a loss.
6. **verify / commit / memory** — as BUILD. Verify includes correctness didn't
regress (the fast wrong answer is still wrong).
## Rules
- One