Optimization Pattern — Skills

Pattern: Optimization
Component: skills.md
Version: 1.1 | Updated: 2026-07-16

---

Skill	Path / domain	Why required
Profiling & tracing	Performance skill	Attribute dominant cost before coding
Benchmark design	Perf / LLMOps cost skills	Fair A/B, warm-up, confidence intervals
Feature flags	Release / platform skill	Kill-switch and progressive exposure
Correctness oracles	Testing skill	Fail closed when faster ≠ correct

Skill activation triggers

1. Profiling — always before a hypothesis; re-run after candidate if bottleneck may have shifted
2. Benchmark design — whenever baseline or candidate results are produced
3. Feature flags — before any production-facing change lands
4. Oracles — every candidate evaluation; not optional for caching/concurrency changes

Verification that skills worked

• Profiling: dominant cost named with ≥30% share or explicit “no dominant cost” stop
• Benchmark: control and candidate share workload ID and warm-up policy
• Flags: report names the flag and disable procedure
• Oracle: zero failures, or recommendation is forced KILL