What problem does it solve? Performance tuning often relies on intuition, producing unverified speedup claims and changes that regress behavior. This Skill replaces guesswork with a scientific loop: it establishes a trustworthy baseline, profiles the dominant bottleneck, generates falsifiable hypotheses, and banks only changes a benchmark proves faster and a correctness check proves safe. ## Core Features & Use Cases - Measurement-first pipeline: No optimization runs before a reproducible baseline exists; every speedup number comes from a deterministic measurement harness, never from model narration. - Domain-generic contracts: A per-run Optimization Contract defines the metric, measurement command, correctness command, and allowed change space, so the same pipeline optimizes source code, container images, LLM hyperparameters, build pipelines, or composite systems. - Gated iterative loop: Approval gates guard the contract, hypothesis backlog, and each banked change, with hard blocks on noisy benchmarks (CV ceiling) and sub-threshold gains. - Use Case: Ask to make a tokenizer 2x faster with p95 under 120ms — the skill profiles the hotspot, ranks Amdahl-bounded hypotheses, tests four diversified candidate implementations in isolated worktrees, and banks only measured winners. ## Quick Start Ask the assistant to optimize a specific target with a measurable goal, for example: optimize src/parser/tokenizer.py to be 2x faster with p95 latency under 120ms.
