What problem does it solve? Performance work often fails because teams optimize the wrong layer, compare inconsistent measurements, or ship speculative caches and indexes without proof. This Skill enforces a measurement-first method: fix comparison conditions, capture a baseline, identify the dominant constraint, apply one minimal fix, and rerun the identical workload to prove the improvement. ## Core Features & Use Cases - Baseline and comparison discipline: Locks version, environment, data, concurrency, and warm/cold state before measuring, with enough repetitions to expose variance and tail latency. - Surface-specific evidence maps: Guides evidence collection for browser (Web Vitals, traces), service (p50/p95/p99 latency, saturation), database (query plans, scans), memory (heap profiles), build (bundle chunks), and prompt-cache (provider token usage) scenarios. - Verified improvement only: Rejects unproven optimizations, requires absolute and relative change reporting, and keeps only wins backed by reruns and correctness checks. - Use Case: A checkout p95 regression triggers an alarm. Use this Skill to baseline against the last healthy window, trace the repeated database query causing the bottleneck, fix only that query, and confirm p95/p99 improvement without correctness errors. ## Quick Start Use the performance optimization skill to diagnose the p95 latency regression in the checkout flow and verify a fix against the production baseline.
