What problem does it solve? Unstructured performance tuning wastes time on guesses, promotes changes that break correctness, and cannot distinguish real wins from measurement noise. This Skill turns "make it faster" into a bounded, evidence-driven loop with a baseline, a correctness gate, and a stopping condition. ## Core Features & Use Cases - Bounded variant search: Requires a defined operation, correctness gate, single metric, baseline with noise floor, and budget before any tuning begins. - Profile-first workflow: Enforces bottleneck identification via profilers, timing, or flame graphs before generating one-hypothesis variants. - Run ledger and promotion gate: Records every variant including rejected ones, and only promotes changes that pass full correctness, reproduce their delta, and have an obvious rollback. - Use Case: A nightly import of 2M rows takes 120 seconds. Profile the stages, test batch-size and worker-count variants one variable at a time, reject a faster variant that fails correctness, and promote the winner with documented commands and numbers. ## Quick Start Ask the assistant to optimize the nightly data import job using the optimization loop, starting with a baseline measurement and a correctness gate.
