What problem does it solve? Optimizing a small hot function often devolves into guesswork: plausible ideas get shipped unmeasured, and roughly a third of well-reasoned optimization hypotheses turn out to be wrong. This Skill provides a disciplined, benchmark-driven loop that finds the fastest implementation of a hot function by testing one hypothesis per variant, gating on correctness, and explaining results with JIT/native disassembly. ## Core Features & Use Cases - Hypothesis-driven variant benchmarking: Build a BenchmarkDotNet (or criterion/google-benchmark) harness where each variant changes exactly one thing, with a correctness gate that compares every variant against the baseline across sizes, seeds, and edge inputs. - Disassembly-guided analysis: Read JIT disassembly to explain why numbers differ — bounds checks, inlining failures, register spills, static-init guards — and follow a prioritized hypothesis ladder from algorithmic waste removal through SIMD vectorization to domain instructions like GFNI and PCLMUL. - Convergence audit and shipping: Run mechanical audits (loop-invariant arguments, size cliffs, serial-recurrence SIMD re-checks) before declaring done, then ship the winner with tiered runtime dispatch, parity tests, and honest before/after end-to-end measurements. - Use Case: A developer suspects a CRC or encoding kernel is slow. The Skill guides them to copy it into a self-contained benchmark, test variants like table lookups and AVX2 vectorization, keep a findings log of confirmed and refuted hypotheses, and port the 2x winner back with capability-gated dispatch. ## Quick Start Ask the AI to set up a micro-benchmark loop for your hot function and iterate on optimization variants until improvements fall below 2-3%.
