evolve

What problem does it solve? Manually optimizing a hot function is slow and speculative, and naive automated evolution cheats by returning wrong answers quickly. This Skill runs OpenEvolve against a function with a rigorously designed judge — a frozen reference implementation, seeded fuzzing, and speedup-ratio scoring — so evolution finds genuinely faster code instead of code that does less work. ## Core Features & Use Cases - Fitness screening: Decides up front whether a task suits evolution (pure function, measurable metric, frozen meaning, fast evaluation) and refuses bad fits like UI, stateful code, or taste-based changes. - Judge design rules: Encodes the harness pattern — hard-coded expected answers, same-seed fuzzing across candidates, in-process speedup ratios, and compiler output fed back into the prompt as artifacts. - Guided run and verification: Covers config setup (claude_code provider, language/file_suffix), baseline noise measurement, background runs with logs, and a strict post-run procedure: understand the winner, re-measure, port in repo style, run the full test suite, and let the user decide. - Use Case: You have a hot string-normalization function in a Swift package. The Skill builds a lab directory with the current implementation, a judge harness, and an evaluator, runs 20 OpenEvolve iterations via the Claude Code CLI, and hands you a verified faster candidate with before/after timings. ## Quick Start Ask the AI to use the evolve skill to find a faster implementation of a specific pure function in your project.