Auto Skill Improver

# Auto Skill Improver > Most prompt iteration is theater. This is how you stop lying to yourself about whether anything got better. Auto Skill Improver is a **benchmark-first loop** for improving Claude Code, Cowork, and OpenClaw skills through controlled mutation and honest keep/discard decisions — not prompt vibes. It is built for people who have real skill files or prompt specs and want to know — with evidence — whether a change actually made things better. --- ## The Problem The failure mode is always the same. You mutate an instruction. You read the new version. It sounds more sophisticated, so you keep it. Nothing measurably improved. Auto Skill Improver resists that trap by treating skill iteration as an empirical process, not an editorial one. --- ## The Benchmark-Mutation Loop The core method — directly inspired by Andrej Karpathy's autoresearch work — applies the same measurement discipline to skill files and prompt specs: 1. **Classify** the skill type 2. **Build a benchmark** that can actually distinguish success from failure 3. **Establish a baseline** before you change anything 4. **Mutate one thing** at a time 5. **Keep only what improves the score** 6. **Stop** when the benchmark saturates or stops teaching you anything The principle we borrowed from Karpathy is not "let an AI improve itself." It is the more disciplined idea underneath it: use a bounded environment, a real benchmark, a repeatable loop, and make decisions based on evidence — not instinct. --- ## Where It Works Best Auto Skill Improver produces real signal on skills with: - **Ambiguous output contracts** — the skill doesn't clearly define what good looks like - **Missing fallback behavior** — the skill collapses when a dependency is unavailable - **Conflicting instruction layers** — one rule contradicts another - **Dependency or portability problems** — the skill breaks outside its original environment - **Weak evidence discipline** — the skill invents certainty when d