Type: url
Original: https://zenn.dev/shintaroamaike/articles/df3ecc0ddee047
Status: generated
Language: ja
Summary
The author presents AutoHarness, a set of Claude Code skills inspired by Google DeepMind’s AutoHarness paper, which analyses a project to generate a natural-language rule file and a verification script and then updates them when rule violations or feedback occur. A small self-reported trial (n=5 per condition) found rules such as using pathlib.Path and Decimal followed only with the harness, while bug-fix rates were similar, and a June 2026 revision frames CLAUDE.md as the describe layer and hooks as the enforce layer, suggesting the verification script be wired to a PostToolUse hook.