Your SKILL.md is production config. Test it like one.

Maintaining agent skills—instruction files for tools like Claude Code or Cursor—often involves careless edits that can lead to silent behavioral regressions. Standard text diffs fail to capture how an agent's emergent behavior changes, and traditional prompt evals often miss subtle trajectory errors. To address this, the developer behind 'skilldiff' proposes treating SKILL.md files as production configurations that require rigorous testing. The tool runs skills in a real agent harness against a fixture repository, comparing the old and new versions based on actual observations like files changed, commands run, and tool calls made. By using plain YAML assertions, developers can catch forbidden actions or unexpected tool sequences before they reach production. This approach provides a deterministic way to validate agent behavior, ensuring that updates to instructions do not inadvertently break established workflows or introduce unwanted side effects in autonomous coding environments.