Test a skill before you trust it

Run the same prompts through Claude Code with your skill and without it, and score every answer in plain code. You see if the skill helps, which rule it breaks, and whether the fix worked.

Our run, 2026-09-28. The same 6 prompts through Claude Code, with the skill and without it, scored in plain code: 5 of 6 commits followed every rule with the skill, 0 of 6 without.