A data-driven investigation into why best practices should live in CLAUDE.md rather than Claude Code skills. The author ran 51 multi-turn evals across 4 configurations (Superpowers, plain skills, CLAUDE.md, CLAUDE.md+hint) and found that plain skills are only invoked 6% of the time in multi-turn sessions, while CLAUDE.md guidelines are always in context. The key insight: skills and CLAUDE.md are both just prompts — the difference is reliability of delivery. Superpowers works not because of skills but because its SessionStart hook front-loads instructions, achieving 66% invocation. The recommendation is clear: put coding standards, TDD rules, and debugging protocols in CLAUDE.md (100% presence, no activation gap), and reserve skills for on-demand procedural recipes like scaffolding or migrations.
Table of contents
How skills actually work in Claude CodeThe activation gapThe multi-turn evalResultsWhy this happensSkills are recipes, CLAUDE.md is the health codeFull turn-by-turn resultsMethodologyWhat to do with thisSort: