A Thinking Cap for Claude Code, Evaluated and Found Wanting

What it does

This repo turns 39 established mental models—from first-principles reasoning to OODA loops and TRIZ—into Claude Code skills you invoke by name. Instead of ad-hoc prompting, you get structured scaffolds for debugging, architecture decisions, risk analysis, and strategy. A meta-skill called thinking-model-router picks the right framework for your problem.

The interesting bit

The author built a replication-gated evaluation pipeline called “Elevate-or-Kill” to test whether these skills actually improve Claude’s accuracy, then published the inconvenient truth: zero skills currently pass the bar for a proven, replicated lift. That transparency is rarer than the frameworks themselves.

Key highlights

• 39 skills covering decision-making, systems thinking, cognitive bias, estimation, and innovation
• Grounded in established work (Munger, Meadows, Kahneman, Goldratt, Boyd)
• Includes meta-skills for routing and combining models
• MIT licensed; zero configuration beyond invoking by name
• Evaluation scorecard is public and brutally honest about null results

Caveats

• No skill has demonstrated a robust, replicated accuracy improvement over placebo prompts
• The closest candidate, thinking-scientific-method, showed mixed replication results and remains unproven
• Treat these as structured reasoning aids, not empirically validated performance enhancers

Verdict

Worth exploring if you want Claude Code to reason through established frameworks rather than freestyle. Skip it if you need guaranteed empirical proof that the scaffolding improves model output.