Why it matters
Most AI prompt packs claim to make models smarter but never test the claim. Claude Code Thinking Skills took the opposite approach: it ran all 39 frameworks through a length-controlled, replication-gated "Elevate-or-Kill" evaluation pipeline, then published the results, including the inconvenient finding that no skill currently holds a proven, replicated accuracy gain. That transparency is rare, and it matters when you're deciding whether to add a dependency.
The big picture
The collection covers six domains: decision-making and analysis (first principles, pre-mortem, Kepner-Tregoe), cognitive and behavioral (Bayesian, debiasing, steel-manning), systems and strategy (OODA, theory of constraints, Cynefin), problem-solving and innovation (TRIZ, five-whys-plus, scientific method), estimation and risk (Fermi estimation, red team, Lindy effect), and product (jobs-to-be-done, effectuation). Three meta-skills, including thinking-model-router, help you pick and combine frameworks.
How it works
Each skill is a SKILL.md file in Claude Code's skill system, which loads only when invoked, so unused frameworks cost nothing. You call any skill by name in the Claude Code chat, or you start with thinking-model-router and let it route the problem to the right framework. The plugin marketplace install (/plugin marketplace add tjboudreaux/cc-thinking-skills) is the recommended path; manual copy to ~/.claude/skills/ also works.
By the numbers
The closest candidate to a proven skill is thinking-scientific-method (hypothesis-differential debugging), which scored +5.3 percentage points (p=0.061, n=150) in its primary run and +8.0pp (p=0.001) in a separate replication. Per the README, because a significant replication cannot rescue a primary run that fails the p<0.05 gate, its verdict remains DIRECTIONAL-NOT-REPLICATED. A pre-registered, larger-N study is flagged as future work.
Yes, but
Treat these as well-grounded structured-reasoning scaffolds, not as proven accuracy boosters. The frameworks themselves are real and established; the empirical case for a model-accuracy lift is not yet there, and the project says so plainly. For teams that want structure in how Claude reasons through architecture decisions or incident postmortems, that scaffolding is still valuable even without a replicated accuracy gain.
The bottom line
If you want Claude Code to reason with named, established frameworks instead of ad-hoc heuristics, this is the most rigorously documented collection available. Start with thinking-model-router, pick the framework that fits your problem, and read the Elevate-or-Kill Scorecard before drawing conclusions about accuracy gains.