Claude Code Thinking Skills

28 Mental Models and Critical-Thinking Frameworks for Claude Code

Claude Code Thinking Skills is a collection of 28 mental-model and critical-thinking frameworks for Claude Code that give Anthropic's AI coding agent structured ways to reason about decisions, debugging, systems, risk, and strategy. Each skill packages a proven thinking framework — from first-principles reasoning to the theory of constraints — into a Claude Code skill you can invoke by name, and the whole collection is backed by a transparent, replication-gated evaluation pipeline.

At a Glance

What it is	A library of 28 mental-model and critical-thinking skills for Claude Code.
Who it's for	Engineers, founders, and analysts who want Claude Code to reason with structured frameworks instead of ad-hoc heuristics.
How to start	Install via the plugin marketplace, then invoke thinking-model-router to be routed to the right skill.
Evidence	Every skill ran through a replication-gated Elevate-or-Kill evaluation pipeline. The honest headline: zero skills currently hold a robust, replicated ELEVATE verdict — and we publish that result rather than hide it.
Entry point	thinking-model-router → START HERE

Why This Project Is Different

Most "AI prompt pack" repositories claim their content makes models smarter and never test the claim. This project did the opposite: it built an objective, length-controlled, replication-gated evaluation harness and evaluated the 39-skill legacy catalog. The result is documented openly, including the inconvenient finding that no skill yet meets the bar for a proven, replicated accuracy gain.

That rigor is the point. These skills are useful structured-reasoning scaffolds grounded in established frameworks, and the evaluation methodology is honest enough to tell you exactly how strong the evidence is. Transparency over hype is the standard here.

Features

• 28 Thinking Frameworks — A curated library of mental models for decision-making, debugging, and strategy.
• Eval-Backed and Honest — Built and tested with a rigorous, replication-gated evaluation pipeline; see the Elevate-or-Kill Scorecard.
• Battle-Tested Foundations — Grounded in frameworks from cognitive science, systems thinking, and strategic analysis (Munger, Meadows, Kahneman, Goldratt, Altshuller/TRIZ, Boyd/OODA).
• Claude Code Native — Designed specifically for Claude Code's skill system and invocation model.
• Zero Configuration — Install and invoke skills by name; no setup required.

Available Skills

All 28 shipped skills, grouped by domain. The meta-skill thinking-model-router is the recommended entry point.

Decision Making & Analysis

Skill	Description	Best For
thinking-first-principles	Break problems into fundamental truths	Innovation, challenging assumptions
thinking-second-order	Think beyond immediate consequences	Strategic decisions, policy changes
thinking-pre-mortem	Imagine failure and work backward	Project kickoffs, risk assessment
thinking-kepner-tregoe	Systematic rational process for complex analysis	High-stakes decisions, root cause analysis
thinking-reversibility	Classify decisions by reversibility (Type 1/2)	Commitment sizing, risk assessment
thinking-opportunity-cost	Evaluate choices by what you give up	Resource allocation, prioritization

Cognitive & Behavioral

Skill	Description	Best For
thinking-bounded-rationality	Make good-enough decisions under constraints	Time pressure, satisficing
thinking-socratic	Systematic questioning framework	Requirements, debugging, coaching
thinking-probabilistic	Calibrated probability estimation	Forecasting, uncertainty quantification
thinking-steel-manning	Argue the strongest opposing position	Debate, decision validation

Systems & Strategy

Skill	Description	Best For
thinking-systems	Analyze interconnected systems	Complex debugging, architecture
thinking-ooda	Rapid decision-making for dynamic situations	Incident response, competitive scenarios
thinking-theory-of-constraints	Identify and manage bottlenecks	Performance optimization, throughput
thinking-cynefin	Classify problems by complexity domain	Methodology selection, approach matching

Problem Solving & Innovation

Skill	Description	Best For
thinking-map-territory	Recognize limits of mental models	Expectation mismatches, abstractions
thinking-circle-of-competence	Know the boundaries of expertise	Delegation, learning decisions
thinking-triz	Resolve technical contradictions	Engineering design, innovation
thinking-five-whys-plus	Enhanced root cause analysis with bias guards	Debugging, incident postmortems
thinking-scientific-method	Hypothesis-differential debugging	Fault localization, ambiguous symptoms
thinking-thought-experiment	Structured imagination for exploration	Architecture, edge cases, philosophy

Estimation & Risk

Skill	Description	Best For
thinking-margin-of-safety	Build in buffers for uncertainty	Risk management, system design
thinking-lindy-effect	Older things likely to last longer	Technology selection, durability
thinking-via-negativa	Improve by removing, not adding	Simplification, robustness
thinking-red-team	Attack your own plans adversarially	Security review, plan validation

Product & Innovation

Skill	Description	Best For
thinking-jobs-to-be-done	Understand the job customers hire products for	Product development, feature design
thinking-effectuation	Start with means, not goals	Startups, innovation, uncertainty

Meta-Skills

Skill	Description	Best For
thinking-model-router	START HERE - Route to the right model by domain	Entry point for all thinking skills
thinking-model-combination	Combine multiple models for richer analysis	Complex problems, high-stakes decisions

How the Skills Were Evaluated

Honesty about evidence is a core feature of this project, so the evaluation results are reported plainly.

• The evidence base: Historical coverage is heterogeneous and remains provisional. The corrected portfolio-v1 gate made zero model calls because power, dataset, and judge-calibration requirements were not met.
• The headline result: Zero skills currently hold a robust, replicated ELEVATE verdict. All 28 shipped skills are manual-only; none is proven to improve model accuracy.
• The closest historical candidate: thinking-scientific-method recorded a provisional +4.0pp fault-localization lift in its larger-N July artifact. Although its recomputed McNemar result was significant, the effect is below the predeclared +5pp utility margin and the study has scoring, control, denominator, and raw-archive defects. It is directional evidence only, not ELEVATE.
• What that means for you: Treat these skills as structured-reasoning scaffolds, not guaranteed accuracy improvements. The audit preserves the useful frameworks while keeping unsupported performance claims out of the product.

Read the evidence yourself:

• Decision-ready audit — catalog dispositions, study citations, and explicit evidence gaps.
• Canonical evidence registry — machine-readable authority for counts, claims, and product dispositions.

The shipped catalog now contains 28 skills. Eleven unsupported or overlapping skills were removed at the evidence-backed cutover; their unique mechanisms were absorbed into surviving skills and their historical evidence remains preserved.

Usage

Once installed, invoke any skill by name in Claude Code. If you are not sure which framework fits, start with thinking-model-router:

> Use the thinking-model-router to pick the right framework for this problem
> Use first-principles thinking to analyze this architecture decision
> Apply the pre-mortem framework to this project plan
> Help me use probabilistic reasoning to evaluate this hypothesis
> Use the theory of constraints to find our bottleneck

Detailed Skill Descriptions

First Principles Thinking

Strip away assumptions to reveal fundamental truths, then rebuild solutions from basics. Championed by Elon Musk and rooted in Aristotle's philosophy.

When to use:

• Conventional approaches have failed
• You're told something is "impossible"
• Need innovation, not incremental improvement

Probabilistic Reasoning

Estimate uncertainty, update priors with evidence, and expose assumptions and calibration.

When to use:

• Estimating probabilities or likelihoods
• Interpreting test results or metrics
• Making decisions with incomplete information

Systems Thinking

View problems as part of interconnected wholes with feedback loops and emergent properties. Essential for debugging complex distributed systems.

When to use:

• Debugging spans multiple components
• Fix in one place breaks another
• Behavior seems emergent or unexpected

Theory of Constraints

Every system has exactly one constraint limiting throughput. Optimizing anything else is wasted effort. Based on Eliyahu Goldratt's work.

When to use:

• Performance optimization
• Process improvement
• Resource allocation
• Identifying bottlenecks

Scientific Method / Hypothesis-Differential Debugging

Localize an ambiguous bug by enumerating falsifiable hypotheses, ranking them by likelihood x cheapness-to-check, and making the cheapest discriminating observation first. This is the most empirically scrutinized skill in the collection (final verdict: DIRECTIONAL-NOT-REPLICATED — see evaluation results).

When to use:

• A symptom could plausibly come from several files/functions/components
• You can inspect code, logs, diffs, traces, or tests now
• You need to localize the fault before applying root-cause analysis

Cynefin Framework

Classify problems by the relationship between cause and effect: Clear, Complicated, Complex, or Chaotic. Each domain requires a different approach.

When to use:

• Choosing methodologies
• Understanding why approaches fail
• Crisis management

Jobs to Be Done

Customers don't buy products—they hire them to do jobs. Understanding the job unlocks innovation.

When to use:

• Product development
• Feature prioritization
• Understanding customer behavior

Red Team Thinking

Attack your own plans before adversaries do. The best defense is knowing your weaknesses.

When to use:

• Security review
• Pre-launch preparation
• Plan stress-testing

FAQ

What are Claude Code thinking skills?

Claude Code thinking skills are 28 reusable mental-model and critical-thinking frameworks packaged as Claude Code skills. Each one gives the AI agent a structured method — such as first-principles reasoning, probabilistic updating, or the theory of constraints — for analyzing a specific kind of problem. You invoke them by name to steer how Claude approaches decisions, debugging, and strategy.

Which thinking skill should I start with?

Start with thinking-model-router. It's the meta-skill entry point that reads your problem and routes you to the most relevant framework, so you don't need to memorize all 28. If you already know your need — for example debugging, risk, or prioritization — you can invoke the specific skill directly.

Do these skills actually improve Claude's accuracy?

No skill is currently proven to improve accuracy. Every skill was run through a replication-gated evaluation, and zero skills hold a robust, replicated ELEVATE verdict. The closest candidate, thinking-scientific-method, scored +5.3pp (p=0.061, n=150) on its fresh primary run — directional but short of the p<0.05 gate — with a significant +8.0pp (p=0.001) replication; because a significant replication can't rescue a primary that fails the gate, its verdict is DIRECTIONAL-NOT-REPLICATED. Treat the skills as solid structured-reasoning scaffolds, not a guaranteed accuracy boost.

What is the model-router skill?

thinking-model-router is a meta-skill that acts as the front door to the collection. Given a problem description, it identifies the domain (decision-making, systems, estimation, debugging, and so on) and points you to the most appropriate thinking framework. It exists so newcomers can get value without studying the entire catalog.

Are these skills based on real research?

Yes. The frameworks draw on established work from thinkers including Charlie Munger (mental models), Donella Meadows (systems thinking), Daniel Kahneman (dual-process cognition), Eliyahu Goldratt (theory of constraints), Genrich Altshuller (TRIZ), and John Boyd (OODA loop). Beyond their source theory, the skills were also subjected to this project's own length-controlled, replication-gated evaluation pipeline, with all results published in the Elevate-or-Kill Scorecard.

Are the skills free to use?

Yes. The entire collection is released under the MIT License, so you're free to use, modify, and distribute it.

Notes

• These skills are structured-reasoning scaffolds, not guaranteed accuracy improvements.
• No skill is currently proven to improve model accuracy. All skills were evaluated but none achieved a robust, replicated ELEVATE verdict.
• Skills must be installed via the plugin marketplace before use. After installation, invoke them by name in Claude Code.
• Evaluation results are publicly available, including the audit and evidence registry.