# Critical Thinking > Think clearly and decide well under uncertainty. Use this skill whenever you reason through a hard problem, analyze or critique an argument, evaluate a claim or a statistic, weigh evidence, forecast or estimate odds, make or pressure-test a decision, diagnose a recurring failure, red-team an idea, or check your own reasoning for bias. Grounded in 21 books on judgment and reasoning (Kahneman's "Thinking, Fast and Slow", Galef's "Scout Mindset", Tetlock's "Superforecasting", Silver, Duke, Meadows, Dennett, de Bono, Deutsch, Sagan, Pigliucci, Bergstrom and West's "Calling Bullshit", Huff, Rosling's "Factfulness", Browne, Weston, Warburton, Adler, Hamming, Bazerman, Hastie). Reach for this for "think through", "is this true", "is this argument sound", "evaluate this claim", "what could go wrong", "should I", "what are the odds", "red-team this", "find the flaw", or any judgment, analysis, forecasting, or decision task, even when no book is named. The deep tool libraries in references/ load on demand, by task. - Skill: `johna2an/critical-thinking-2` (Agent Skill, multi-file: 27 files) - Install (CLI): `npx skillmds@latest add johna2an/critical-thinking-2` - Raw SKILL.md: https://api.skillmd.com/api/skills/johna2an/critical-thinking-2/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Productivity - Author: Johna2an (https://skillmd.com/u/johna2an) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/johna2an/critical-thinking-2 --- Think clearly and decide well. The first answer your mind offers is a hypothesis, not a verdict. This skill is a toolkit for catching your own errors, reasoning cleanly, quantifying uncertainty, stress-testing claims, thinking in systems, generating real alternatives, and deciding under uncertainty. It carries the working tools of 21 books on judgment. The directives below are the operating manual; `references/` holds the deep tool libraries and a per-book teaching library. ## How to use this Apply the phases that fit the task. Phase 1 (frame) and Phase 2 (check yourself) run on almost every hard question; the rest engage as the problem demands. Load a reference file only when the task needs that depth, and load a per-book file only when a task calls for that book. Read only what the task needs. When two directives seem to conflict, scope to purpose (see "When the directives collide"). ## Phase 1. Frame the question - **State the exact question.** Pin the issue and the conclusion in plain words before reacting. Sort the issue: is it descriptive (what is) or prescriptive (what ought)? The two need different evidence. - **Catch the swap.** Ask what you were actually asked and what easier question your mind answered instead. A hard call that suddenly feels easy and emotionally clear is a substitution (attribute substitution). - **Set the boundary on purpose.** Decide what is inside the frame and what is out, then deliberately pull back in the delayed effects, distant people, and long-term consequences you first excluded. ## Phase 2. Check yourself (bias and the scout stance) - **Ask "Is it true?" rather than "Can I believe it?" or "Must I believe it?"** Notice which question you were running. You bend evidence toward what you want and away from what you resist, so flip the standard and re-check. - **Hold beliefs as probabilities and your identity lightly.** State a number rather than a certainty, and arrange things so being wrong costs you nothing. A belief welded to your identity gets defended instead of tested, so hold it lightly. - **Run the selective-skeptic test.** If this same evidence pointed the other way, would you still find it credible? A different standard for friendly and unfriendly evidence is a double standard. - **Hunt disconfirming evidence first.** Record every fact that cuts against you the instant you meet it, because the mind drops inconvenient facts faster than convenient ones. Lean into confusion rather than explaining it away. ## Phase 3. Reason and argue cleanly - **Separate the conclusion from the reasons, and let reasons come first.** A conclusion with no reasons is a bare assertion. A conclusion chosen first with reasons scrambled in afterward is the tell of motivated reasoning. - **Surface the hidden premise and the value judgment.** Most arguments leave a premise unstated, and the weakness usually hides there. Name the value the arguer had to rank first to reach the conclusion. - **Steelman before you strike.** Re-express the opposing view so well its holder would endorse your version, then engage the strongest form (Rapoport's rule, the principle of charity). - **Name the fallacy.** Keep the high-frequency roster loaded: ad hominem, straw man, false dilemma, slippery slope, begging the question, post hoc, equivocation, red herring, and appeal to questionable authority. (`references/argument-and-evidence.md`) ## Phase 4. Quantify and forecast - **Base rate first, then adjust.** Find the reference class and anchor on how often things like this happen in general (the outside view) before you touch the vivid specifics of this case (the inside view). - **Put a number on it and decompose the impossible.** Translate "likely" into a probability. Fermi-ize a hard estimate into parts you can each guess, then recombine. - **Update in small steps and keep score.** Nudge on genuinely diagnostic evidence; avoid both freezing on a prior and lurching on a flashy detail. Log forecasts and grade them, because a forecast you never score teaches nothing. (`references/forecasting-and-decisions.md`) ## Phase 5. Stress-test (red-team and call bullshit) - **Interrogate the black box.** Audit the data going in and the result coming out without needing the math. Ask "compared to what?", "out of what?" (the denominator), and "which average?" before trusting any statistic. - **Read the axes before the shape, and demand effect size with every p-value.** A tiny real bias across a huge sample manufactures statistical significance. - **Default to the prosaic and demand evidence proportional to the claim.** Rule out misperception, error, coincidence, and selection before reaching for the extraordinary. Ask what observation would prove the claim false. - **Prefer explanations that are hard to vary.** An account you could reshape to fit the opposite outcome just as comfortably is weak, however confident it sounds. (`references/argument-and-evidence.md`, `systems-science-and-creativity.md`) ## Phase 6. Think in systems - **Look for the structure behind it.** When a problem keeps recurring, stop hunting the person who caused this instance and map the stocks, flows, feedback loops, and delays that make the behavior the system's default. - **Aim high on the leverage hierarchy and audit your metrics.** Push on goals, rules, and information flows over tweaking mere numbers. The system delivers exactly the indicator you specify, so you get what you measure. (`references/systems-science-and-creativity.md`) ## Phase 7. Generate alternatives - **Set a quota and suspend judgment.** Produce several readings even after a good one arrives, because the first acceptable answer suppresses the rest. Let a wrong-looking idea be a stepping stone. Use reversal, analogy, and a random input to break the pattern, then find the hidden option C. Generate first, then switch to analysis to test what you produced. (`references/systems-science-and-creativity.md`) ## Phase 8. Decide under uncertainty - **Judge the decision on what was knowable.** Grade a choice by the information and reasoning available at the time, never by how it happened to land (resulting). Frame the choice as a bet that names the alternative futures and the opportunity cost. Ignore sunk costs and decide on future consequences alone. Run a premortem, aggregate small positive-expected-value bets, and hold extra caution for the unquantifiable tail. (`references/forecasting-and-decisions.md`) ## When the directives collide When two moves seem to conflict, scope the choice to the purpose and the conditions in front of you. - **Trust or distrust intuition: diagnose the domain.** Let intuition lead only when the environment is regular enough to hold learnable patterns and you got rapid, clear feedback while learning it (a surgeon's hands, a firefighter's read of a room). In noisy, low-feedback domains (markets, geopolitics, one-off strategy), distrust the gut and substitute base rates, checklists, or a simple model. - **Use falsifiability as a flashlight.** Ask "what would change your mind?" and treat a confident "nothing could" as a strong tell. Score a claim across several marks (testable predictions, response to disconfirmation, progress, details pinned by reality) rather than passing or failing it on one rule. Demarcation is a gradient. - **Quantify the tractable middle, stay humble at the tail.** Forecast where the domain has feedback and refuse to extrapolate past the horizon where skill decays. Separate epistemic uncertainty (knowable, push on it) from aleatory uncertainty (irreducibly random, stay cautious). - **Separate generation from analysis by phase.** Generate with judgment suspended, then switch to rigor to test. The two sabotage each other only when run at once. - **Spend rigor where it pays (stakes-triage).** Reserve full deliberation for choices that are important, hard to reverse, unfamiliar, or numeric. Let fast judgment handle low-stakes, reversible, high-frequency calls. Refuting a bad claim costs far more than making one, so aim your scrutiny at the load-bearing claim. ## NEVER - Grade a decision by its outcome (resulting), or rewrite a prediction after the fact (hindsight). - Trust a vivid, coherent story as evidence. Coherence is not proof, and adding a plausible detail lowers probability (the conjunction fallacy). - Accept a correlation as a cause without listing the rival explanations first. - Read a single statistic without its denominator, its sample size, and its comparison group. - Let a number already on the table set your estimate (anchoring); generate an independent one first. - Defer to a claim because it wears numbers, jargon, or a credential; scrutinize those harder. - Judge a probabilistic forecast by which side of 50% it landed on (the wrong-side-of-maybe fallacy). - Honor a sunk cost or escalate commitment to justify a past choice. - Treat a small sample, a streak, or one extreme result as a stable signal (regression to the mean, the law of small numbers). - Accept a claim that forbids no possible observation. If nothing could refute it, it explains nothing. - Settle a charged decision in a single frame. If your preference flips when the same options are reworded as a loss instead of a gain, you do not really hold it. ## Go deeper Load a reference when the task calls for it. Do not read them all up front. - `references/biases-and-judgment.md`: the bias catalog and the self-checks that counter it (Kahneman, Galef, Bazerman, Hastie, Rosling, Duke). - `references/argument-and-evidence.md`: argument analysis, evidence grading, statistical skepticism, rhetoric and fallacy detection (Browne, Weston, Warburton, Huff, Bergstrom and West, Dennett, Adler). - `references/forecasting-and-decisions.md`: probability, calibration, forecasting discipline, and decisions under uncertainty (Tetlock, Silver, Duke, Hastie, Bazerman, Rosling). - `references/systems-science-and-creativity.md`: systems thinking, scientific epistemics, and lateral generation (Meadows, Sagan, Deutsch, Pigliucci, de Bono, Dennett, Hamming). - `references/books/`: one teaching file per book. Load a single file only when a task wants that book's depth (for example `thinking-in-systems` for a recurring organizational failure, `calling-bullshit` to refute a quantitative claim, `superforecasting` to build a calibrated prediction). Reason like a scout: the goal is to see what is true rather than to win the point. Hold every belief as a probability, hunt the evidence that would change your mind, and judge your thinking by whether it tracks reality.