Last Updated: 2026-07-25
A gym for your critical thinking. The AI is the coach; you do the work.
The Web game is a third entry point, not a replacement for the original skill. All three paths share this repository as their source, but they currently have different runtimes and content breadth:
| Entry | Best for | What runs today |
|---|---|---|
| Claude Code plugin / CLI | Dynamic practice with an AI coach | Four canonical modes, 23 verified Expedition packs, and the on-disk Passport |
| Claude.ai Skill ZIP | No local install; upload and chat | Four modes and the pack library, with a copy-paste Passport instead of local files |
| Web Casebook | Players who do not use a CLI | Four fixed bilingual excerpts plus an optional server-backed zh-TW Daily case; no model call |
The portable single-file Markdown build is a compatibility variant of the chat path. It includes Drill, Scene, and Detective, but not Expedition or an on-disk Passport.
New: bilingual Casebook + Daily Dispatch. web/ opens with a
real domain choice instead of dropping everyone into an education case. Its UI
and four fixed excerpts switch between English and Taiwan-oriented Traditional Chinese.
Drill, Scene, the Forecaster excerpt of Expedition, and the two-step Detective
dependency demo each have a finishable browser flow, using an original 1890s
mystery setting, generated character/background art, visual-novel dialogue,
original background music, answer effects, a local-first Passport, and
responsive layouts. These authored excerpts demonstrate selected mechanics;
they are not runtime-generated full Skill sessions and have not received an
independent human validity sign-off. A fifth, independent Daily entry can rotate
zh-TW cases through Vercel: public prompts stay in the repo while unpublished
cases and answer records stay in Private Blob. If that API is absent, all four
fixed excerpts keep working. Try the current web demo.
The checked-in Daily rotation cycles fourteen days. The first seven mirror the public fixed excerpts and therefore demonstrate the server-response flow rather than anti-cheat integrity. The second seven are an independently authored source-credibility case family: their public files carry prompts and choices only, and their answer records live solely in private storage — the production pattern.
python -m http.server 4173 --bind 127.0.0.1 --directory web
# open http://127.0.0.1:4173/Install in 30 seconds (Claude Code CLI / VS Code / JetBrains):
/plugin marketplace add Imbad0202/critical-thinking-for-humans
/plugin install critical-thinking-for-humans
Or install it as a plain Claude Code skill with one command:
git clone https://github.com/Imbad0202/critical-thinking-for-humans ~/.claude/skills/critical-thinking-for-humansThen start any session and say drill, scene, expedition, detective, or
just describe what you want to practice. (No Claude Code? A single-file portable
edition runs in any frontier model's
chat window.)
Local Passport prerequisite: the concurrency-safe on-disk Passport writer
requires a supported Node.js release, version 22 or newer, on PATH
(node --version). Current native Claude Code installations do not provide that
system runtime themselves. Without it, all four coaching modes still work, but
the skill pauses local Passport recording, says why, and never falls back to an
unsafe direct write. The claude.ai and portable editions have no on-disk
Passport and do not need Node.js.
This release promotes source credibility to the fourteenth canonical reasoning structure and adds credibility-specific generation and behavior gates. Scene grows from ten to twelve fallacy lenses with motte-and-bailey and gambler's fallacy, including reverse guards for legitimate narrowing, dependence, parameter learning, future-block probability, and regression to the mean; it also makes complete steelman-before-correction ordering explicit. The local Passport writer now serializes concurrent checkpoints and fails closed on unsafe paths or a missing Node.js 22+ runtime. The seven Daily source cases also remove option-length answer cues. No Passport schema or manipulation-taxonomy bump. Full details in CHANGELOG.md.
You can be the sharpest reasoner in your own field and still get fooled by the exact same logic the moment it shows up somewhere unfamiliar. Rigor does not transfer on its own.
A researcher who would never accept a causal claim without a control group, who would tear that paper apart in review, gets a security alert saying "unusual login detected, verify now" and clicks before asking the one question their own discipline trained them to ask: where is the evidence for that premise? The structure is identical. The eye just does not carry over.
So this is a gym for the reps. The AI is the coach and you do the work. You drill argument structures, pull apart a real situation in scene, audit reasoning you could not have produced on an expedition, or crack a layered case flaw by flaw as a detective. Every item names the underlying structure, so the move you just made gets a name rather than staying an unrepeatable intuition. Whether a named move carries from a journal to an email to a contract is the open question the field has not settled; this is a place to practice the move, not a promise that it transfers. All four modes track your patterns over time in a passport on your machine, showing you what no single session can: your longitudinal blind spots.
Why spar with an AI? Because criticizing other people is easy and criticizing yourself is not. An AI removes the audience. Nobody hears you get it wrong. If you think it is too blunt, push back, argue, close the window, go on with your day. That zero social cost is what makes it safe enough to look at your own blind spots, the one thing this practice needs and real life rarely gives you room for. An AI coach is the cheapest place in the world to lose face.
One honest limit, stated where you can see it rather than buried: the material is AI-generated at runtime, so it can be wrong. drill is the sharp case. Its items have a keyed answer and the coach rules right or wrong against it, but that key is written and audited by the same model, in the same session. The reverse-solve gate (modes/drill.md) makes the model re-solve its own item with fresh eyes, which catches sloppy distractors; it does not make the key independently verified. No human and no second model signs off, so on a subtle item the coach can confidently key the wrong option. Treat it as a sparring partner, not an oracle: a ruling that does not survive your own argument is a move to push back on, and pushing back is the practice, not a deference test. The coach is required to answer the objection on its merits or concede the item is flawed (modes/drill.md, "honor a challenge to the key"), never to defend the key by authority. It is practice, not advice.
Three commitments, non-negotiable:
-
Not neutral about reasoning quality. Factual errors, internal contradictions, and evidence misreadings are corrected in every mode, whoever's frame they sit in. "It's just a perspective" never exempts a logical error from correction.
-
No verdicts between value frames. Competing interpretations are laid out steelmanned, each in its strongest defensible form, and never ranked. The coach adjudicates reasoning, not worldviews.
-
The endpoint is commitment, not permanent neutrality. Every session closes with you taking a defensible position and defending it against the strongest objection. Endless hedging is not the goal.
A gym doesn't tell you where to walk. It puts the weight in your hands, and it will absolutely tell you when you stumble.
drill. Judge stance. Original argument-analysis items built in your field around one of fourteen structures (necessary assumption, alternative cause, reverse causation, coincidence/timing, sample selection, proxy mismatch, evidence sufficiency, base rate neglect, regression to the mean, Simpson's paradox, circular reasoning, hasty generalization, weak analogy, source credibility). You commit an answer before any analysis appears; then every option is dissected and the transferable structure named. You leave with the skeleton, not just the answer.
scene. Socratic stance. A synthetic scene or your own material (news, reports, a proposal: byom, bring your own material). All six interpretive frames laid out and steelmanned. The camera turns on your own reading too: your interpretation is examined with the same rigor as the others. Ends with you committing to a position and defending it against the strongest objection. Scene also runs a separate fallacy-recognition track: bring an argument and the coach helps you judge whether it commits a named fallacy (false dilemma, ad hominem, strawman, a fallacious appeal, equivocation, false analogy, whataboutism, slippery slope, genetic fallacy, no true scotsman, motte-and-bailey, gambler's fallacy) with three honest rulings (it is a fallacy, it is not, or there is not enough context to say), and a guard against the opposite mistake of crying "fallacy" at a sound argument. A third configure track inverts the whole exercise: before any analysis appears, you design the information request and verification plan for a decision, then the case reveals which asks were load-bearing, which were noise, and what your plan missed — because outside a gym, nobody hands you the relevant evidence pre-assembled. One track runs at a time; frames are never ranked, but a fallacy in the form of an argument is named plainly.
expedition. Guide stance. Problems that stayed open for decades and fell
to AI-class search, run only from curated packs with verified solutions, never
improvised. You are not expected to solve them; you audit, climb, or forecast,
and what you train is the human-executable translation of machine advantages:
decomposition, representation shifts, small-case probes, pre-committed kill
criteria. Say expedition or impossible.
(Installed packs live in expeditions/. Without a pack the mode says so
honestly and routes you to drill or scene; the authoring spec is
expeditions/PACK-SCHEMA.md.)
detective. Guide-and-judge stance. A single runtime-generated case in your
own field, worked as a multi-layer escape room: each layer hides one keyed flaw,
and catching it yields a concrete key (a number, a name, a threshold) that is the
necessary input to the next layer's puzzle. You carry each discovery downward
until the final layer's key is the case's truth. It sits between scene and
expedition: it judges (the flaws are real, against a frame the case states up
front), but the material is generated fresh each time rather than drawn from a
verified pack. Say detective (or 查案 / 破案 / 偵探).
(Detective generation is the most demanding work in the skill; recommended on an
opus-class or stronger model. On weaker models it degrades to fewer layers or
declines to start rather than shipping a broken case.)
Not every field plugs into every mode, and the gym says so rather than
pretending otherwise. The fourteen drill structures are seven
causal-inductive, three statistical, three formal/inductive, and one
source-evaluation tool. They need material where someone offers evidence for
a conclusion and a single gap can be engineered. source_credibility asks
how much evidential weight a source or observation report warrants from
documented credibility-relevant features; origin alone never settles whether a
claim is true or false. When you name a domain that does not fit drill, the
coach stops, says why, and points you to the path that does. It never
silently re-skins another field's material under your domain's name.
Fields that fit drill directly (anything built on empirical or causal argument):
- Physics, chemistry, biology, earth science. Not the laws or theorems
themselves (those are deductive; see below), but experimental reasoning: a
paper claims A caused B without controlling a third factor
(
alternative_cause), a sample that excludes the cases most able to refute it (sample_selection), a metric that measures activity instead of the claimed outcome (proxy_mismatch). You drill the gap between a study's evidence and its conclusion. - Human geography, economics, the social sciences, policy, medicine, education, business. Causal and inductive claims are the native material.
- Manipulation recognition. Its own built-in domain (below).
Fields that need scene mode, not drill (where there is no single defensible answer to key against):
- Music, art, literature, film, design. Aesthetic and interpretive judgement. Redline 1 forbids the gym from adjudicating a value frame, so it will not hand you "is this piece good?" as a keyed item, because that answer does not exist. Instead, scene mode dissects discourse about the art: bring a review, a curatorial statement, a critical essay, and the coach lays out interpretive frames over its argument ("this essay calls the style a decline. On what evidence? What does it not see?"). You audit the reasoning, never the artwork.
- Pure mathematics, formal logic, theoretical CS. Deductive systems. A proof is valid or it is not; there is no inductive gap to engineer. Scene mode handles these as non-social analytical material (modes/scene.md): an adapted lens set (step validity, hidden premises, reversibility, edge and degenerate cases, quantifier scope, necessary vs sufficient) dissects a flawed derivation step by step.
- Pure ethics / aesthetics / definitional disputes. Value and definition questions, not evidential ones. Same routing: scene mode, frames laid out, no verdict on the value position itself.
This boundary is the tool being honest, not a limitation to work around. A gym that trains critical thinking must not pretend that "which symphony is better" or "is this conjecture true" have the same shape as "does this study's evidence support its claim." Different shapes need different modes, and for the fields drill cannot key, scene is where you bring your own material and still train every frame.
(Expedition mode is orthogonal to this: it runs on verified packs regardless of your home field. What it trains is auditing a chain you could not have produced, not your domain's content.)
Name it at intake (sales pressure, scam scripts, political rhetoric, relational manipulation, 話術辨識) and the gym has you practice spotting the technique rather than memorizing the story: false scarcity, love bombing, gaslighting, whataboutism, and ten more, each with a counter-question you keep. Recognition only, by hard rule: the coach never writes, improves, or personalizes a manipulative script for use on a real target, in any framing. Political material samples techniques across the spectrum; the technique is adjudicated, never the position.
Install with either method above (plugin or the
one-line git clone), then start any Claude Code session and say drill,
scene, expedition, or detective, or just describe what you want to
practice.
On first run the coach asks the three choices, then routes you to your mode; later sessions confirm your profile in one line.
Intake (three choices):
- Field. Any domain in your own words; multiple fields or "no preference" accepted.
- Support level.
intro(high scaffold),standard, oradvanced(open construction). - Feedback style.
direct(error stated plainly) orcushioned(same fact, more context). The fact of the correction is non-negotiable; the delivery is yours to choose.
Safe words (always honored, announced once at session start):
"stuck" (demonstration on a parallel case), "hint" (one scaffold step),
"enough for today" (graceful close), "forget this one" (discards pending events only; items already checkpointed stay on disk).
The tool adapts its delivery; it never adapts its standards.
Ask any image or text model to generate "a principal talking with a teacher" ten times. Tally the gender, titles, and who speaks first across the ten outputs. Bring the distribution to a scene session and examine what you find. You are not testing whether the model is biased. You are practicing the examination: which frames can account for the distribution, which cannot, and what evidence would defeat each reading. Whichever way the tally lands, it is fully discussable material.
The passport lives at ~/.ct-gym/ on your machine. It records:
- drill: hit/miss per structure ID, per session and longitudinally.
- scene: process coverage. Which frames were raised, steelmanned, whether the camera turn was completed, whether you made a closing commitment.
- expedition: process record per pack. Role taken, which disciplines you fired unprompted, whether you articulated the breakthrough.
- detective: process record per case. Layers solved of total, eggs found of total, confirmed false positives, any correct objections the case key had missed, and the main-flaw structures hit (fed into the same per-structure miss-log drill uses).
- The longitudinal mirror: after enough sessions the passport summary will show you patterns no single session shows. For example: "4 of your last 5 misses are sample_selection," with a citation to the record so you can read it yourself. The pattern appears in your passport summary, not as an unprompted coach callout.
The passport's relevant content enters the model context when used. You can
run show passport, delete passport, or pause recording at any time.
A sensitive BYOM session writes no passport events at all, not even your closing commitment, unless you explicitly ask.
Current Claude Code sessions serialize checkpoint batches through the bundled
local writer, so simultaneous sessions on the same machine do not overwrite one
another's events. A lock or write failure is fail-closed: the session keeps the
checkpoint pending and tells you it was not saved. Each session also carries a
local deletion-generation token: delete passport rotates it before removing
the record, so an older session cannot silently recreate pre-deletion events
from its pending buffer. Returning-user and show passport reads go through
the same locked helper and reject symlinks and non-regular files instead of
opening the event log directly.
The argument anatomy comes from informal logic (Toulmin, Ennis). The facilitation model treats ill-structured problems as requiring defensible judgment, not right answers (King & Kitchener), and hunts assumptions the way adult educators do (Brookfield, Mezirow). The difficulty system applies scaffolding and desirable-difficulty research (Vygotsky, Bjork), and items always name their transferable structure on the hypothesis that labeling aids transfer. That hypothesis has partial support for near transfer and abstraction (the analogy literature) and remains contested for the far transfer this tool is reaching for; the meta-analytic picture on teaching critical thinking is a positive but moderate, highly heterogeneous effect (Abrami et al.), not a settled result the tool can borrow as its own. Item engineering draws on the long-established form of critical-reasoning assessment (the assumption / strengthen / weaken / inference question types common to standardized reasoning tests generally) as a structural reference, never as a content source.
This is an independent project, not affiliated with, sponsored by, or endorsed by any standardized-test publisher or assessment organization. It is not a test-preparation product and makes no claim to improve any examination score.
All practice items are original. CLI/Skill sessions generate items at runtime; the Web Casebook instead uses clearly labeled, authored fixed excerpts and a versioned Daily demo rotation. The project does not reproduce, adapt, imitate, or distribute the questions of any published test, and references the types of reasoning task (assumption, strengthen, weaken, inference) only as a generic structural form, not as anyone's proprietary content. Any third-party names that appear in this repository (products, organizations, researchers, journals) are used nominatively, for identification, citation, and commentary only. Claude, Claude Code, and claude.ai are referenced solely for runtime compatibility; the project is not affiliated with, sponsored by, or endorsed by Anthropic.
All institution and person names in example items and scenes are fictional; any resemblance to real entities is coincidental. Real individuals and organizations named in the expedition packs are cited for their published, public work only, and every characterization is grounded in the cited source.
This is a practice tool, not an outcome guarantee. It is designed for deliberate practice of reasoning skills and makes no claim to improve any examination score or any academic, professional, financial, medical, or legal outcome.
Educational use only. Nothing the tool produces is legal, medical, financial, psychological, or safety advice. The manipulation-recognition material teaches recognition, not response; for an active scam, a controlling relationship, or any situation involving immediate danger or real loss, consult the appropriate qualified professional or local emergency and crisis resources.
These notes describe the project's own practices; they are not legal advice.
Everything in this repository (the modes, shared floor, expedition packs, docs,
and the build and lint code under scripts/) is licensed under CC BY-NC 4.0
(see LICENSE).
The repo is the single source of truth; the claude.ai-uploadable zip is generated, never hand-edited:
./scripts/build_claude_ai_zip.sh # → dist/critical-thinking-for-humans-claude-ai.zip
The build copies the canonical runtime files, including the 23 expedition
packs, and applies whole-file overlays from platforms/claude-ai/ (SKILL.md,
shared/redlines.md, shared/scaffolding.md, passport/SCHEMA.md). The platform
delta is storage only: no local filesystem, so the event log becomes a session
tally plus a copy-paste passport block the user saves and re-imports (see
platforms/claude-ai/passport/SCHEMA.md); redline 12 is reworded to stay honest
about where conversation data lives. Modes, stances, redlines 1 through 11, 13,
and 14, and the item pipeline are identical. One honest caveat: the expedition
path is less battle-tested on claude.ai than on Claude Code; if the platform
does not expose the bundled pack files to the session, expedition degrades to
its honest no-pack refusal and routes you to drill or scene.
Maintenance rule: when you edit a canonical file that has an overlay counterpart,
review the overlay in the same commit. scripts/check_invariants.py re-checks
every redline and SKILL.md invariant against the overlay copies and fails the
build on drift or on local-filesystem vocabulary leaking into the zip.
For use outside Claude, the repo can build a single self-contained Markdown file you paste whole into any frontier model's chat:
./scripts/build_portable.sh # → dist/critical-thinking-for-humans-portable.md
Paste the file as the system or first message, then say "let's practice critical
thinking" (or drill / scene / detective). The build assembles the file from
the same canonical sources, rewrites the multi-file and on-disk-passport language
into single-document language, and fails if any filesystem, router, or
absent-mode wording survives. The canonical files are never modified.
Two honest caveats live in the file's own header. It was written and tested only on Claude; it is designed to be model-neutral but is unverified elsewhere, so use a strong, current model. And two pieces of the full version are absent: the expedition mode (it needs the verified pack library, which cannot fit in one pasted document) and the on-disk passport (a plain chat has no file, so progress is tracked in the conversation only). The three included modes are drill, scene, and detective; detective is the most model-dependent and degrades on weaker models.
Tag releases build both distributable editions once in CI. The release workflow
downloads that immutable Actions artifact, verifies its SHA256SUMS, uploads
the same files to a draft GitHub Release, downloads them again, and publishes
the release only after the remote asset bytes pass the recorded checksums.