No lectures — just critical questions that expose fallacies. Scans any argument across 12 dimensions, names the logical fallacies by sentence, and trains you to ask the questions that break weak reasoning.
📚 Reading Notes PPT: "Asking the Right Questions — A Guide to Critical Thinking"
English | 中文 | 🌐 Landing Page
📄 "Here's the proposal deck — poke holes in it." Paste the text and get a 12-dimension diagnostic report: conclusion clarity, reason sufficiency, hidden assumptions, evidence tier, and an overall Strong / Medium / Weak rating.
🕵️ "Is there anything logically wrong with this?" Sentence-level fallacy scan across 12 classic fallacies — each hit quotes the original line, explains why it fails, and offers a rewrite.
❓ "Train me. Ask me the hard questions." Socratic mode walks you down a 5-layer ladder — surface facts → reasons and evidence → hidden assumptions → alternative explanations → verification — in 5 rounds or fewer.
🔎 "Audit this decision before I commit." A 12-item decision quality checklist plus reversibility rating, worst-case test, expected-regret test, and an explicit Kill Switch — the conditions under which you should walk away.
📊 "Does this data actually say what they claim?" Splits any chart or report into what the data tells you / what it doesn't / the traps hiding inside, checked against 7 data traps and a 7-level evidence hierarchy.
⚔️ "Prepare me for both sides of this debate." Balanced pro/con tables with best evidence, likely rebuttals, and your counter-rebuttals — plus the key clash points and the fallacies you're most likely to commit yourself.
📋 "Give me a question list for tomorrow's review." Checklist mode outputs a prioritized set of must-ask questions, each tagged with the dimension it tests and what it means if nobody can answer it.
💼 "How should I think about this at work?" 8 workplace scenarios pre-mapped to the dimensions that matter most: proposal review, interview evaluation, data reporting, conflict mediation, strategic decisions, report auditing, managing up, and performance evaluation.
Built on Asking the Right Questions (Browne & Keeley). Every scan walks all 12 dimensions and rates each ✅ Pass /
| # | Dimension | What It Tests |
|---|---|---|
| ① | Issue & Conclusion | Is the issue descriptive or prescriptive? Is the conclusion identifiable? |
| ② | Reasons | Are reasons independent, relevant, sufficient (≥2), and verifiable? |
| ③ | Ambiguity | Do key terms have operational definitions, or is the argument riding on vague words? |
| ④ | Value Assumptions | Which value won the trade-off — efficiency vs fairness, short vs long term? |
| ⑤ | Descriptive Assumptions | What unstated belief bridges reasons to conclusion? When does it break? |
| ⑥ | Reasoning Fallacies | Any of the 12 fallacies in the inference chain? |
| ⑦ | Evidence Credibility | Which tier of the evidence pyramid? Sample size, method, conflicts of interest? |
| ⑧ | Rival Causes | Confounders, reverse causation, third factors, or plain coincidence? |
| ⑨ | Statistical Traps | Base rate, denominator, time window, mean vs median, framing? |
| ⑩ | Omitted Information | Counter-evidence, costs, long-term effects, failure cases, missing stakeholders? |
| ⑪ | Reasonable Conclusions | What other conclusions do the same reasons support? Absolute, conditional, or probabilistic? |
| ⑫ | Decision Quality | Alternatives compared, worst case, reversibility, kill switch, expected regret? |
Each entry ships with a definition, a recognition pattern, typical phrasings, workplace examples, and ready-to-use correction language.
| Fallacy | One-Line Tell |
|---|---|
| Ad Hominem | Attacks the speaker's background instead of the claim |
| Straw Man | "So you're saying…" followed by something they never said |
| Slippery Slope | A → B → C → D with no evidence for any link |
| False Dichotomy | "Either… or…" when a middle path exists |
| Appeal to Authority | "Big companies do it" replaces analysis of fit |
| Circular Reasoning | The reason is the conclusion rephrased |
| Hasty Generalization | "All / every time / always" from a handful of cases |
| False Cause | B happened after A, therefore A caused B |
| Red Herring | Answers a question nobody asked |
| Appeal to Emotion | Guilt and fear substituted for evidence |
| Bandwagon | Popularity treated as proof of correctness |
| False Analogy | "X is just like Y" where the key features differ |
Guided mode never dumps a wall of questions. Maximum 2 questions per round, maximum 5 rounds, with feedback after every answer and a synthesis of your weakest link at the end.
Layer 1 Surface facts "What exactly? Where's the data?"
Layer 2 Reasons & evidence "Why do you believe that? Based on what?"
Layer 3 Hidden assumptions "What are you taking for granted? Does it hold?"
Layer 4 Alternatives "What other explanation or option exists?"
Layer 5 Verification "How would you test this? What does it cost?"
The 12-item decision checklist ends with something most frameworks skip: an explicit Kill Switch — the pre-committed conditions under which you re-evaluate or abandon the decision, so sunk cost can't hold you hostage. Output includes decision type (reversible / irreversible, risk level), an X/12 completeness score, and a verdict of Execute / Reinforce First / Hold.
| Data Trap | Catch It By Asking |
|---|---|
| Base rate fallacy | "What's the denominator? What's the absolute number?" |
| Mean vs median | "What's the median? What does the distribution look like?" |
| Cherry-picked time window | "Why start here? Show me YoY and MoM." |
| Survivorship bias | "What about the ones that failed or churned?" |
| Simpson's paradox | "Does it still hold when you split the groups?" |
| Missing data | "Which metrics went down and weren't shown?" |
| Correlation as causation | "What's the mechanism? Could C be causing both?" |
Evidence is ranked across 7 tiers, from meta-analysis and randomized controlled trials down to expert opinion and personal anecdote — so "I've done this before" and "our A/B test showed" never carry the same weight.
Each scenario comes pre-loaded with its priority dimensions, a must-ask question list, the fallacies that show up most often there, and the exact phrasing to use.
| Scenario | Priority Dimensions | Most Common Trap |
|---|---|---|
| Proposal review | ② ⑥ ⑦ | Appeal to authority |
| Interview evaluation | ⑥ ⑦ ⑧ | Halo effect / confirmation bias |
| Data reporting | ⑨ ⑩ ⑧ | Selective presentation |
| Conflict mediation | ③ ④ ⑤ | Straw man |
| Strategic decision | ⑪ ⑫ ⑩ | Slippery slope |
| Report auditing | ① ② ⑥ | Unclear conclusion |
| Managing up | ② ⑦ ④ | Self-centered reasoning |
| Performance evaluation | ⑦ ⑥ ⑧ | Recency effect |
| Dimension | mu-critical-thinking | Generic ChatGPT prompt | Toulmin Model | Six Thinking Hats |
|---|---|---|---|---|
| Diagnostic framework | 12 fixed dimensions, full coverage | Varies every run | 6 components | 6 perspectives |
| Fallacy identification | 12 fallacies, sentence-level quoting | Occasional, unnamed | Not covered | Not covered |
| Repair suggestions | Correction script per fallacy | Generic advice | Not covered | Not covered |
| Evidence grading | 7-tier hierarchy | None | Qualitative only | None |
| Data trap detection | 7 named traps | None | None | None |
| Decision auditing | 12-item checklist + Kill Switch | None | None | Partial (black hat) |
| Interactive training | Socratic mode, 5 rounds max | Unbounded chat | Static model | Meeting protocol |
| Debate preparation | Symmetric pro/con + clash points | One-sided by default | Single-argument | Not designed for it |
| Workplace grounding | 8 mapped scenarios | Generic | Academic | Generic |
| Anti-over-diagnosis rule | Yes — says "no fallacies found" when true | Tends to invent problems | N/A | N/A |
| Output format | Fixed report templates | Free text | Diagram | Discussion notes |
| Setup cost | Drop in a folder | Rewrite prompt each time | Learn the theory | Train the whole team |
Three interaction modes:
| Mode | Behavior | When It Activates |
|---|---|---|
| 🚀 Quick (default) | Paste → auto 12-dimension scan → diagnostic report | Text pasted with no mode specified |
| 💬 Guided | Socratic dialogue, decision audit, or data coaching | You say "train me", "audit this", "teach me to read this" |
| 📋 Checklist | Structured question list, no back-and-forth | You say "give me a question list" / "what should I ask at the review" |
Six workflows:
| Workflow | Scenario | Trigger |
|---|---|---|
| Argument Quality Assessment | You have a proposal, memo, or claim and want to know if it holds up | Paste any text with no mode specified → auto 12-dimension scan |
| Logical Fallacy Identification | You sense something is off but can't name it | "Is there anything logically wrong with this?" |
| Socratic Questioning Training | You want to build the reflex, not just get an answer | "Train me" / "ask me questions" / "help me prepare follow-ups" |
| Decision Quality Audit | A consequential call is about to be made | "Audit this decision" / "should we do this?" |
| Data Interpretation Coach | A chart or report looks too good to be true | "Is this data reliable?" / "how should I read this?" |
| Debate Preparation | You need both sides before you pick one | "Prepare pro and con arguments for me" |
| Item | Description |
|---|---|
| Type | Markdown-based AI skill — no code, no runtime, no build step |
| Dependencies | None. Zero npm/pip packages, zero API keys |
| Compatible with | Claude Code, Claude Desktop, and any agent supporting the Skill / SKILL.md convention |
| Package size | ~60 KB total (SKILL.md + 4 reference files) |
| Structure | 1 SKILL.md + 4 references (twelve-dimensions, logical-fallacies, evidence-evaluation, workplace-scenarios) |
| Input | Plain text, proposals, reports, meeting notes, data descriptions, chart summaries |
| Output | Structured Markdown reports — diagnostic table, fallacy scan, decision audit, data reading, debate prep |
| Language | Chinese-first content; works with English input and output |
| Progressive loading | Reference files are read only when the matching scenario fires, keeping context lean |
| Version | 2.0.0 |
| License | MIT |
Cross-skill linkages (optional, if you have them installed):
mu-pyramid-principle— critical thinking deconstructs the problem, pyramid principle restructures the expressionmu-humanizer-minesweeping— strips the AI flavor out of the resulting text
1. Install — clone into your skills directory
git clone https://github.com/muippt/mu-critical-thinking.git ~/.claude/skills/mu-critical-thinkingUsing a different agent? Just drop the folder wherever your tool loads skills from. Project-level works too:
.claude/skills/mu-critical-thinking.
2. Verify — restart your agent and confirm the skill is picked up
List my available skills
3. Run — paste something and let it work
Paste any proposal, report, or claim → automatic 12-dimension diagnostic report
Or invoke a specific workflow:
Scan this for logical fallacies: "Every major company is doing this, so we should too — if we don't, we'll fall behind and eventually lose the market entirely."
Audit this decision: we're going to double the team size next quarter to hit the roadmap.
Train me with Socratic questions on: remote work lowers team productivity.
- Fully local — the skill is plain Markdown read by your agent. Nothing is installed, nothing runs.
- Zero network calls — no API endpoints, no external services, no remote fetches at any point.
- Zero telemetry — no usage tracking, no analytics, no phone-home. There is no code that could collect anything.
- Zero dependencies — no npm, pip, or third-party packages, so no supply-chain surface.
- Your text stays where your agent is — the skill adds no storage, no logging, and no persistence of the arguments you analyze.
- Fully auditable — 5 Markdown files, all human-readable. Read the whole thing in fifteen minutes.
- No credentials required — no API keys, tokens, or accounts.
- Judges arguments, not people — an explicit rule prohibits moral verdicts on positions; the skill evaluates reasoning quality only.
If this helps you catch one bad argument before it becomes a bad decision, a star would mean a lot.
Twelve dimensions, twelve fallacies, seven data traps — turning "something feels off" into "here's exactly what's wrong, and here's how to fix it."
🎓 Signatory Author of Tsinghua University Press / 2026 Dangdang Influential Author / AI & Large Model Business HR Specialist at a Leading Tech Company / National Level-1 HR Manager / Level-2 Psychological Counselor / Self-taught Designer
📚 Author of Visual Team Management. Clients include ByteDance, Tencent, Baidu, China Mobile, SMG, BOE…
💡 WeChat Official Account / Xiaohongshu: muippt
🧰 More skills: mu-skill-hub · GitHub
MIT © 2025 muippt
Acknowledgments
- Asking the Right Questions: A Guide to Critical Thinking by M. Neil Browne and Stuart M. Keeley — the methodological foundation of this project. The 12-dimension assessment framework, the descriptive vs prescriptive issue distinction, the value/descriptive assumption analysis, and the rival-causes discipline all derive from their work. This skill is an operational adaptation for workplace scenarios; the original book remains the canonical source and is well worth reading in full.
- The classical Socratic method, for the progressive questioning model.
- The evidence-based practice community, for the evidence hierarchy that informs the 7-tier pyramid.
- Everyone who filed an issue, asked an uncomfortable question, or argued back. That's the whole point.
Note: Much of this project was co-created with AI assistance. If you believe your work has been used without proper attribution, please open an issue.