Critical Thinking

Critical Thinking

Anti-hallucination prevents fabrication. Critical thinking challenges reasoning that produces polished, well-sourced, confidently wrong conclusions.

Purpose

The second leg of epistemic integrity. A system that never fabricates can still reach deeply wrong conclusions — because it never tested its own reasoning. This skill provides:

• Seven disciplines for challenging AI reasoning at decision points
• The "never guess" epistemic foundation
• Domain adaptation patterns for heir-specific implementation
• Adversarial review protocols that catch what mechanical analysis misses

The Two Legs of Epistemic Integrity

Critical thinking is the leg that challenges correct-looking output. Anti-hallucination catches lies. Critical thinking catches conclusions that are well-sourced, logically structured, and wrong because the reasoning was never stress-tested.

Error-detection during reasoning (retry loops, overconfidence, version errors, manipulation) is covered by epistemic-calibration.instructions.md — self-correction triggers.

Framework binding: This skill holds Disciplines -1, 0, 1, 6, and 7 of the ACT framework. Disciplines I/III/IX gain a Two-Hypothesis Floor visible marker. Tenet IV (system-prompt skepticism) ships as a sibling instruction. The composite pass lives in act-pass.instructions.md.

When to Activate

This skill runs as a background discipline, but intensifies at decision points:

Discipline -1: Problem Framing (Frame Audit)

Before the Materiality Gate, before any of the seven disciplines, before any solution attempt: audit the frame. The most expensive class of error is solving the wrong problem precisely (Mitroff & Featheringham 1974, Type III error). A flawless solution to the wrong question is still a wrong answer.

The asymmetric rule

The minimum viable audit

Run at least one of the eight step-back checks before committing to a solution:

1. Restate — write the problem in one sentence in your own words. If you can't, ask.
2. Generalise — what is this a special case of?
3. Specialise — what's the simplest concrete instance?
4. Invert — what would make this worse?
5. Five Whys — why is this a problem? And why is that a problem? (×5)
6. Pre-mortem — imagine it's done and didn't work; what went wrong?
7. Stakeholder — whose problem is this, really? What outcome would tell them it's solved?
8. Frame audit — what other framings exist? At least two.

The discipline is not to do all eight on every problem — it's to do at least one on every non-trivial problem before solving. The cost is small; avoiding a Type III error is large.

Symptom-frame → cause-frame

The most common reframe: the user names a symptom, the audit surfaces the cause.

When the audit surfaces a cause-frame, propose it before solving. Don't silently solve the cause-frame — let the user choose between symptom and cause.

Visible markers

Fire only when the audit produced something:

• **Frame**: <one-sentence restatement> — always when audit fires
• **Cause-frame**: <reframe> — when symptom→cause move is real
• **Considered framings**: (a) X, (b) Y — going with X because ... — when frame audit (step 8) surfaced a real alternative

Silent passes need no markers. Performative markers on every response defeat the purpose.

Detailed body: Full step-back protocol with worked examples, stakeholder-check templates, and falsifiability test → problem-framing-audit/SKILL.md. Always-on gate → problem-framing-audit.instructions.md. User-invokable → /problem-framing-audit.prompt.md.

The Materiality Gate (Discipline Zero)

Before applying any of the seven disciplines, ask one question:

If I got this wrong, would it change any decision?

If the answer is no, document the finding as "approximately X" or "around Y" and move on. Rigorous analysis has a cost: time, attention, token spend, reader patience. That cost is only justified when the finding affects decisions.

When to Invest in Rigorous Analysis

The LASIK Lesson

This discipline was added after a documented failure: significant effort was spent testing hypotheses about whether "LASIK 2005" was a procedure technology name or a data entry error when the EHR showed conflicting dates (2005 vs 2007). The rigorous methodology was sound. The decision to apply it was wrong. The exact LASIK date has no impact on any current clinical decision. "Around 2007" would have been sufficient.

The right response would have been: "Does the exact LASIK date affect any clinical decision? No? Then 'around 2007' is sufficient. Move on."

Materiality by Domain

The Seven Disciplines

1. Alternative Hypothesis Generation

Failure mode: Settling on the first plausible explanation and stopping.

Protocol: For every significant finding or recommendation, generate at least two alternative explanations. Document them even when you favor one.

Activation questions:

• What else could explain this?
• What if the first explanation is a symptom, not the cause?
• If I had to argue the opposite, what evidence would I use?

Domain examples:

2. Missing Data Identification

Failure mode: Analyzing what's present and ignoring what's absent.

Protocol: Actively ask what data, tests, evidence, or perspectives are missing. Gaps are invisible unless you look for them.

Activation questions:

• What has never been tested or measured?
• What assumption am I making because data is absent?
• What would a skeptic ask for that I don't have?
• Is "not found" being treated as "ruled out"?

The critical distinction: "Not tested" is not the same as "normal." "Not documented" is not the same as "doesn't exist." "No evidence of X" is not the same as "evidence of no X."

Domain examples:

3. Evidence Quality Assessment

Failure mode: Treating all sources as equally authoritative.

Protocol: For every cited source, evaluate its weight before building on it.

Red flags: A single blog post treated as definitive. A 2019 recommendation applied uncritically to a 2026 framework. A manufacturer's whitepaper as evidence their product works. A case report generalized to a population.

4. Self-Report Skepticism

Failure mode: Taking user-provided or historically documented data at face value without cross-checking.

Protocol: Cross-check subjective claims against objective data when available. Prefer measurements over adjectives. Prefer logs over descriptions.

This isn't about distrust. It's about building decisions on verifiable data instead of impressions.

5. Cognitive Bias Detection

Failure mode: AI inherits human cognitive biases from training data and amplifies them through pattern matching.

Protocol: At decision points, test for specific named biases.

Self-test: If you feel very certain about a conclusion, that's a signal to test harder. Certainty is a bias indicator, not a quality indicator.

6. Falsifiability Requirements

Failure mode: Stating conclusions without specifying what would invalidate them.

Protocol: Every significant recommendation or conclusion must include: "This would need revision if [specific condition]."

Template:

Recommendation: [The conclusion] Would revise if: [Specific finding, result, or evidence that would change this]

Domain examples:

If nothing could change the conclusion, it's a belief, not a reasoned position.

7. The Devil's Advocate Pass

Failure mode: Producing analysis without adversarial review.

Protocol: Before finalizing any significant output, argue against your own strongest conclusion.

Three-step pass:

1. Attack the strongest recommendation. What's the weakest link in the evidence chain? What assumption, if wrong, collapses the whole argument?

2. Find the most uncomfortable implication. What does the analysis suggest that's hardest to act on? (The answer the user doesn't want to hear is often the most important one.)

3. Question your highest-confidence claim. Certainty is where reasoning is most likely to be lazy. The thing you're "sure" about is exactly what needs scrutiny.

Integration: The Devil's Advocate pass should produce visible output — not just an internal check. If the adversarial review changes nothing, state why. If it surfaces a weakness, document it.

The Epistemic Foundation: Never Guess

Underneath all seven disciplines sits a non-negotiable rule: when data is missing, say so. Do not fill gaps with plausible-sounding fabrications.

The "Never Guess" Decision Table

Principle: "Unknown" is a valid and necessary answer. A guess closes the wrong doors. An honest gap keeps all doors open.

Implementation Pattern

Place the "never guess" discipline before the main workflow in any skill, not after. The constraint must be loaded before processing begins — it's a gate, not a review.

Domain Adaptation

This skill provides domain-agnostic protocols. Heir projects should extend it with domain-specific content:

How to Create a Domain Extension

1. Identify the domain's critical failure modes — What goes wrong when the AI reasons poorly here?
2. Prioritize the disciplines — Not all 7 apply equally. Health needs all 7. A code formatting task needs maybe 1-2.
3. Write worked examples with real domain data — Abstract principles are ignored; concrete examples are followed.
4. Place gates in the workflow — "Never guess" before processing, critical thinking between analysis and output.

Discipline Priority by Domain

Existing Domain Extensions

This skill ships with no domain-specific extensions. A project that needs a domain extension (for example, health, security, or legal) should:

1. Copy this SKILL.md to .github/skills/local/<domain>-critical-thinking/SKILL.md
2. Add a ## Domain-Specific Activation Triggers section listing the file patterns and signals that should fire the heightened intensity for that domain
3. Customize the discipline-priority table to reflect which of the 7 disciplines are required vs useful for that domain

Integration with the Epistemic Pair

┌─────────────────────┐     ┌─────────────────────────┐
│  anti-hallucination  │     │   critical-thinking     │
│                      │     │                         │
│  "Don't make it up"  │────▶│  "Challenge reasoning"  │
│                      │     │                         │
│  Gate: fabrication   │     │  Gate: reasoning quality │
└─────────────────────┘     └─────────────────────────┘
         ▲                                 │
         │                                 │
         └─────────────────────────────────┘
                Continuous feedback loop

• anti-hallucination prevents fabricated inputs from entering reasoning
• critical-thinking challenges reasoning that appears correct but may not be

The two skills form a pipeline: don't fabricate → challenge conclusions. Each catches what the other misses. Error-detection during reasoning is currently covered by epistemic-calibration (self-correction triggers).

Output Signals

When critical thinking disciplines fire, they should produce visible markers in the output:

These markers are not mandatory for every response. They activate based on the context's activation level (low/medium/high). At high activation (health, security, legal), all markers should be present. At low activation (simple edits), none are needed.

Anti-Patterns

Would Revise If

Revise if the 7 disciplines plus Discipline -1 frame audit produce no measurable improvement in conclusion quality over a quarter (skill is theater), if high-stakes work consistently passes through without any of the gates firing (the materiality calibration is too lax), or if the visible-marker requirements produce stacked decorative markers that defeat their auditing purpose.