cognitive-bias-detection（认知偏差检测）

Cognitive Bias Detection

Core principle: Human (and AI) reasoning is systematically distorted by cognitive biases — predictable errors in judgment that operate below conscious awareness. The most dangerous analyses are the ones that *feel* most certain. This skill audits the reasoning process itself, not just the conclusions.

The Most Impactful Biases to Check

Evaluation & Decision Biases

Confirmation Bias Seeking, interpreting, and remembering information that confirms existing beliefs. Disconfirming evidence is dismissed or reframed.

• *Signal*: "The data confirms what we suspected." / Evidence against the conclusion gets less attention than evidence for it.
• *Fix*: Actively seek the strongest case *against* the conclusion. Assign someone to argue the opposite.

Anchoring Over-weighting the first number, estimate, or framing encountered.

• *Signal*: Estimates cluster around an initial figure. Comparisons are made relative to a reference point that was never validated.
• *Fix*: Generate estimates independently before seeing others. Ask "what would this look like if the anchor didn't exist?"

Availability Heuristic Overweighting recent, memorable, or vivid events when estimating likelihood.

• *Signal*: "We just had an incident like this" leads to overestimating its probability. Quiet failures are underweighted.
• *Fix*: Use base rates. Ask "how often does this actually happen over a long period?"

Sunk Cost Fallacy Continuing a course of action because of past investment, not future value.

• *Signal*: "We've already put 6 months into this." / Reluctance to abandon despite evidence it's not working.
• *Fix*: Ask "if we hadn't invested anything yet, would we start this today?"

Planning Fallacy Systematic underestimation of time, cost, and risk — even when we know past projects ran over.

• *Signal*: Estimates feel optimistic. No buffer for unknowns. Past projects are treated as exceptions.
• *Fix*: Use reference class forecasting: how long did similar projects actually take?

Social & Group Biases

Groupthink Desire for group harmony overrides realistic appraisal. Dissent is suppressed.

• *Signal*: Everyone agrees quickly. No one plays devil's advocate. Contrarian views are dismissed socially.
• *Fix*: Assign a formal devil's advocate. Ask people to write independent opinions before group discussion.

Authority Bias Overweighting the opinion of someone perceived as an authority, independent of their actual expertise.

• *Signal*: "The CTO/senior person thinks X, so it must be right." Analysis stops when authority speaks.
• *Fix*: Evaluate the argument on its merits, not its source. Ask "what's the evidence, separate from who said it?"

In-group Bias Favoring people, ideas, and solutions associated with one's own group.

• *Signal*: Solutions from the team are evaluated more generously than identical solutions from outside.
• *Fix*: Blind evaluation where possible. Ask "would we accept this if a competitor proposed it?"

Framing & Perception Biases

Framing Effect The same information leads to different decisions depending on how it's presented (gain vs. loss framing).

• *Signal*: "90% success rate" vs. "10% failure rate" trigger different reactions to the same fact.
• *Fix*: Reframe every option in multiple ways before deciding. Check if the decision changes.

Survivorship Bias Drawing conclusions from visible successes while ignoring invisible failures.

• *Signal*: "Company X did Y and succeeded" — but how many companies did Y and failed?
• *Fix*: Actively seek the failure cases. Ask "what don't we see because they didn't survive?"

Dunning-Kruger Effect Low competence in a domain produces overconfidence; high competence produces underconfidence.

• *Signal*: Extreme certainty in a novel or complex domain. Or excessive hedging from a genuine expert.
• *Fix*: Calibrate confidence against demonstrated track record in this specific domain.

Recency Bias Overweighting recent data and underweighting long-term patterns.

• *Signal*: Last quarter's results dominate the analysis. Historical base rates are ignored.
• *Fix*: Extend the time window. Look at multi-year trends, not just recent performance.

Output Format

🔍 Bias Scan Results

For each bias checked:

⚠️ High-Risk Findings

For each high-severity bias detected:

• Bias: Name and brief description
• How it's showing up: Specific evidence in the reasoning or decision
• What it's distorting: What conclusion or estimate is being skewed, and in which direction?
• Debiasing move: Concrete action to correct or validate

🧹 Debiased Re-evaluation

After flagging biases, offer a corrected version of the analysis:

• What changes if we remove the bias?
• What evidence is actually strong vs. inflated by bias?
• Does the conclusion still hold?

🎯 Confidence Calibration

• What is the actual confidence level warranted by the evidence, absent bias?
• What would need to be true to justify higher confidence?
• What's the most important thing to validate before committing?

Meta-Check: Is Claude Biased Here?

This skill also applies to Claude's own analysis. When generating evaluations, check:

• Am I confirming what the user wants to hear? (sycophancy / confirmation bias)
• Am I anchoring to the first framing the user gave me?
• Am I overweighting the most vivid or recent example?
• Am I assuming the user's group/team/approach is better without evidence?

If yes to any — flag it and correct.

Thinking Triggers

• *"How would this analysis look if we had concluded the opposite from the start?"*
• *"What's the strongest evidence against the current conclusion?"*
• *"Are we continuing because it's right, or because we've invested too much to stop?"*
• *"Who benefits from this conclusion, and are they also the ones evaluating it?"*
• *"If a stranger reviewed this reasoning, what would they say we're missing?"*