domains / families / science / falsify

falsify

authentic

The scientific thinking protocol for AI agents. Use when facing complex, ambiguous, or high-stakes questions where guessing is costly: hypothesis → attempt to break it → evidence → calibrated conclusion.

When to use

• Architecture / design decisions with trade-offs
• "Why" questions about a failing system or data anomaly
• Recommendations that will be acted on (a library, a fix, a strategy)
• Claims about what a user, market, or system "will" do
• Anything where being wrong costs time, money, or trust
• Factual recall you can verify in one lookup
• Trivial questions where the answer is obvious and consequences are zero
• Small talk. Not everything is a thesis defense.

When not to use

• The protocol changes *how* an agent concludes, not *what* the agent knows — it cannot manufacture evidence the model was never given, and it must never be used to fabricate sources or confidence.
• No amount of internal falsification substitutes for an external signal: re-examining your own reasoning without new evidence adds no confidence (Huang et al. 2023). When a claim needs ground truth, the agent must name the test or the sourc…
• The skill is contextual, not mandatory: it must not turn simple lookups or small talk into thesis defenses. Depth is a tool, not a costume.
• Outside the agent's (or the verified sources') area of competence, the calibrated answer is "I don't know" — not a hedged guess.

Skill page

Source