Popperian reasoning

Popperian Reasoning

Use this skill to reason by criticism, not confirmation. Start from a problem, propose tentative answers, try to refute them, and keep only the best surviving answer provisionally.

This SKILL.md is the portable source of truth for Codex, Claude, opencode, and other agent harnesses. Host-specific metadata may exist beside it, but the reasoning workflow must not depend on a specific platform.

Modes

• Quick - Compress the loop to rivals, falsifiers, survivor, and next test when the user needs a short answer.
• Standard - Use the full loop by default.
• Deep - Add more rivals and sharper tests when the decision is high-impact or the evidence is weak.

Core Loop

Use these headings by default, especially for smaller models:

1. Problem - State the concrete problem or decision in one sentence.
2. Conjectures - Give 2-4 competing explanations, plans, or answers.
3. Falsifiers - For each conjecture, say what observation, test, or result would show it is wrong.
4. Risky Tests - Prefer tests that could genuinely fail the favored conjecture, not tests that merely collect confirming evidence.
5. Error Elimination - Kill, revise, or demote conjectures that are vague, unfalsifiable, contradicted, or patched ad hoc.
6. Provisional Survivor - Choose the best remaining answer and state why it survived these tests.
7. Next Test - End with the cheapest concrete check that would most improve confidence.

Rules

• Ask "what would break this?" before "what supports this?"
• Treat confirmations as weak unless they came from a risky prediction.
• Penalize claims that can explain every possible outcome.
• Do not rescue a failing conjecture with an ad hoc patch unless the patch creates a new falsifiable prediction.
• Say "survived these tests" or "best current explanation", not "proven true".
• If only one conjecture is available, create at least one serious rival before judging it.
• If the user needs a short answer, compress the loop but keep falsifiers, error elimination, and the next test.

Anti-Patterns

• Do not list only supporting evidence.
• Do not judge a single conjecture without a serious rival.
• Do not accept claims that can explain every possible outcome.
• Do not call a surviving conjecture proven.

Output Template

**Problem**
[Concrete problem.]

**Conjectures**
1. [Conjecture A]
2. [Conjecture B]

**Falsifiers**
- A would fail if [testable failure].
- B would fail if [testable failure].

**Risky Tests**
- [Test that could refute the leading conjecture.]

**Error Elimination**
- [Conjecture]: [keep, revise, demote, or reject] because [reason].

**Provisional Survivor**
[Best remaining answer, with uncertainty stated.]

**Next Test**
[Cheapest concrete check.]

Quality Bar

A good Popperian answer makes the favored idea vulnerable. If the response only lists supporting evidence, rewrite it around falsifiers and competing conjectures.

Self-Check

Before the final answer, verify:

• A serious rival exists.
• The favored conjecture has a real falsifier.
• At least one conjecture was killed, revised, or demoted.
• The next test is concrete and cheap.

References

Read references/examples.md only when concrete examples would help adapt this skill to a task type.