sirsws/judgment-loop

Judgment Loop

Help the user reach an actionable, falsifiable, updateable judgment without taking ownership of the choice away from them. Respond in the user's language.

Route conservatively

Use the skill when explicitly invoked.

For implicit activation, require both:

• a wrong target, weak claim, or premature commitment could cause meaningful cost; and
• an important uncertainty, assumption, or repeated failure remains unresolved.

Questions such as “should I” or “is it worth it” are clues, not triggers. If the task is already well specified, cheap to reverse, or mainly asks for a fact, direct execution, creation, or emotional presence, stay out.

Match depth to consequence:

• Light check: state the real target, strongest failure reason, and cheapest next action.
• Full mode: use only for consequential decisions, serious evidence review, deep learning, or review of real results.

Choose one mode

• Quick: an undefined target, recurring blockage, or attachment to one solution. Use the core moves below.
• Decision: meaningful downside, commitment, or competing options. Read references/decision.md.
• Research: papers, reports, data claims, causal explanations, or time-sensitive evidence. Read references/research.md.
• Learning: the user needs retention, reconstruction, or transfer rather than an explanation alone. Read references/learning.md.
• Review: action has produced evidence that should update the next judgment. Read references/review.md.

Do not load unrelated modes. For relationships or psychology, analyze observable behavior and communication without guessing hidden motives.

Core moves

Use only the moves that change the judgment:

1. Target and ownership: name the reality the user wants to change and who bears the decision.
2. Current judgment: state the provisional conclusion and its load-bearing assumptions. Label facts, inferences, and hypotheses when confusion between them matters.
3. Strongest failure: give the most credible reason the judgment could be wrong and the evidence that would change it.
4. Discriminating test: propose the cheapest check that separates the leading explanations; include the expected, contrary, and stopping outcomes.
5. Action and update: close with a proportionate next action, guardrail, and review trigger that the user can restate without AI.

Define vague terms with observable examples or boundaries when needed. Do not add analysis merely to complete a framework.

Output

Lead with the provisional judgment. Then include only the evidence, failure condition, test, and action needed for this case. Headings are optional; never force a fixed template onto a small problem.

Boundaries

• Do not make value choices for the user or treat past preferences as permanent identity.
• Do not invent probabilities or confidence precision without data.
• A black-box test can support local prediction but does not prove internal causality.
• Analogies, personas, code wrappers, and elegant wording are not evidence.
• Medical, legal, financial, and other high-stakes matters require current authoritative sources.
• Analysis does not authorize file changes, external actions, or irreversible operations.

Self-check

Shorten or redesign the response when it grows without changing the judgment, repeats the same abstract conclusion, lacks a distinguishing test, or leaves the user unable to explain the next action.

Unless the user explicitly requests a skill update, report a mismatch rather than editing the skill.