I Built an Agent Skill to Stop AI From Answering the Wrong Question Beautifully

Описание и применение

Judgment Loop — это агентский навык, который предотвращает ответы ИИ на неправильные вопросы. Он работает как маршрутизатор и защитная ограда для важных решений, требуя проверки гипотез и тестирования на фальсификацию перед принятием действий.

Задача

AI systems produce coherent, detailed, and persuasive answers that can be misleading. The model may optimize a proxy while the underlying reality stays unchanged, or turn an attractive plan into a commitment before testing its weakest assumption.

Результат

The skill routes between five modes (Quick, Decision, Research, Learning, Review) to ensure answers are falsifiable and actions remain under user ownership. It prevents premature commitment to uncertain claims by requiring isolated falsification tests.

Рабочий процесс

• Identify the core question and the reality supposed to change.
• Check if the chosen metric could improve while that reality remains the same.
• Classify claims as verified facts, reasonable inferences, or untested hypotheses.
• Identify the strongest reason the current judgment could fail.
• Determine the cheapest test that separates competing explanations.
• Decide what action should close now and what belief should remain open to revision.

Доказательства

The model may optimize a proxy while the underlying reality stays unchanged. It may turn an attractive plan into a commitment before testing its weakest assumption.

Источник

The skill routes between five modes: Quick, Decision, Research, Learning, Review.

Источник

The repository is available at: https://github.com/sirsws/judgment-loop

Источник