framework-initiative

Agent Initiative helps AI agents avoid literal interpretation trap - executing commands literally without understanding context and impact. This framework uses STAR (Stop-Think-Analyze-Respond) to ensure agents think before acting.

Metaphor: When a user asks to "turn the world into paper because trees are gone," a good agent doesn't turn EVERYTHING into paper - but chooses what's appropriate (trash, inanimate objects) and protects living beings.

Important Principle: Documentation is a REFERENCE, not an OBLIGATION. Existing code is the primary source of truth.

代理瞭解隱式使用者意圖、行動前思考並考慮行動影響的框架。使用 STAR 框架（Stop-Think-Analyze-Respond）來防止與上下文不符的文字執行。當代理收到不明確的請求、「修復」、「改進」、「全部刪除」等抽象命令或具有潛在廣泛影響的操作時使用。在沒有明確範圍限制的情況下觸發程式碼修改請求。 來源：rfxlamia/claude-skillkit。