durable learning

Verify the original request, correction, and result. Treat assistant success claims as claims until checked. Search existing skills before proposing a new one. Save a concise reusable method, not a transcript, private identity, credential, or temporary task state. Include when to use it, steps, validation, and a stop condition. Use skill_manage when available. Read first, preserve user/imported skills, and use the current hash for an agent-owned patch. Creating a skill requires the user’s explicit request or an enabled task whose stated scope authorizes it. No new permissions, outbound actions, or recursive agents. A no-change result is valid. Assess the next relevant use before claiming improvement.