skill-evolve

原文涵盖的主题

• skill-evolve
• The three buckets (route every candidate before doing anything)
• The selectivity gate
• Workflow
• Hard rules
• Scope & security

Close the loop between using skills and improving them: the friction you hit this session (a wrong command, a missing recipe, a stale doc) is the highest-signal feedback there is. Capture it selectively and ship it back into the skills repo - most sessions produce 0-2 real skill improvements, and a long list is a smell. Adjacent skills: vd:rule-miner (standing-behavior corrections → CLAUDE.md/AGENTS.md rules), vd:skill-management (lifecycle ops), vd:create-skill (new skill, evals first), memory (project/user facts). skill-evolve only edits the content of existing skills from this session's evidence. 1. Skill-content improvement - a SKILL.md / reference is wrong, incomplete, or missing a reusable recipe you just proved out → this skill. 2. Standing-behavior correction - "the agent keeps doing X wrong across tasks" → vd:rule-miner. Do not write rules here. 3. Project / task fact - a brand name, this repo's layout, a one-off config → memory, never into a skill (skills are shared/public). If a candidate doesn't clearly fit bucket 1, it does not belong in a skill. Keep a bucket-1 candidate only if all three hold: (a) General - it helps a future, unrelated session, not just this task. (b

这里只展示节选，请以完整原文和当前要求为准。