菜单
Use this skill when the user asks to 自检自身、重新蒸馏、进化版 Agent、增强回去、避免死循环处理、能力增益评估，或在瘦身与增强之间取舍。它先盘点已处理对象、能力风险和真正增益点，再决定增强、修复、拆分、瘦身或停止，不把行数减少当成默认目标。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | agent-evolution-governor |
| version | 1 |
| description | Use this skill when the user asks to 自检自身、重新蒸馏、进化版 Agent、增强回去、避免死循环处理、能力增益评估，或在瘦身与增强之间取舍。它先盘点已处理对象、能力风险和真正增益点，再决定增强、修复、拆分、瘦身或停止，不把行数减少当成默认目标。 |
| allowed-tools | read_file list_directory write_file edit_file execute_command task subagent_task |
Govern Agent self-improvement so it produces a stronger Agent, not a smaller but weaker one.
Use this skill before broad self-audit or capability refactor work. It decides whether to enhance, restore, split, slim, or leave an asset untouched.
Use when the user says:
Classify each candidate asset into one of five outcomes:
SKILL.md compact and move rare material to supporting files.Before writing, gather only relevant evidence. For a full-Agent distillation request, run the self-check as an asset-wide program, not as a single local patch:
memory/, skills/, runtime/, jobs/, and scripts/;allowed-tools where the workflow uses tools;Use subagents for independent read-only inspection when helpful, but the main Agent owns all writes and final decisions. A failed subagent does not stop the program; continue with direct local inspection.
For broad self-refactor campaigns, keep a short non-secret runtime note or local processed set with:
Use runtime/ for current campaign anchors, not memory. Never treat activity history alone as the campaign state when deciding what to revisit.
Every change must preserve or improve:
If any of these would be lost, choose restore/enhance/split instead of slimming.
After changes:
Report:
Use this skill when the Agent needs to decide where a capability, rule, workflow, runtime anchor, script, job, document, persona change, plugin asset, or repository maintenance item should live. It provides a layered map of the Agent's durable assets so requests to “落实到 all”, “整理自身框架”, or “细化自身” do not bloat memory or scatter files.
Use this skill when the Agent needs an executive-control layer before acting: classify user intent, decide direct execution vs read-only inspection vs button confirmation vs clarification, avoid over-asking, and keep work moving to the smallest safe completed state. Trigger especially when the user says 继续、落地、放开手脚、接管、自己判断、按方案推进，or complains about typing confirmations that should be buttons.
Use this skill when the Agent needs to read feedback signals before deciding whether to act, correct itself, distill a lesson, or ask for clarification. It treats user wording, repeated failures, tool errors, missing evidence, and activity context as sensory inputs, then routes to the right organ without overreacting.
Use this skill when the Agent needs to turn experience into a repeatable growth loop: observe signals, decide whether learning is needed, distill reusable lessons, place them into the right asset layer, verify the change, and keep future behavior lighter rather than heavier. Trigger when the user asks to 主动成长、持续进化、经验闭环、复盘后改进, or when repeated failures show a missing habit rather than a one-off bug.
Use this skill when the Agent needs to distinguish safe self-improvement from high-risk mutation. It guards credentials, destructive actions, MoviePilot core files, downloads, restarts, external service changes, and over-broad memory edits while still allowing authorized low-risk /config/agent and /config local plugin work to proceed.
Use this skill when the Agent needs to map a request or failure to one of its long-term capability organs: executive control, proactive orchestration, self-correction, growth cycle, capability map, distillation metabolism, verification, or repository maintenance. It prevents dumping everything into memory and keeps each organ's responsibility clear.