技能修剪

工程开发 社区

技能解读

当前原文的自动摘译,尚未人工审校。

何时调用

每季度的节奏。在scripts/skill_linter.py --check-orphans报告孤立参考文件之后。当 SKILL.md 或参考文件增长超过其上限并需要修剪时。经过去专业化清理后,删除与已退休挑战相关的内容。

修剪标准(四个信号 - 与 /skill-update 形状相同,倒置)

当参考/场景/行满足以下任一条件时,它就是修剪候选者: 孤立 - 在过去 60 天内未从任何 SKILL.md 或其他参考文件链接。仅在失败的交战中引用 — 出现在以 status=BLOCKED 结束的运行中的 attack-chain.md 中,从未出现在成功的链中。与新内容相矛盾——后来的场景/模式取代了它;旧条目不再反映当前技术。

安全规则

切勿使用“annotation. Never prune content cited in a still-open engagement'sOUTPUT_DIR/attack-chain.md”删除文件。切勿针对单所有者规则(暴力破解、输出规则、环境读取器、技能更新)修剪规范主文件。偏向于保留技术丰富的内容而不是操作知识。

程序

运行scripts/skill_linter.py --check-orphans以显示孤儿。对于每个候选文件或块,评估四个信号。制定删除计划 — 显示要删除的文件/行,并为每个项目提供一行理由。仅在计划获得批准后应用删除(技能修剪在调用期间不会自动删除)。重新运行scripts/skill_linter.py以确认更改没有破坏其他链接并且没有重新引入重复项。

输出

简明变更报告:已删除。文件或块+每一项的一行理由。保留(标记为后续)。未能满足所有四个信号但值得下季度重新审视的项目。没有李子。明确说明何时没有任何内容值得删除。

流狐档案 运行与权限由流狐静态检查推断,不代表作者声明或实际测试结果。
领域
工程开发
作者声明的智能体
未找到明确声明;不据此推断已兼容或已测试
静态检查
88 / 100 · 启发式扫描,不代表运行安全
运行环境证据
未执行;当前没有隔离运行测试证据
作者 / 版本 / 许可证
@transilienceai · 未声明许可证
上下文用量评级
低消耗
接入复杂程度
即装即用
是否需要外部接口密钥
不需要
兼容的系统
未声明(默认跨平台)
运行环境要求
未检出明确要求
检测到的文件与系统行为
  • 只读
  • 允许写入 / 修改
网络行为
仅限本地
安装命令数
无(仅作为资料)

需要注意: 未声明工具权限范围;实际权限由运行环境决定。

作者说明文件(原文) SKILL.md

# Skill Prune

Inverse of `/skill-update`. Removes content rather than adding it. Run during quarterly maintenance, after engagements, or when `scripts/skill_linter.py` reports orphans / duplicates.

## When to invoke

- Quarterly cadence.
- After `scripts/skill_linter.py --check-orphans` reports orphan reference files.
- After a SKILL.md or reference file grows past its cap and needs trimming.
- After a de-specialization sweep, to drop content tied to retired challenges.

## Prune criteria (the four signals — same shape as /skill-update, inverted)

A reference / scenario / line is a **prune candidate** when it satisfies any of:

1. **Orphan** — not linked from any SKILL.md or other reference file in the last 60 days.
2. **Referenced only by failed engagements** — appeared in `attack-chain.md` of runs that ended `status=BLOCKED`, never in a successful chain.
3. **Contradicted by newer content** — a later scenario / pattern supersedes it; the older entry no longer reflects current technique.
4. **Redundant with newer content** — same technique covered more clearly elsewhere.

Removing content fails any of these → keep it.

## Safety rules

- **Never** prune a file with `<!-- KEEP: <reason> -->` annotation.
- **Never** prune content cited in a still-open engagement's `OUTPUT_DIR/attack-chain.md`.
- **Never** prune the canonical-home file for a single-owner rule (brute-force, output-discipline, env-reader, skill-update).
- Bias toward keeping technique-rich content over operational lore.

## Procedure

1. Run `scripts/skill_linter.py --check-orphans` to surface orphans.
2. For each candidate file or block, evaluate the four signals.
3. Build a deletion plan — show files / lines to remove with one-line rationale per item.
4. Apply deletions only after the plan is approved (skill-prune does not auto-delete during invocation).
5. Re-run `scripts/skill_linter.py` to confirm the change broke no other links and didn't reintroduce duplicates.

## Output

Concise change report:

- **Removed.** File or block + one-line rationale per item.
- **Kept (flagged for follow-up).** Items that failed all four signals but are worth revisiting next quarter.
- **No prunes.** State explicitly when nothing warranted removal.

## Anti-Patterns

- Pruning content because it's old — age alone is not a signal; relevance is.
- Pruning a reference because its parent SKILL.md is bloated — fix the SKILL.md instead.
- Removing a single-owner canonical file (brute-force / output-discipline / env-reader).
- Pruning during an active engagement that may still cite the content.

讨论

基于 GitHub Discussions。登录 GitHub 即可参与讨论、点赞、订阅更新。