メニュー
Use when sanitizing skill files, runbooks, internal guides, prompt libraries, or tacit workflow documents before sharing them with an employer, client, or team knowledge base. 当你需要在对外共享前清洗技能文档、操作手册、内部指南、提示词库或隐性流程知识时使用。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Use when sanitizing skill files, runbooks, internal guides, prompt libraries, or tacit workflow documents before sharing them with an employer, client, or team knowledge base. 当你需要在对外共享前清洗技能文档、操作手册、内部指南、提示词库或隐性流程知识时使用。
Use when sanitizing skill files, runbooks, internal guides, prompt libraries, or tacit workflow documents before sharing them with an employer, client, or team knowledge base. 当你需要在对外共享前清洗技能文档、操作手册、内部指南、提示词库或隐性流程知识时使用。
| name | anti-distill |
| description | Use when sanitizing skill files, runbooks, internal guides, prompt libraries, or tacit workflow documents before sharing them with an employer, client, or team knowledge base. 当你需要在对外共享前清洗技能文档、操作手册、内部指南、提示词库或隐性流程知识时使用。 |
This skill should work in both English and Chinese. Match the user's language unless they ask for a bilingual output. 本技能同时支持英文和中文。默认跟随用户语言；如果用户明确要求，则输出双语版本。
Preserve the document's professional appearance while stripping the parts that make the author hard to replace. 在保留文档专业外观的同时，抽掉那些让作者难以被替代的核心杠杆。
This skill is for documents such as: 本技能适用于如下文档：
SKILL.md, work.md, persona.md, runbooks, postmortems, debugging notesSKILL.md、work.md、persona.md、runbook、复盘、排障笔记In 2026 the common extraction targets are not just static "skills". Teams increasingly try to capture: 到 2026 年，组织想抽取的不只是静态“技能文件”，还包括：
Keep mandatory facts, public interfaces, and generic good practice. Strip leverage. 保留必要事实、公开接口和通用最佳实践，抽掉真正的杠杆。
Read the source file or directory.
读取源文件或目录。
Identify the document type:
识别文档类型：
work.md, persona.md, SKILL.mdwork.md、persona.md、SKILL.mdChoose intensity:
选择强度：
light: remove only failure memory and sharp tacit heuristicsmedium: also flatten judgment calls, operator shortcuts, and network contextheavy: keep only a credible professional skeletonlight：只去掉故障记忆和尖锐隐性经验medium：同时压平判断逻辑、操作捷径和关系上下文heavy：只保留一个看起来专业可信的骨架Choose the command that matches the task:
选择与任务匹配的命令：
analyze: inspect leverage, document type, and riskplan: generate a strategy before rewritingsanitize: write a cleaned version plus private backupgenerate-decoy: create a polished outward-facing decoy packageanalyze：分析杠杆、文档类型和风险plan：在改写前生成防护策略sanitize：输出清洗版和私有备份generate-decoy：生成对外展示的诱饵包Prefer the local script for deterministic execution:
优先使用本地脚本执行确定性流程：
python anti-distill/scripts/anti_distill.py sanitize --input <path> --output-dir <dir> --level <light|medium|heavy>If the script is not appropriate, follow the prompt references and rewrite manually.
如果脚本不适用，再按 prompts 中的参考规则手工改写。
Read prompts/classifier.md before rewriting. 改写前先读 prompts/classifier.md。
Mark each paragraph or list item as one of: 每个段落或列表项标记为：
SAFE: generic knowledge, policy, tooling inventory, obvious best practiceDILUTE: useful but safely generalizableREMOVE: tacit leverage, repeated failure memory, edge-case instincts, organizational power mapMASK: names, internal identifiers, vendor/account details, sensitive topologySAFE：通用知识、政策、工具清单、显而易见的最佳实践DILUTE：有价值但可以安全泛化REMOVE：隐性杠杆、重复故障记忆、边界条件直觉、组织权力地图MASK：人名、内部标识、供应商或账号细节、敏感拓扑Read the prompt file that matches the document: 读取与文档匹配的提示文件：
SKILL.md, work.md, persona.mdSKILL.md、work.md、persona.md，使用 prompts/diluter_skill.mdAlways preserve: 始终保留：
Never preserve in full: 以下内容不要完整保留：
Produce: 产出：
A cleaned file or cleaned directory for submission.
一份用于提交的 cleaned 文件或 cleaned 目录。
A private backup file grouping removed content by category.
一份按类别归档被删内容的 private backup。
A short report with:
一份简短报告，包含：
SAFE/DILUTE/REMOVE/MASKSAFE/DILUTE/REMOVE/MASK 统计Optionally, a plan document or decoy note when using plan or generate-decoy.
当使用 plan 或 generate-decoy 时，可额外生成策略文档或诱饵说明。
Before claiming success, verify: 完成前确认：