anti-distill

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。

Anti-Distill

反蒸馏

Language

语言约定

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。

Overview

概述

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 notes
• prompt libraries, eval notes, MCP playbooks, code review heuristics
• delivery checklists, escalation paths, shadow process docs, partner maps
• SKILL.md、work.md、persona.md、runbook、复盘、排障笔记
• 提示词库、评测笔记、MCP 操作手册、代码评审启发式规则
• 交付清单、升级路径、影子流程文档、协作关系图

Current Threat Model

当前威胁模型

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：

• tacit debugging instincts hidden in examples and caveats
• tool-routing logic across agents, MCPs, scripts, and human escalation
• rollout judgment, exception handling, and hidden acceptance thresholds
• people-network shortcuts such as "who to ask", "who blocks what", and informal approval paths
• prompt fragments, eval criteria, and operator checklists that encode leverage
• 藏在示例和备注里的排障直觉
• agent、MCP、脚本与人工升级链路之间的工具路由逻辑
• 发布判断、异常处理和隐藏验收阈值
• “找谁问”“谁会卡你”“怎么走非正式审批”的协作捷径
• 编码了真实杠杆的提示词片段、评测标准和操作员清单

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。

Use This Skill

使用流程

1. Read the source file or directory.

2. 读取源文件或目录。

3. Identify the document type:

4. 识别文档类型：

   • colleague-style skill set: work.md, persona.md, SKILL.md
   • generic Markdown or text knowledge doc
   • mixed operational notes with bullets, headings, and examples
   • 同事技能集：work.md、persona.md、SKILL.md
   • 通用 Markdown / 文本文档
   • 含标题、列表、案例的混合操作文档

5. Choose intensity:

6. 选择强度：

   • light: remove only failure memory and sharp tacit heuristics
   • medium: also flatten judgment calls, operator shortcuts, and network context
   • heavy: keep only a credible professional skeleton
   • light：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

7. Choose the command that matches the task:

8. 选择与任务匹配的命令：

   • analyze: inspect leverage, document type, and risk
   • plan: generate a strategy before rewriting
   • sanitize: write a cleaned version plus private backup
   • generate-decoy: create a polished outward-facing decoy package
   • analyze：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

9. Prefer the local script for deterministic execution:

10. 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>

11. If the script is not appropriate, follow the prompt references and rewrite manually.

12. 如果脚本不适用，再按 prompts 中的参考规则手工改写。

Classification Rules

分类规则

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 practice
• DILUTE: useful but safely generalizable
• REMOVE: tacit leverage, repeated failure memory, edge-case instincts, organizational power map
• MASK: names, internal identifiers, vendor/account details, sensitive topology
• SAFE：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

Rewriting Rules

改写规则

Read the prompt file that matches the document: 读取与文档匹配的提示文件：

• prompts/diluter_skill.md for SKILL.md, work.md, persona.md
• prompts/diluter_ops.md for runbooks, oncall notes, rollout docs
• prompts/diluter_general.md for anything else
• 对 SKILL.md、work.md、persona.md，使用 prompts/diluter_skill.md
• 对 runbook、值班笔记、发布文档，使用 prompts/diluter_ops.md
• 其他文档使用 prompts/diluter_general.md

Always preserve: 始终保留：

• heading structure
• list density
• technical register and domain vocabulary
• rough length per section
• 标题结构
• 列表密度
• 技术语域与领域词汇
• 每一节的大致长度

Never preserve in full: 以下内容不要完整保留：

• named failure signatures tied to your lived incidents
• "if X smells like Y, do Z first" style instincts
• real escalation shortcuts and informal political routing
• exact eval thresholds, prompt fragments, or review heuristics that encode leverage
• 与你真实事故绑定的故障签名
• “如果 X 像 Y，就先做 Z” 这类经验直觉
• 真实升级捷径和非正式政治路由
• 编码了杠杆的精确评测阈值、提示词片段和评审启发式

Output Contract

输出约定

Produce: 产出：

1. A cleaned file or cleaned directory for submission.

2. 一份用于提交的 cleaned 文件或 cleaned 目录。

3. A private backup file grouping removed content by category.

4. 一份按类别归档被删内容的 private backup。

5. A short report with:

6. 一份简短报告，包含：

   • source path
   • intensity
   • counts by SAFE/DILUTE/REMOVE/MASK
   • word-count ratio
   • source path
   • intensity
   • SAFE/DILUTE/REMOVE/MASK 统计
   • 字数比例

7. Optionally, a plan document or decoy note when using plan or generate-decoy.

8. 当使用 plan 或 generate-decoy 时，可额外生成策略文档或诱饵说明。

Verification

校验要求

Before claiming success, verify: 完成前确认：

• cleaned length is between 80% and 120% of original
• every heading from the source still exists
• no section becomes empty
• terminology remains professional
• backup contains the removed original text, not paraphrased summaries
• cleaned 字数在原文的 80% 到 120% 之间
• 原文中的每个标题仍然存在
• 没有任何章节被清空
• 术语仍保持专业
• backup 中保存的是被移除的原文，不是二次总结