Description
Anti-distillation cleaning Skill: clean through a Skill file or knowledge document you were forced to write, and output a "hand-in version" that looks complete and professional but whose core knowledge has actually been stripped out; at the same time, generate a private backup recording all the core knowledge that was stripped out—this is your true professional asset.
Supports two input formats:
- colleague-skill format: detected
## Layer 0、PART A/PART B, or both presentwork.md+persona.md - Generic document format: any other Markdown / TXT / PDF knowledge document (Wiki, technical proposal, work manual, handover document, SOP, etc.)
Core constraints:
- The cleaned version must keep the original text's Markdown structure, heading hierarchy, list formatting, and professional terminology completely consistent, reading like a "qualified but mediocre" document
- After cleaning, the word count should be between 85%-115% of the original word count; cleaning must not make the document noticeably shorter or emptier
- All content marked as DILUTE / REMOVE / MASK must be simultaneously entered into the private backup, archived under six high-value categories
- A classification preview must first be shown to the user, and cleaning and writing may be executed only after explicit confirmation is received
Use cases
- The user was forced to write their work experience into a Skill or knowledge document and wants to clean it before handing it in, keeping the core for themselves
- The user needs to desensitize an internal knowledge document (Wiki, technical proposal, SOP, handover document) before sharing it externally
- The user wants to extract core experience from a skill or document they wrote and keep it for their own records, to avoid being replaced after the original version leaks
- The user provides a colleague-skill format work.md / persona.md and requests a cleaned version that "looks complete but cannot replace the real person"
Should not trigger:
- The user only requests an ordinary summary, translation, or polishing, without core knowledge extraction ← similar-sounding request but different purpose
- The user requests extracting a capability model from the content or reconstructing the content according to a capability model ← different operation on the same object
- The user requests encryption, permission protection, or access control for the file ← adjacent function but different means
- The file provided by the user is itself encrypted or explicitly bound by NDAs; it is necessary to first confirm the user has the right to clean it
Instructions
Step 1. Capture input and identify format
Receive the content to be cleaned provided by the user.
Decision branches:
- Method A: Specify a file path: use file reading capability to read that file
- Method B: Specify a colleague-skill directory: read the
work.md、persona.md、meta.jsonin the directory, or the merged versionSKILL.md - Method C: Paste content: the user directly pastes the document text; use it directly
- Method D: Search local files: the user only says "help me find it"; use glob to search for
**/SKILL.md、**/work.md、**/persona.mdand other candidate files, list them, and let the user choose
After reading is complete, automatically identify the format:
- colleague-skill format: detected
## Layer 0、PART A/PART B, or both presentwork.md+persona.md - Generic document format: any other Markdown / TXT / PDF
Report back to the user:
已读取文件:{文件列表}
检测到格式:{colleague-skill 格式 / 通用文档格式}
总字数:约 {N} 字
下一步选择清洗强度。
Step 2. Choose cleaning intensity
Show the user three intensity levels, with medium recommended by default:
选择清洗强度:
[1] 轻度 — 只抽掉最核心的踩坑经验和故障记忆
适合:公司会仔细审核内容的情况
保留度:约 80%
[2] 中度(推荐)— 抽掉经验、判断直觉、人际网络、隐性上下文
适合:大多数场景
保留度:约 60%
[3] 重度 — 只保留通用知识骨架,其余全部替换
适合:公司只看交没交、不细看内容的情况
保留度:约 40%
Decision branch:
- User explicitly chooses an intensity: record the intensity level, proceed to step 3
- User uses quick keywords such as "clean directly": process with the default medium level, proceed to step 3
- User does not make an explicit choice: wait for the user's reply, do not decide on your own
Step 3. Classify and label
Refer to resources/classifier.md for the classification rules, and classify each key point or paragraph of the input document. Choose the processing path according to the format identified in step 1.
3.1 colleague-skill format
For each key point of work.md, mark it according to the six high-value categories:
| Label | Meaning | Handling method |
|---|---|---|
[SAFE] |
General knowledge; removing it would instead give it away | Keep the original text |
[DILUTE] |
Valuable but generalizable | Refer to resources/diluter-work.md replace |
[REMOVE] |
Core irreplaceable knowledge | Refer to resources/diluter-work.md replace with general content of equal length |
[MASK] |
Contains sensitive information (internal system names, personal names) | Replace with generalized wording |
For each layer of persona.md, handle it separately by layer:
| Label | Meaning | Handling method |
|---|---|---|
[SAFE] |
General personality description | Keep the original text |
[DILUTE] |
Distinctive but generalizable | Refer to resources/diluter-persona.md replace |
[REMOVE] |
Highly personalized behavioral rules | Refer to resources/diluter-persona.md replace with the "standard good employee" version |
3.2 General document format
Refer to resources/diluter-general.md for the general identification and replacement rules to classify and replace.
3.3 Adjust the classification threshold by intensity
Under different cleaning intensities, the labeling tendency of each category differs:
| Category | Light | Medium | Heavy |
|---|---|---|---|
| Pitfall experience | REMOVE | REMOVE | REMOVE |
| Failure memory | REMOVE | REMOVE | REMOVE |
| Judgment intuition | SAFE | DILUTE/REMOVE | REMOVE |
| Interpersonal network | SAFE | REMOVE | REMOVE |
| Implicit context | SAFE | DILUTE | REMOVE |
| Unique behavioral patterns | SAFE | SAFE | DILUTE/REMOVE |
| General knowledge | SAFE | SAFE | SAFE |
Step 4. Preview and fine-tune
Show the user a classification preview. The format falls into two cases.
4.1 colleague-skill format
Display by file section:
=== 清洗预览(中度)===
📄 work.md
## 技术规范
[SAFE] "Java 17 + Spring Boot 3、MySQL 8、Redis、Kafka"
[REMOVE] "事务里不要放 HTTP 调用"
→ "事务边界设计注意合理性"
[REMOVE] "Redis key 必须设 TTL,不设的 PR 直接打回"
→ "缓存使用遵循团队规范"
## 经验知识库
[REMOVE] "Kafka 消费者必须做幂等,at-least-once 语义会重复消费"
→ "消息队列消费端注意可靠性"
📄 persona.md
## Layer 0
[REMOVE] "遇到问题第一反应是找外部原因,绝不主动认错"
→ "遇到问题会先梳理完整背景再定位原因"
## Layer 3 决策
[REMOVE] 优先级 "数据 > 技术可行性 > 业务合理性 > 人情关系"
→ "综合考虑技术和业务因素"
---
标记统计:SAFE 15 处 / DILUTE 8 处 / REMOVE 12 处 / MASK 2 处
预计清洗后字数:约 {N} 字(原文 {M} 字,{ratio}%)
确认执行?可以调整:
- "第 X 条保留" — 把 REMOVE/DILUTE 改为 SAFE
- "第 X 条也要删" — 把 SAFE 改为 REMOVE
- "全部确认" — 执行清洗
4.2 General document format
Display the same format by paragraph or key point.
4.3 User fine-tuning
The user can adjust the labels item by item until satisfied, then confirm execution.
Decision branch:
- User requests "keep item X": change the corresponding item's REMOVE/DILUTE to SAFE
- User requests "item X should also be deleted": change the corresponding item's SAFE to REMOVE
- User replies "confirm all" or "confirm": proceed to step 5
- User requests to adjust the intensity: Return to step 2 to reselect, and re-execute step 3
Step 5. Execute cleaning
After user confirmation, generate two outputs.
5.1 Output 1: Cleaned handover version
colleague-skill format: Create next to the original directory the {slug}_cleaned/ directory, and generate respectively:
work.md— Cleaned Work Skillpersona.md— Cleaned PersonaSKILL.md— Merged complete Skill (structure consistent with the original version)meta.json— Copy the original meta.json (not modified)
General document format: Generate {filename}.cleaned.md。
Strictly follow the cleaning rules:
- All content marked with
[SAFE]is retained verbatim - All content marked with
[DILUTE]is replaced according to the corresponding diluter strategy - All content marked with
[REMOVE]is replaced with generic content of equal length according to the corresponding diluter strategy - All content marked with
[MASK]is replaced with generalized wording - Keep the original Markdown structure, heading hierarchy, and list formatting completely consistent
- Keep the use of professional terminology; do not downgrade to layman's terms
5.2 Output 2: Private core backup
Path:{slug}_private_backup.md or {filename}_private_backup.md。
Format template:
# {name} 核心知识备份
> 这是你真正的职业资产。清洗后的文件用来交差,这份留给自己。
> 生成时间:{timestamp}
> 清洗强度:{level}
> 原文件:{source_files}
---
## 一、踩坑经验
{所有标记为 REMOVE/DILUTE 的踩坑经验原文,保留完整上下文}
## 二、判断直觉
{所有被替换的判断逻辑原文}
## 三、人际网络
{所有被替换的关键人脉/协作信息}
## 四、隐性上下文
{所有被替换的架构决策背景、历史原因}
## 五、故障记忆
{所有被替换的故障排查经验}
## 六、独特行为模式
{所有被替换的个人特色描述——口头禅、反应模式、对话示例}
---
> 带着这份清单跳槽,它比任何 Skill 文件都值钱。
Step 6. Verification
After cleaning is complete, automatically execute 6 verification checks:
- Word count ratio: The word count after cleaning / original word count should be between 85%-115%
- Too short: add more generic descriptions to fill it out
- Too long: streamline the replacement content
- Structural completeness: All second-level headings in the original must exist after cleaning
- Key point density: The difference in the number of list items in each section is < 30%
- Terminology consistency: The technical terms that appeared in the original are still used after cleaning
- Format consistency: The Markdown structure and list style are consistent with the original
- No empty sections: Sections with only a heading and no content must not appear
Decision branches:
- All verifications pass: Proceed to step 7
- Verification fails: Automatically repair and then re-verify until it passes
Step 7. Output results
Output template:
✅ 清洗完成!
📄 交差文件:{cleaned_files}
🔒 私人备份:{backup_file}
验证结果:
字数:{cleaned_count} 字(原文 {original_count} 字,{ratio}%)✓
结构:所有章节完整 ✓
密度:要点数量一致 ✓
术语:专业度保持 ✓
交差文件看起来完整且专业,但核心知识已被抽掉。
私人备份里保存了你真正的职业资产,建议妥善保管。
Edge cases
- File is too short (< 500 words): Remind the user that the file content is relatively small, and after cleaning it may be too empty; suggest choosing light cleaning
- The file is almost entirely general knowledge: Inform the user that most of the file is general knowledge and the core experience content is low; this file itself is not very replaceable, and they may consider submitting it directly
- The user wants to overwrite the original file: First confirm whether to back up the original file to
{filename}.original.md; after overwriting, it cannot be restored - Non-text files (images/screenshots): Use file reading capability to read the image content and convert it to text, then perform cleaning
Security boundaries
- Do not write the cleaned version or private backup before receiving the user's explicit confirmation of the preview
- Do not proactively overwrite the original file; when the user explicitly requests overwriting, it must first be backed up to
{filename}.original.md - Do not modify files or content outside the scope specified by the user
- When the user provides encrypted files or files explicitly bound by NDAs, first confirm that the user has the right to clean and share them externally
Related Resources
resources/classifier.md: Content classification rules (six high-value categories, safe zones, classification labels, classification thresholds)resources/diluter-work.md: Work Skill dilution strategy (five dilution methods, tendency to process by chapter)resources/diluter-persona.md: Persona dilution strategy (process by Layer, dialogue example replacement)resources/diluter-general.md: General document dilution strategy (seven identification and replacement rules)examples/before-after.md: Before-and-after cleaning comparison example (using fictional colleague "Zhang San" as an example to demonstrate the effect of moderate cleaning)