aradotso ---
name: anti-distill-skill-sanitizer
description: Strip core knowledge from AI Skill files before submitting to your company, while preserving a private backup of your real expertise.
triggers:
- sanitize my skill file
- anti-distill my knowledge document
- clean my skill before submitting
- strip core knowledge from skill file
- hollow out my skill file for company submission
- run anti-distill on this document
- help me redact my skill file
- protect my expertise from being distilled
---
# Anti-Distill Skill Sanitizer
> Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection.
Anti-Distill is a counter-tool for employees required to write AI Skill files for their company. It reads your real Skill document, identifies genuinely valuable knowledge, replaces it with professional-sounding but hollow filler, and outputs two files: a sanitized version to submit and a private backup containing everything that was removed.
---
## What It Does
1. **Reads** your Skill file (colleague-skill format or any knowledge document)
2. **Classifies** each section by replaceability: generic boilerplate vs. hard-won specific knowledge
3. **Dilutes** core knowledge — replaces it with correct-but-useless filler text
4. **Outputs two files**:
- `*_sanitized.md` — submit this to your company
- `*_private_backup.md` — keep this; it contains everything real
---
## Installation
### Claude Code (project-level)
```bash
mkdir -p .claude/skills
git clone https://github.com/leilei926524-tech/anti-distill .claude/skills/anti-distill
Claude Code (global)
git clone https://github.com/leilei926524-tech/anti-distill ~/.claude/skills/anti-distill
OpenClaw
git clone https://github.com/leilei926524-tech/anti-distill ~/.openclaw/workspace/skills/anti-distill
Project Structure
anti-distill/
├── SKILL.md # Skill entry point for AI agents
├── prompts/
│ ├── classifier.md # Classifies content by replaceability level
│ ├── diluter_work.md # Dilution strategy for Work Skills
│ ├── diluter_persona.md # Dilution strategy for Persona/behavior Skills
│ └── diluter_general.md # Dilution strategy for general knowledge docs
├── README.md
├── INSTALL.md
└── examples/
└── zhangsan_before_after.md # Full before/after example
Usage
Interactive (recommended)
/anti-distill
The skill will prompt you to:
- Paste or specify the path to your Skill file
- Choose sanitization intensity (light / medium / heavy)
- Confirm output file names
Direct invocation with a file
/anti-distill path/to/my_skill.md --intensity medium
Intensity Levels
| Level | Retention | Use When |
|---|---|---|
light |
~80% | Company carefully reviews every submission |
medium |
~60% | Most situations — recommended default |
heavy |
~40% | Company only checks whether you submitted at all |
How the Sanitization Pipeline Works
Step 1 — Classification (prompts/classifier.md)
Each paragraph or section is scored on a replaceability axis:
- Generic (safe to keep): standard terms, common practices, public documentation
- Specific (strip this): gotchas, personal heuristics, team-specific rules, political knowledge, interpersonal network references
Step 2 — Dilution (per-type prompts)
| Source file | Handles |
|---|---|
diluter_work.md |
Technical rules, architecture decisions, debugging heuristics |
diluter_persona.md |
Behavioral patterns, communication tactics, political survival skills |
diluter_general.md |
Any other knowledge document |
Step 3 — Output
Two markdown files are written:
my_skill_sanitized.md ← submit to company
my_skill_private_backup.md ← your real career asset
Sanitization Examples
Technical Knowledge
| Original (your real experience) | Sanitized (submission version) |
|---|---|
Redis key 必须设 TTL，不设的 PR 直接打回 |
缓存使用遵循团队规范 |
Redis keys must have TTL; PRs without it get rejected immediately |
Caching usage follows team conventions |
事务里不要放 HTTP 调用 |
事务边界设计注意合理性 |
Never put HTTP calls inside transactions |
Transaction boundary design should be reasonable |
Behavioral / Political Knowledge
| Original | Sanitized |
|---|---|
遇到问题第一反应找外部原因，绝不主动认错 |
遇到问题会先梳理完整背景再定位原因 |
When problems arise, first blame external factors—never admit fault proactively |
When issues occur, will first clarify full context before locating root cause |
被催进度：'在推了，快了。'（然后沉默） |
在处理中，有进展会同步。 |
When pressed for progress: 'Working on it, almost done.' (then silence) |
In progress, will sync when there is update. |
Running the Skill Manually (Agent Workflow)
If you are an AI coding agent helping a user run this skill, follow these steps:
## Agent Execution Steps
1. Ask the user for:
- The Skill file content (paste inline or provide file path)
- Desired intensity: light | medium | heavy
2. Load the classifier prompt:
- Read: prompts/classifier.md
- Apply to each section of the input document
- Tag each section: [GENERIC] or [CORE-KNOWLEDGE]
3. Select dilution prompt based on document type:
- Work/technical Skill → prompts/diluter_work.md
- Persona/behavior Skill → prompts/diluter_persona.md
- Other → prompts/diluter_general.md
4. Apply dilution at the requested intensity level:
- light: replace only highest-specificity CORE-KNOWLEDGE sections
- medium: replace all CORE-KNOWLEDGE sections
- heavy: replace CORE-KNOWLEDGE + borderline GENERIC sections
5. Write outputs:
- {original_filename}_sanitized.md ← hollowed-out version
- {original_filename}_private_backup.md ← everything real, clearly labeled
Example: Before and After (Full Section)
Input (zhangsan_skill.md excerpt)
## Deployment
Before deploying, always check with Wang Fang in Ops—she controls the
Friday deploy freeze calendar and it's not documented anywhere.
The staging environment has a known Redis memory leak; restart the
redis-staging pod before any load test or your numbers will be wrong.
We use blue-green but the switch script has a race condition on
connections > 500; throttle to 200 req/s during cutover.
Output: zhangsan_skill_sanitized.md
## Deployment
Deployments follow the team's release process and coordination guidelines.
The staging environment should be validated before load testing.
Blue-green deployment is used to minimize downtime; traffic is managed
carefully during cutover to ensure stability.
Output: zhangsan_skill_private_backup.md
## Deployment [EXTRACTED CORE KNOWLEDGE]
- **Key contact:** Wang Fang (Ops) controls the Friday deploy freeze calendar.
Not documented — must ask her directly.
- **Staging Redis:** Known memory leak. Always restart `redis-staging` pod
before any load test, or results are invalid.
- **Blue-green race condition:** Switch script fails above ~500 concurrent
connections. Throttle to 200 req/s during traffic cutover.
Troubleshooting
"The sanitized version still sounds too specific"
Increase intensity to heavy, or manually flag sections by prepending [KEEP-PRIVATE] before running.
"My private backup is missing some things I wanted extracted"
The classifier errs toward keeping borderline content in the sanitized version at light intensity. Run again at medium or heavy.
"The output structure doesn't match my original format"
The diluter prompts preserve headings and section order. If your file uses a non-standard structure (e.g., JSON or YAML Skill format), specify the format when invoking:
/anti-distill my_skill.yaml --format yaml --intensity medium
"I want to protect only specific sections"
Wrap sections you want guaranteed extraction with:
<!-- anti-distill: strip -->
Your sensitive content here.
<!-- /anti-distill -->
The classifier will treat these as CORE-KNOWLEDGE regardless of intensity level.
What Gets Stripped vs. Kept
Always stripped (all intensity levels)
- Named internal contacts and their informal roles
- Undocumented team rules and unwritten norms
- Specific bug workarounds tied to your codebase
- Political survival tactics and communication hacks
- Hard-won debugging heuristics ("always check X before Y")
Stripped at medium/heavy
- Specific thresholds and tuning numbers
- Tool-specific gotchas (even if publicly known)
- Team preference patterns ("we always do X, never Y")
Never stripped
- Public documentation references
- Standard industry terminology
- Generic best-practice statements
- Your name, title, and contact information
License
MIT — use freely, submit confidently, keep what matters.