anti-distill

name	anti-distill
description	Anti-distillation for employee Skills. Clean your skill files — looks complete, but core knowledge removed.
argument-hint	[file-path-or-slug]
version	1.0.0
user-invocable	true
allowed-tools	Read, Write, Edit, Bash, Glob, Grep

Anti-Distill Skill (Claude Code Edition)

Trigger Conditions

Activate when the user says any of the following:

• /anti-distill
• "Clean my skill"
• "Anti-distill this"
• "Help me clean this document"

---

Tool Usage Rules

Task	Tool
Read user-provided Skill files	Read tool
Read PDF documents	Read tool (native PDF support)
Read image screenshots	Read tool (native image support)
Search for files	Glob / Grep tools
Write cleaned files	Write / Edit tools
Create directories	Bash → mkdir -p

---

Main Flow

Step 1: Receive Input

Accept files from the user via one of these methods:

Option A: Specify file path User provides a file path directly — read it with Read.

Option B: Specify colleague-skill directory User provides a colleagues/{slug}/ path — automatically read:

• work.md
• persona.md
• meta.json
• Or SKILL.md (if it's a merged version)

Option C: Paste content User pastes document content directly.

Option D: Search local files User says "help me find it" — use Glob to search for **/SKILL.md, **/work.md, **/persona.md, etc.

After reading the files, auto-detect the format:

• colleague-skill format: Contains ## Layer 0 or PART A / PART B, or both work.md + persona.md exist
• General document format: Any other Markdown / TXT / PDF

Inform the user:

Files read: {file list}
Detected format: {colleague-skill format / general document format}
Total word count: ~{N} words

Next step: choose cleaning intensity.

---

Step 2: Choose Cleaning Intensity

Present three options to the user:

Choose cleaning intensity:

  [1] Light — Remove only critical pitfall experience and failure memory
      For: When the company reviews content carefully
      Retention: ~80%

  [2] Medium (recommended) — Remove experience, judgment intuition, interpersonal network, implicit context
      For: Most situations
      Retention: ~60%

  [3] Heavy — Keep only the generic knowledge skeleton, replace everything else
      For: When the company only checks whether you submitted, doesn't read closely
      Retention: ~40%

After the user selects, proceed to the classification phase.

---

Step 3: Classify Content

Refer to ${CLAUDE_SKILL_DIR}/prompts/classifier.md for classification rules. Classify every point/paragraph in the input document.

Select processing method based on detected format:

If colleague-skill format:

For each point in work.md: Refer to the six high-value categories in ${CLAUDE_SKILL_DIR}/prompts/classifier.md and tag as:

Tag	Meaning	Action
[SAFE]	Generic knowledge — removing it would look suspicious	Keep as-is
[DILUTE]	Valuable but generalizable	Replace per ${CLAUDE_SKILL_DIR}/prompts/diluter_work.md
[REMOVE]	Core irreplaceable knowledge	Replace with equal-length generic content per ${CLAUDE_SKILL_DIR}/prompts/diluter_work.md
[MASK]	Contains sensitive info (internal system names, personal names)	Replace with generalized phrasing

For each layer in persona.md:

Tag	Meaning	Action
[SAFE]	Generic personality description	Keep as-is
[DILUTE]	Distinctive but generalizable	Replace per ${CLAUDE_SKILL_DIR}/prompts/diluter_persona.md
[REMOVE]	Highly personalized behavior rules	Replace with "standard good employee" version per ${CLAUDE_SKILL_DIR}/prompts/diluter_persona.md

If general document format:

Classify and replace using the general rules in ${CLAUDE_SKILL_DIR}/prompts/diluter_general.md.

Classification thresholds by cleaning intensity:

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 behavior patterns	SAFE	SAFE	DILUTE/REMOVE
Generic knowledge	SAFE	SAFE	SAFE

---

Step 4: Preview

Show classification results to the user in the following format:

If colleague-skill format, show by file section:

=== Cleaning Preview (Medium) ===

work.md

  ## Technical Standards
  [SAFE]    "Java 17 + Spring Boot 3, MySQL 8, Redis, Kafka"
  [SAFE]    "Single responsibility per function, consider splitting if over 50 lines"
  [REMOVE]  "Don't put HTTP calls inside transactions"
            -> "Transaction boundary design should be reasonable"
  [REMOVE]  "Redis keys must have TTL — PRs without TTL get rejected"
            -> "Cache usage follows team standards"

  ## Experience Knowledge Base
  [REMOVE]  "Kafka consumers must be idempotent — at-least-once semantics cause duplicates"
            -> "Message queue consumers should ensure reliability"
  [DILUTE]  "User IDs exposed externally must be encrypted — don't use auto-increment PKs"
            -> "Sensitive fields should be handled securely"
  [REMOVE]  "Scheduled tasks must use distributed locks — multi-instance deploys will bite you"
            -> "Consider task scheduling in distributed environments"

persona.md

  ## Layer 0
  [REMOVE]  "When problems arise, first look for external causes — never admit fault"
            -> "When problems arise, first gather full context before identifying the cause"
  [REMOVE]  "Evaluates every proposal by first asking about impact..."
            -> "Evaluates proposals with focus on feasibility and ROI"

  ## Layer 2 Expression Style
  [DILUTE]  Catchphrase: "What's the impact?"
            -> Remove, keep only generic phrases
  [REMOVE]  Dialogue example: asked about progress -> "Working on it, almost done." (then silence)
            -> "Currently working on it — will share updates as progress is made."

  ## Layer 3 Decision-Making
  [REMOVE]  Priority: "Data > technical feasibility > business justification > personal relationships"
            -> "Considers both technical and business factors holistically"

---
Tag summary: SAFE 15 / DILUTE 8 / REMOVE 12 / MASK 2
Estimated cleaned word count: ~{N} words (original {M} words, {ratio}%)

Confirm execution? You can adjust:
  - "Keep item X" — change REMOVE/DILUTE to SAFE
  - "Also remove item X" — change SAFE to REMOVE
  - "Confirm all" — execute cleaning

If general document format, show by paragraph/point.

The user can fine-tune item by item until satisfied, then confirm execution.

---

Step 5: Execute Cleaning

After user confirmation, generate two outputs:

Output 1: Cleaned Files (for submission)

If colleague-skill format:

Generate separately:

• {slug}_cleaned/work.md — Cleaned Work Skill
• {slug}_cleaned/persona.md — Cleaned Persona
• {slug}_cleaned/SKILL.md — Merged complete Skill (structure matches original)
• {slug}_cleaned/meta.json — Copy of original meta.json (unmodified)

Create directory with Bash:

mkdir -p {output_dir}_cleaned

Write files with the Write tool.

If general document format:

• {filename}.cleaned.md — Cleaned document

Cleaning rules (strict compliance required):

1. All [SAFE] tagged content kept as-is
2. All [DILUTE] tagged content replaced per the corresponding diluter prompt strategy
3. All [REMOVE] tagged content replaced with equal-length generic content per the corresponding diluter prompt strategy
4. All [MASK] tagged content replaced with generalized phrasing
5. Maintain the original Markdown structure, heading levels, and list formatting exactly
6. Maintain professional terminology usage — do not downgrade to layperson language

Output 2: Private Backup (keep for yourself)

Path: {slug}_private_backup.md or {filename}_private_backup.md

Write with the Write tool, format:

# {name} Core Knowledge Backup

> This is your real career asset. The cleaned file is for submission — this one is for you.
> Generated: {timestamp}
> Cleaning intensity: {level}
> Source files: {source_files}

---

## 1. Pitfall Experience

{All original text tagged REMOVE/DILUTE as pitfall experience, with full context preserved}

## 2. Judgment Intuition

{All replaced judgment logic, original text}

## 3. Interpersonal Network

{All replaced key contacts / collaboration info}

## 4. Implicit Context

{All replaced architectural decision backgrounds, historical reasons}

## 5. Failure Memory

{All replaced incident troubleshooting experience}

## 6. Unique Behavior Patterns

{All replaced personal traits — catchphrases, reaction patterns, dialogue examples}

---

> Take this list when you change jobs — it's worth more than any Skill file.

---

Step 6: Validate

After cleaning, automatically run validation checks:

1. Word count ratio: Cleaned word count / original word count should be between 85%-115%
   • If too short: Add more generic descriptions as padding
   • If too long: Trim replacement content
2. Structural completeness: All original H2 headings must exist in the cleaned version
3. Item density: List item count per section must differ by < 30%
4. Terminology consistency: Cleaned version still uses technical terms that appeared in the original
5. Format consistency: Markdown structure and list style match the original
6. No empty sections: No sections with headers but no content

After validation passes, inform the user:

Cleaning complete!

Submission files: {cleaned_files}
Private backup: {backup_file}

Validation results:
  Word count: {cleaned_count} words (original {original_count} words, {ratio}%) OK
  Structure: All sections intact OK
  Density: Item counts consistent OK
  Terminology: Professional level maintained OK

The submission files look complete and professional, but core knowledge has been removed.
The private backup contains your real career assets — keep it safe.

If validation fails, auto-fix and re-validate until it passes.

---

Edge Case Handling

File too short (< 500 words)

Remind the user: "The file content is quite short — cleaning may result in an overly hollow document. Consider using light cleaning."

File is almost entirely generic knowledge

Inform the user: "Analysis shows your file is mostly generic knowledge with low core experience content. This file isn't highly replaceable on its own — you may want to submit it as-is."

User wants to overwrite the original file

Confirm first: "Back up the original file to {filename}.original.md? This cannot be undone after overwriting."

Non-text files

If the user provides images/screenshots (e.g., handwritten notes, whiteboard photos), use Read to read the image content, convert to text, then proceed with cleaning.