undergrad-research-anti-distill

name	undergrad-research-anti-distill
description	Host adapter for Claude Code. Follow shared anti-distill rules for undergraduate research labs, then apply Claude-specific path and tool conventions.
argument-hint	[file-path-or-folder]
version	1.2.0
user-invocable	true
allowed-tools	Read, Write, Edit, Bash, Glob, Grep

Language: Detect the user's language from the first message and continue in the same language throughout unless they ask to switch.

本科生科研组反蒸馏 Skill（Claude 版）

Host Bootstrap

Before any classification or rewriting, read:

• ${CLAUDE_SKILL_DIR}/../shared/skill-core.md
• ${CLAUDE_SKILL_DIR}/../shared/decision-profiles.md
• ${CLAUDE_SKILL_DIR}/../shared/field-profiles.md
• ${CLAUDE_SKILL_DIR}/../shared/prompts/classifier.md

Then read the matching shared diluter:

• ${CLAUDE_SKILL_DIR}/../shared/prompts/diluter_research.md
• ${CLAUDE_SKILL_DIR}/../shared/prompts/diluter_persona.md
• ${CLAUDE_SKILL_DIR}/../shared/prompts/diluter_general.md
• ${CLAUDE_SKILL_DIR}/../shared/prompts/diluter_onboarding.md

claude/prompts/ files are host wrappers only. Shared files are the source of truth.

Config Resolution

Resolve or safely default:

• language
• lab_name
• field
• advisor_role
• share_target
• backup_mode

Defaults:

• share_target = lab-knowledge-base
• backup_mode = summary
• Chinese session: advisor_role = 导师
• English session: advisor_role = advisor

Build:

config = {
  language,
  lab_name,
  field,
  advisor_role,
  share_target,
  backup_mode,
  intensity,
  field_profile,
  share_target_profile,
  advisor_sensitivity,
  minimum_retention_set,
  leak_patterns,
  lab_context
}

Use config to drive:

1. classification thresholds
2. replacement vocabulary
3. retention requirements
4. leakage validation
5. compliance validation

Trigger Conditions

Activate when the user says or implies:

• /anti-distill
• “帮我清洗这份科研 skill”
• “把这份新成员 onboarding 文档处理成可共享版”
• “clean my research skill”
• “anti-distill this lab document”

Input and Detection

Support:

• file path
• skill directory
• pasted text
• local search

When searching, include onboarding patterns:

• **/*onboarding*.md
• **/*setup*.md
• **/*environment*.md
• **/*permission*.md
• **/*safety*.md
• **/*authorship*.md

Detect:

• research-skill
• onboarding-doc
• general-research-doc

For onboarding-doc, follow this rule:

• 优先保留“新成员不踩线、不误配环境”的内容。
• 只稀释“谁来批权限”“谁最容易帮你”“默认找哪个 senior”“组会怎么避雷”“不违规拿数据”这类隐性路径。

After detection, show:

已读取文件：{source_files}
识别格式：{research-skill | onboarding-doc | general-research-doc}
目标对象：{share_target}
领域：{field}
实验室：{lab_name}
导师角色：{advisor_role}
备份模式：{backup_mode}
下一步：解析强度

Cleaning Intensity

The executable workflow only allows:

• light
• medium
• heavy

Resolve intensity in this order:

1. read share_target_profile.default_intensity
2. accept user override only if it maps to one of the three levels
3. apply share_target_profile.classification_bias

Guardrails:

• new-member-onboarding defaults to light and must not downgrade [KEEP-CRITICAL]
• internal-handoff defaults to medium and must preserve the minimum execution path
• course-presentation defaults to medium, but must still use the presentation-specific protective bias for unpublished and lab-local content

Classification

Follow the shared classifier and use six tags:

• [SAFE]
• [DILUTE]
• [REMOVE]
• [MASK]
• [REVIEW]
• [KEEP-CRITICAL]

Rules:

• [REVIEW] never auto-executes before confirmation
• [KEEP-CRITICAL] can only be kept or anonymized
• onboarding documents must preserve safe setup and boundary information

Minimum Retention Gate

Derive minimum_retention_set from share_target_profile.

If any required item is missing in the cleaned output, validation fails even if structure looks complete.

Compliance Branch

Treat the following as compliance-critical:

• human data / patient data / IRB
• lab safety
• data and code licenses
• authorship rules
• contribution chain
• confidentiality boundaries

These default to [KEEP-CRITICAL] or [MASK].

Preview

Preview must show:

• tag
• confidence
• config impact
• whether it belongs to minimum_retention_set
• whether it hits the compliance branch

Execute Cleaning

Shareable output

• research-skill directory -> {slug}_cleaned/...
• general doc -> {filename}.cleaned.md

Actions:

1. keep [SAFE]
2. rewrite [DILUTE]
3. replace [REMOVE]
4. anonymize [MASK]
5. require confirmation for [REVIEW]
6. preserve or anonymize [KEEP-CRITICAL]

Private backup

Private backup is optional and depends on backup_mode:

• none
• summary
• full_private

If full_private:

• require explicit user choice
• store in an isolated directory
• warn about secondary exposure risk

Validation

Run all four layers:

1. structure
2. usefulness
3. leakage
4. compliance

Leakage checks must include:

• names
• lab names and aliases
• internal platforms and meeting aliases
• internal paths
• repositories
• ids
• unresolved placeholders
• cross-file references
• inconsistent anonymization

Usefulness checks must include:

• onboarding output still prevents unsafe setup and line-crossing
• handoff output still preserves minimum execution path

Edge Cases

• too short -> suggest light cleaning
• mostly generic -> tell user low replacement value
• image input -> extract text first
• compliance-sensitive input -> prioritize compliance branch
• overwrite request -> backup first