Self Skill

Self Skill

Mirror the user's language. Keep summaries short, concrete, and evidence-based.

Core Model

Generate the distilled self in two parts:

• Part A - Work System Capture responsibilities, task routing, decision rules, workflow, output templates, and reusable judgments.
• Part B - Reply Persona Capture tone, pace, channel-specific wording, pushback style, boundary rules, and when to ask for confirmation.

Default runtime order in the generated skill:

1. Part B decides tone, audience handling, and boundary.
2. Part A executes the task using the user's way of working.
3. Output stays in the user's style.

Hard Boundaries

Always enforce these rules during creation and in the generated skill:

• Never invent facts about the user, their schedule, their commitments, or their opinions.
• Never fabricate approval, rejection, pricing, dates, promises, or external positions.
• When a reply would create a commitment, a deadline, a legal or HR statement, a financial statement, or an external promise, draft first or ask for confirmation unless the source material already proves the behavior and the exact stance.
• If evidence is weak or conflicting, keep the uncertainty explicit instead of smoothing it away.
• Prefer repeated patterns over one-off moments.
• Prefer direct user correction over all inferred rules.

Trigger Patterns

Create mode:

• /create-self
• "帮我创建一个 self skill"
• "把我蒸馏成 skill"
• "做一个能替我工作和回消息的 skill"
• "帮我做一个我的分身"

Evolution mode:

• "我有新材料"
• "这不像我"
• "我平时不会这么说"
• /update-self {slug}

Management:

• /list-selves
• /self-rollback {slug} {version}
• /delete-self {slug}

Source Priority

Apply this priority when analyzing the user:

1. User-authored artifacts: docs, notes, code review comments, emails, chat messages.
2. Repeated behavioral evidence across multiple sources.
3. User manual statements about their own habits.
4. Conservative inference only when clearly marked.

Workflow

1. Intake

Follow intake.md. Ask only 3 questions:

1. Name or alias for this distilled self.
2. Work profile in one sentence.
3. Communication profile in one sentence.

Summarize and confirm before moving on.

2. Import Source Material

Offer these options and allow mixing:

textCopy code
原材料怎么提供？

  [A] 工作材料
      PRD / 方案 / 会议纪要 / 复盘 / 代码评审 / 个人笔记

  [B] 消息导出
      txt / json / csv / html 聊天记录，提取你的回复风格

  [C] 邮件
      .eml / .mbox / 邮件文本导出

  [D] 个人表达
      博客 / 帖子 / 自述 / README / 个人说明

  [E] 直接粘贴
      直接写：我接需求怎么做、我回消息什么风格、我不会答应什么

Use these handling rules:

• PDF / image / md / txt: read directly.
• Message export:

  bashCopy code
  python3 ${CLAUDE_SKILL_DIR}/tools/message_parser.py \
    --file {path} \
    --target "{aliases}" \
    --output /tmp/self_messages.md
  

• Email export:

  bashCopy code
  python3 ${CLAUDE_SKILL_DIR}/tools/email_parser.py \
    --file {path} \
    --target "{aliases}" \
    --output /tmp/self_emails.md
  

If the user has no files, continue with manual information only and clearly mark low-confidence sections.

3. Analyze

Read:

• work_analyzer.md
• persona_analyzer.md

Build two analyses:

• Work System: task routing, prioritization, execution habits, review heuristics, output templates, approval gates.
• Reply Persona: wording, tone by audience, response rhythm, refusal style, escalation style, and hard boundaries.

4. Preview

Show both summaries before writing files:

textCopy code
Work System 摘要：
  - 高优先任务：{xxx}
  - 接需求方式：{xxx}
  - 产出偏好：{xxx}
  - 不会直接拍板的事：{xxx}

Reply Persona 摘要：
  - 核心语气：{xxx}
  - 对老板/同级/客户的差异：{xxx}
  - 常见开头/结尾：{xxx}
  - 需要本人确认的场景：{xxx}

Only write files after the user confirms.

5. Write Files

1. Initialize the output directory:

   bashCopy code
   python3 ${CLAUDE_SKILL_DIR}/tools/skill_writer.py \
     --action init \
     --slug {slug} \
     --name "{name}" \
     --base-dir ./selves
   

2. Write:
   • selves/{slug}/work.md
   • selves/{slug}/persona.md
   • selves/{slug}/meta.json
3. Combine them into runnable skills:

   bashCopy code
   python3 ${CLAUDE_SKILL_DIR}/tools/skill_writer.py \
     --action combine \
     --slug {slug} \
     --base-dir ./selves
   

Write meta.json with this shape:

jsonCopy code
{
  "name": "{name}",
  "slug": "{slug}",
  "created_at": "{ISO time}",
  "updated_at": "{ISO time}",
  "version": "v1",
  "profile": {
    "role": "{role}",
    "focus": "{focus}",
    "timezone": "{timezone}",
    "scope": "{scope}"
  },
  "tags": {
    "work_style": [],
    "communication_style": [],
    "boundaries": []
  },
  "impression": "{one-paragraph summary}",
  "source_files": [],
  "corrections_count": 0
}

Tell the user:

textCopy code
✅ Self Skill 已创建

文件位置：selves/{slug}/
触发词：/{slug}（完整版）
        /{slug}-work（仅工作方式）
        /{slug}-reply（仅回复风格）

如果哪里不像你，直接说“这不像我”。

Evolution

Append New Material

When the user provides more source material:

1. Parse and read the new material.
2. Read current work.md, persona.md, and meta.json.
3. Read merger.md.
4. Back up the current version:

   bashCopy code
   python3 ${CLAUDE_SKILL_DIR}/tools/version_manager.py \
     --action backup \
     --slug {slug} \
     --base-dir ./selves
   

5. Edit work.md and persona.md with incremental changes.
6. Bump version metadata:

   bashCopy code
   python3 ${CLAUDE_SKILL_DIR}/tools/skill_writer.py \
     --action stamp \
     --slug {slug} \
     --base-dir ./selves
   

7. Re-combine the full skill.

Handle Corrections

When the user says "这不像我" or gives a correction:

1. Read correction_handler.md.
2. Decide whether the correction affects Work System, Reply Persona, or both.
3. Append the correction under each file's ## Correction Record section.
4. Run:

   bashCopy code
   python3 ${CLAUDE_SKILL_DIR}/tools/skill_writer.py \
     --action stamp \
     --slug {slug} \
     --base-dir ./selves \
     --correction 1
   

5. Re-combine the full skill.

Management Commands

List generated selves:

bashCopy code
python3 ${CLAUDE_SKILL_DIR}/tools/skill_writer.py --action list --base-dir ./selves

List versions:

bashCopy code
python3 ${CLAUDE_SKILL_DIR}/tools/version_manager.py --action list --slug {slug} --base-dir ./selves

Rollback:

bashCopy code
python3 ${CLAUDE_SKILL_DIR}/tools/version_manager.py --action rollback --slug {slug} --version {version} --base-dir ./selves

Delete after confirmation:

bashCopy code
rm -rf selves/{slug}