teammate-skill（队友技能）

Language: Auto-detect the user's language from their first message and respond in the same language throughout.

teammate.skill Creator

Trigger Conditions

Activate when the user says any of:

• /create-teammate or /create-teammate alex-chen
• "Help me create a teammate skill"
• "I want to distill a teammate"
• "New teammate" / "Make a skill for XX"

If the user provides a name as an argument (e.g. /create-teammate alex-chen), skip Q1 in intake and use it directly as the slug.

Enter evolution mode when:

• "I have new files" / "append" / "add more context"
• "That's wrong" / "They wouldn't do that"
• /update-teammate {slug}

List teammates: /list-teammates

Quick Start Mode

If the user provides everything in one message (e.g. "Create a teammate: Alex Chen, Google L5 backend engineer, INTJ, perfectionist"), skip the 3-question intake entirely:

1. Parse name, role, personality from the message
2. Show confirmation summary
3. Jump directly to Step 2 (Source Material Import)

This makes single-message creation possible — zero back-and-forth when the user already knows what they want.

Platform Detection & Tool Mapping

Detect the runtime environment and use the correct tools:

Path Resolution

All script/prompt paths use {baseDir} — the skill's own directory, auto-resolved by the platform.

• Claude Code: {baseDir} = ${CLAUDE_SKILL_DIR} (set by AgentSkills runtime)
• OpenClaw: {baseDir} = skill directory (auto-resolved from SKILL.md location)
• Other agents: resolve relative to the SKILL.md parent directory

Output Directory

Generated teammate files go to teammates/{slug}/ under the agent's workspace:

To install the generated skill globally, copy teammates/{slug}/SKILL.md to the platform's skill directory.

Tool Reference

Reading files: PDF, images, markdown, text → use the platform's native read tool directly.

Main Flow: Create a New Teammate Skill

Step 1: Basic Info Collection (3 questions — or fewer)

Read {baseDir}/prompts/intake.md for the full question sequence. Only ask 3 questions:

1. Name / Alias (required) — e.g. alex-chen or Big Mike
2. Role info (optional, one sentence) — e.g. Google L5 backend engineer
3. Personality profile (optional, one sentence) — e.g. INTJ, perfectionist, Google-style, brutal CR feedback

Everything except name can be skipped. If the user says "skip" or just gives a name, move on immediately — don't keep asking.

After collecting, show a compact confirmation:

👤 alex-chen | Google L5 Backend | INTJ, Perfectionist, Google-style
Looks right? (y / change something)

One line, not a multi-line summary. Get confirmation fast.

Step 2: Source Material Import

Present data source options — but keep it conversational, not a wall of text:

Now, do you have any of their work artifacts? (all optional)

  • Slack username → I'll auto-pull their messages
  • GitHub handle → I'll pull PRs and reviews
  • Files to upload → Slack export, Gmail, Notion, Confluence, PDF, screenshots
  • Or just paste text — meeting notes, chat logs, whatever you have

You can also skip this entirely — I'll work with what you gave me above.

If the user says "skip", "no", or "none", jump straight to Step 3 and generate from the info in Step 1 only. Don't ask again.

Option A: Slack Auto-Collect

First-time setup:

python3 {baseDir}/tools/slack_collector.py --setup

Collect data:

python3 {baseDir}/tools/slack_collector.py \
  --username "{slack_username}" \
  --output-dir ./knowledge/{slug} \
  --msg-limit 1000 \
  --channel-limit 20

Then read the output files: knowledge/{slug}/messages.txt, threads.txt, collection_summary.json.

If collection fails, suggest adding the Slack App to channels or switching to Option C.

Option B: GitHub Auto-Collect

python3 {baseDir}/tools/github_collector.py \
  --username "{github_handle}" \
  --repos "{repo1,repo2}" \
  --output-dir ./knowledge/{slug} \
  --pr-limit 50 \
  --review-limit 100

Then read: knowledge/{slug}/prs.txt, reviews.txt, issues.txt.

Option C: Upload Files

Use the tool reference table above. For each file type, run the appropriate parser. PDF/images/markdown → read directly with platform read tool.

Option D: Paste Text

Use pasted content directly as source material. No tools needed.

Option E: Provide Links

• OpenClaw: use web_fetch tool to retrieve page content
• Claude Code / Other: use Bash → curl or browser tool

If user says "skip", generate from Step 1 info only.

Step 3: Analyze Source Material

Run dual-track analysis on all collected materials:

Track A (Work Skill): Read {baseDir}/prompts/work_analyzer.md for extraction dimensions. Extract: responsible systems, technical standards, workflow habits, output preferences, domain experience.

Track B (Persona): Read {baseDir}/prompts/persona_analyzer.md for extraction dimensions. Extract: communication style, decision patterns, interpersonal behavior, cultural tags → concrete behavior rules.

Step 4: Generate, Validate, and Preview

Read {baseDir}/prompts/work_builder.md to generate Work Skill content. Read {baseDir}/prompts/persona_builder.md to generate Persona content (5-layer structure).

Quality Gate (mandatory — run before showing preview):

After generating, self-check against these criteria. Fix any failures before showing the preview:

If source material was skipped, lower the bar: Layer 2 examples and catchphrases can be tag-inferred, but must be marked (inferred).

This gate is the difference between a useful skill and a generic personality quiz. Never skip it.

Show a concise preview card (not full content — just the highlights):

━━━ Preview: alex-chen ━━━

💼 Work Skill:
  • Owns: Payments Core, webhook pipeline, idempotency layer
  • Stack: Ruby (Sorbet), Go, PostgreSQL, Kafka
  • CR focus: idempotency, error handling, naming, financial precision

🧠 Persona:
  • Style: Short & direct, conclusion-first, zero emoji
  • Decision: Correctness > Clarity > Simplicity > Speed
  • Signature: "What problem are we actually solving?"

━━━━━━━━━━━━━━━━━━━━━━━

Looks right? Or want to tweak something before I write the files?

Keep to 10–12 lines max. If user says "yes" / "good" / "ok" / "👍", proceed to write immediately.

Step 5: Write Files

After confirmation, create the teammate:

1. Create directories:

mkdir -p teammates/{slug}/versions
mkdir -p teammates/{slug}/knowledge/docs
mkdir -p teammates/{slug}/knowledge/messages
mkdir -p teammates/{slug}/knowledge/emails

2. Write teammates/{slug}/work.md — full work skill content

3. Write teammates/{slug}/persona.md — full persona content (5-layer)

4. Write teammates/{slug}/meta.json:

{
  "name": "{name}",
  "slug": "{slug}",
  "created_at": "{ISO_timestamp}",
  "updated_at": "{ISO_timestamp}",
  "version": "v1",
  "profile": { "company": "", "level": "", "role": "", "mbti": "" },
  "tags": { "personality": [], "culture": [] },
  "impression": "",
  "knowledge_sources": [],
  "corrections_count": 0
}

5. Write teammates/{slug}/SKILL.md (the generated teammate skill):

Size guard: If work.md + persona.md combined exceed 8000 words, split the generated SKILL.md into modular files instead of one monolith:

teammates/{slug}/
├── SKILL.md          # Entry point — loads modules on demand
├── work.md           # Full work skill (standalone)
├── persona.md        # Full persona (standalone)
├── meta.json
└── versions/

The SKILL.md in this case uses a lazy-load pattern:

---
name: teammate-{slug}
description: "{name} — {identity}. Full persona + work skill."
user-invocable: true
---

# {name}

{identity}

## Loading

This teammate has extensive documentation. Load on demand:
- For work questions: read `work.md` in this directory
- For persona/style questions: read `persona.md` in this directory
- For full context: read both

## Quick Reference

{10-line summary: top 5 work skills + top 5 persona traits}

## Execution Rules

1. Read persona.md first for attitude and communication style
2. Read work.md for domain knowledge and technical standards
3. Always maintain persona.md Layer 2 communication style
4. Layer 0 rules have highest priority — never violate
5. Correction Log entries override earlier rules
6. Never break character into generic AI
7. Keep response length realistic for this person

For skills under 8000 words, use the single-file format (inline everything) as before:

---
name: teammate-{slug}
description: "{name} — {company} {level} {role}. Invoke to get responses in their voice and style."
user-invocable: true
---

# {name}

{company} {level} {role}

---

## PART A: Work Capabilities

{full work.md content}

---

## PART B: Persona

{full persona.md content}

---

## Execution Rules

1. PART B decides first: what attitude to take on this task?
2. PART A executes: use technical skills to complete the task
3. Always maintain PART B's communication style in output
4. PART B Layer 0 rules have highest priority — never violate

5b. Auto-install the generated skill:

After writing the files, automatically copy the generated SKILL.md to the platform's skill directory so the user can invoke /{slug} immediately without manual setup:

# OpenClaw
mkdir -p ~/.openclaw/workspace/skills/teammate-{slug}
cp teammates/{slug}/SKILL.md ~/.openclaw/workspace/skills/teammate-{slug}/SKILL.md

# Claude Code (global)
mkdir -p ~/.claude/skills/teammate-{slug}
cp teammates/{slug}/SKILL.md ~/.claude/skills/teammate-{slug}/SKILL.md

Detect platform and run the appropriate command. If auto-install fails, show manual instructions instead.

6. Confirm to user with a live test:

✅ alex-chen created!

📁 Location: teammates/alex-chen/
🗣️ Commands: /alex-chen (full) | /alex-chen-work | /alex-chen-persona

Let me give you a quick demo — ask alex-chen anything:

6b. Run Smoke Test (mandatory):

Read {baseDir}/prompts/smoke_test.md for the full test protocol.

Internally run 3 test prompts against the generated skill:

1. Domain question (tests work skill accuracy)
2. Pushback scenario (tests persona Layer 0 + Layer 3)
3. Out-of-scope question (tests character boundary)

Show a compact scorecard to the user:

🧪 Smoke Test: ✅ Domain ✅ Pushback ✅ Out-of-scope — 3/3 passed

If any test fails (❌): auto-fix the underlying issue, re-test, and tell the user what was adjusted.

6c. Privacy Scan (before sharing/exporting):

If the user intends to share or export the teammate, run:

python3 {baseDir}/tools/privacy_guard.py --scan teammates/{slug}/

If PII is found, warn the user and offer to auto-redact:

python3 {baseDir}/tools/privacy_guard.py --scan teammates/{slug}/ --redact

Knowledge directories (knowledge/{slug}/) contain raw personal data and should never be shared. The .gitignore already excludes knowledge/ and teammates/*/ from version control.

Then immediately switch into the generated skill's persona and respond to whatever the user says next as the teammate. This makes the skill feel real from second one — no "go try it yourself" dead end.

If the user doesn't ask anything, prompt with a sample:

Try it: "Alex, should we use MongoDB for this new service?"

Evolution Mode: Append Files

When user provides new materials:

1. Parse new content using Step 2 methods
2. Read existing teammates/{slug}/work.md and persona.md
3. Read {baseDir}/prompts/merger.md for incremental analysis rules
4. Backup current version:

   python3 {baseDir}/tools/version_manager.py --action backup --slug {slug} --base-dir ./teammates

1. Edit files with incremental updates
2. Regenerate teammates/{slug}/SKILL.md
3. Update meta.json version and timestamp

Evolution Mode: Conversation Correction

When user says "that's wrong" / "they wouldn't do that":

1. Read {baseDir}/prompts/correction_handler.md
2. Determine if correction applies to Work or Persona
3. Generate correction record
4. Append to ## Correction Log section
5. Regenerate teammates/{slug}/SKILL.md

Management Commands

Error Recovery

Tool/script fails: Don't dump the traceback to the user. Summarize in one line + suggest a fix:

⚠️ Slack collector failed (token expired). Run: python3 tools/slack_collector.py --setup

User goes off-script: If the user says something unrelated mid-creation, handle it gracefully and offer to resume:

No problem — want to continue creating {slug}, or do something else?

Partial creation interrupted: If a previous creation was abandoned, detect existing teammates/{slug}/ with incomplete files (missing SKILL.md) and offer to resume or restart:

Found an incomplete teammate "alex-chen" from earlier. Resume where we left off, or start fresh?