Create Teammate

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.