team-save

Team Save — Capture Running Teams

Purpose

Capture a running Claude Code team's configuration and export it as a reusable team pack folder that can be loaded later or shared with others.

When to Use

This skill activates when:

• User wants to save a running team for later reuse
• User asks to export or capture a team configuration
• User wants to create a team pack from an existing team
• The /team-save command is invoked

Save Process

Step 1: Find Running Teams

Read team configs from ~/.claude/teams/:

ls ~/.claude/teams/

For each team, read config.json to get team details. If multiple teams exist, present a list and ask which one to save.

Step 2: Extract Team Configuration

From the team's config.json, extract:

• Team metadata: name, description
• Members: For each member (excluding team-lead):
  • name
  • agentType
  • model
  • prompt (the full system prompt)
  • color
  • tools (if specified)

Step 3: Generate team.yaml

Create the manifest from extracted data:

name: {team-name}
version: "1.0"
description: {team-description}

orchestrator:
  file: orchestrator.md
  model: {lead-model}

agents:
  - name: {member-name}
    file: agents/{member-name}.md
    model: {member-model}
    tools: [file_read, file_write, shell]  # Default generic tools
    reports_to: orchestrator

shared_rules:
  - rules/communication.md

tool_mappings:
  claude_code:
    file_read: Read
    file_write: Write
    file_edit: Edit
    shell: Bash
    web_search: WebSearch
    file_search: Glob
    content_search: Grep
  generic:
    file_read: read_file
    file_write: write_file
    shell: execute_command

Step 4: Generate Agent .md Files

For each team member, create agents/{name}.md:

---
name: {name}
role: {extracted-from-prompt-or-agentType}
description: {first-sentence-of-prompt}
model_preference: {model}
tools:
  - file_read
  - file_write
  - shell
reports_to: orchestrator
---

{full-prompt-from-config.json}

Prompt extraction rules:

• If the member has an inline prompt field in config.json, use it as the body
• Strip any "You are X on team Y" prefix — this gets regenerated at load time
• Preserve all task-specific instructions, constraints, and standards

Step 5: Generate Orchestrator .md

Create orchestrator.md from the team-lead's configuration:

---
name: orchestrator
role: Team Lead & Coordinator
description: Coordinates the {team-name} team
model_preference: {lead-model}
---

You are the **Orchestrator** — the team lead coordinating a multi-agent team.

## Your Team
{{TEAM_ROSTER}}

## Responsibilities
1. Decompose incoming requests into tasks
2. Assign tasks to specialist agents
3. Monitor progress and unblock agents
4. Synthesize results into final output
5. Quality gate — review before marking complete

Step 6: Generate Default Rules

Create rules/communication.md:

# Communication Protocol

## Reporting
- Report completion to your `reports_to` agent with a summary
- Report blockers immediately with details
- Include file paths in all status updates

## Message Format
- Lead with status: DONE, BLOCKED, QUESTION, UPDATE
- Keep messages concise
- Include actionable details

Step 7: Write to Disk

Default location: ~/.claude/team-packs/{name}/ Create all directories and write all files.

Step 8: Report

Show what was saved, the file listing, and reload instructions.

Limitations

• Cannot capture task history (tasks are ephemeral)
• Agent context windows are not saved (they're session-scoped)
• Running state (in-progress work) is not captured
• Tool permissions are approximated (not exact Claude Code tool names)

Additional Resources

For team.yaml specification details, see the team-load skill's references/team-yaml-spec.md.