onboard

name	onboard
description	Interactive agent factory onboarding. Use when the user runs /onboard or wants to create multi-agent applications, set up new specialized agents for a domain, or generate sub-agent configurations and tool stubs.
disable-model-invocation	true
allowed-tools	Read, Write, Edit, Glob, Grep, Bash, Task, AskUserQuestion
argument-hint	Optional domain name (e.g., "HR Assistant")

Agent Factory Onboarding Skill

When the user runs /onboard, follow this process to create a new agent application.

Overview

This skill helps create a multi-agent application by:

1. Collecting configuration (domain + sub-agents)
2. Generating sub-agent files using templates in ./templates/
3. Modifying 6 orchestration agent prompts in parallel using Task subagents

See ./examples/hr_assistant.md for a complete example.

---

Phase 1: Collect Configuration

Step 1: Domain Configuration

Ask: "What domain does this assistant serve? Please provide:

1. Domain name (e.g., 'HR Assistant', 'IT Support')
2. Description (one sentence explaining what the assistant helps with)"

Step 2: Sub-Agents

For each sub-agent, collect:

• key: snake_case identifier (e.g., leave, payroll)
• name: display name (e.g., 'leave-agent')
• description: one-line description
• capabilities: list of things this agent can do
• tools: list of tool functions with name and description
• use_cases: 2-3 use cases with example queries
• when_to_use: scenarios where triage should route here
• when_not_to_use: scenarios where another agent is better

Ask: "Now let's define your specialized agents. For each agent, provide key, name, description, capabilities, tools, use cases, when to route here, and when NOT to route here."

Continue: "Do you have another agent to add? (yes/no)"

Step 3: Confirm Configuration

Present a summary and ask: "Does this look correct? (yes/no)"

---

Phase 2: Generate Sub-Agent Files

After confirmation, generate using templates from ./templates/:

1. app/agent_factory/subagent_config.py - Use template: ./templates/subagent_config.py.md
2. app/agent_factory/agents/sub_agents/{key}_agent.py - Use template: ./templates/sub_agent.py.md
3. app/agent_factory/agents/sub_agents/tools/{key}_tools.py - Use template: ./templates/tools.py.md

---

Phase 3: Modify Orchestration Agent Prompts (PARALLEL)

After generating sub-agent files, modify the 6 orchestration agent prompts in parallel using the Task tool.

Critical: Use Parallel Task Subagents

Launch 6 Task tool calls in a single message. Each subagent modifies one orchestration agent file independently.

Orchestration Agents and Placeholders

Agent	File	Key Placeholders
Triage	triage_agent.py	{domain_purpose}, {agent_summaries}, {routing_examples}, {decision_guidelines}
Plan	plan_agent.py	{domain_purpose}, {agent_tools_summary}, {planning_examples}, {parallel_vs_sequential_guidance}
Replan	replan_agent.py	{agent_tools_summary}, {retry_examples}, {when_to_retry_vs_complete}
Review	review_agent.py	{completeness_criteria}, {domain_specific_quality_checks}, {review_examples}
Clarify	clarify_agent.py	{domain_purpose}, {capabilities_summary}, {clarification_examples}
Summary	summary_agent.py	{domain_purpose}, {formatting_guidelines}, {response_examples}

How to Execute

Use a single message with 6 Task tool calls. Example:

Task 1 (parallel):
  description: "Modify triage agent prompt"
  subagent_type: "general-purpose"
  prompt: |
    Modify the triage agent for {domain_name}.

    Configuration:
    - Domain: {domain_name} - {domain_description}
    - Sub-agents: {sub_agent_summaries with tools and routing info}

    File: app/agent_factory/agents/orchestration/triage_agent.py

    Instructions:
    1. Read the file
    2. Fill {domain_purpose} with domain description
    3. Fill {agent_summaries} with detailed agent descriptions including tools, use cases, routing guidance
    4. Fill {routing_examples} with 5-8 specific examples
    5. Fill {decision_guidelines} with decision-making rules
    6. Write the updated file

    DO NOT modify output format. Write in English.

Task 2 (parallel):
  description: "Modify plan agent prompt"
  subagent_type: "general-purpose"
  prompt: |
    Modify the plan agent for {domain_name}.
    ... (similar structure)

... Tasks 3-6 for replan, review, clarify, summary

Subagent Instructions Template

Each subagent should:

1. Read the orchestration agent file
2. Fill {placeholders} using LLM reasoning based on collected config
3. Modify surrounding text if it improves the prompt
4. DO NOT modify output format sections (captured by schema)
5. Write the updated file
6. ALWAYS write in English regardless of user's language

---

Phase 4: Post-Generation Summary

After all tasks complete, summarize:

## Files Generated/Modified

### Generated:
1. `app/agent_factory/subagent_config.py`
2. `app/agent_factory/agents/sub_agents/{key}_agent.py` (per agent)
3. `app/agent_factory/agents/sub_agents/tools/{key}_tools.py` (per agent)

### Modified (prompts customized for your domain):
4-9. All 6 orchestration agents in `app/agent_factory/agents/orchestration/`

## Next Steps

1. **Implement Tools**: Replace stubs in `tools/{key}_tools.py`
2. **Test**:
   ```bash
   USE_DEMO_OPSAGENT=false uv run uvicorn app.main:app --reload

Would you like help implementing any tools?

---

## Phase 5: Tool Implementation (Optional)

If the user wants help implementing tools:

1. **Gather requirements** (APIs, data sources, SDKs)
2. **Add dependencies**: `uv add <package-name>`
3. **Implement tools**: Replace `NotImplementedError` stubs with actual logic
4. **Test each tool**:
   ```python
   # test_tools.py
   from app.agent_factory.agents.sub_agents.tools.{key}_tools import tool_func
   result = tool_func(param="value")
   print(result)

Run: uv run python test_tools.py 5. Iterate until all tools pass

---

Important Notes

• Use snake_case for keys and function names
• Use PascalCase for class names
• Tool functions must return JSON strings
• Use Annotated type hints for parameters
• Orchestration prompts are filled by Claude (LLM), not string parsing
• DO NOT modify output format sections (captured by schema)
• ALWAYS write prompts in English

---

Example Session

User: /onboard

Claude: What domain does this assistant serve?

User: HR Assistant - Helps employees manage leave, payroll, benefits

Claude: Now define your first agent...

User: leave agent - handles vacation/leave, tools: get_leave_balance, submit_leave_request
       Route here for: PTO, vacation, sick leave
       Don't route here for: pay during leave (that's payroll)

Claude: Do you have another agent?

User: yes, payroll agent - salary/compensation, tools: get_pay_stub, get_bonus_info
       Route here for: salary, pay, bonus
       Don't route here for: leave balance

User: no more agents

Claude: [Shows configuration summary]

User: yes, looks correct

Claude: [Generates sub-agent files]
Claude: [Launches 6 parallel Task subagents to modify orchestration prompts]
Claude: [Shows completion summary with next steps]

See ./examples/hr_assistant.md for the complete generated output.