onboard-agent（船上 Agent）

MANDATORY PREPARATION

Invoke /agent-workflow — it contains workflow principles, anti-patterns, and the Context Gathering Protocol. Follow the protocol before proceeding — if no workflow context exists yet, you MUST run /teach-maestro first.

Bootstrap a new agent workflow from scratch, or add a new agent to an existing system.

Step 1: Establish Conventions

## Workflow Conventions
### Prompt Format
- Delimiter style: [XML tags / markdown headers / triple-dash]
- Section order: [System → Context → Instructions → Input]
- Output format: [JSON with schema / markdown template]

### Tool Conventions
- Naming: [verb_noun / noun.verb / camelCase]
- Description template: [What → When → When Not → Returns]
- Error format: [{ code, message, details }]

### Logging
- Format: [JSON structured]
- Required fields: [workflow_id, step, timestamp, level]

### File Structure
- Prompts: [prompts/workflow-name/v1.md]
- Tools: [tools/tool-name.{ext}]
- Config: [config/environment.yaml]
- Tests: [tests/workflow-name/]

Step 2: Create Initial Structure

project/
├── prompts/          # System prompts, versioned
├── tools/            # Tool definitions
├── config/           # Environment-specific configuration
├── tests/            # Golden test sets and evaluation suites
├── logs/             # Runtime logs (gitignored)
└── .maestro.md       # Workflow context

Step 3: Create the First Agent

1. System prompt: Role definition with constraints
2. 2-3 essential tools: Start with the minimum viable tool set
3. Output schema: Define expected output format
4. One golden test: At least one test case with known-good output
5. Basic error handling: Structured error responses
6. Logging: Structured log output for each run

Step 4: Verify

• Run the agent with the golden test case
• Verify error handling works (send bad input)
• Verify logging captures useful context

Recommended Next Step

After onboarding, run /diagnose for a baseline health check, then /fortify to add production-grade error handling.

NEVER:

• Start building without establishing conventions
• Create tools without descriptions
• Skip the golden test case
• Over-scope the initial agent (start minimal, amplify later)