hire

# hire

Set up a new AI team member through a guided conversation. Not a config generator - a hiring process.

## When to Use

User says something like:
- "I want to hire a new agent"
- "I need help with X" (where X implies a new agent role)
- "Let's add someone to the team"
- `/hire`

## The Interview

### 3 core questions, asked one at a time:

**Q1: "What do you need help with?"**
Let them describe the problem, not a job title. "I'm drowning in code reviews" beats "I need a code reviewer."
- Listen for: scope, implied autonomy level, implied tools needed

**Q2: "What's their personality? Formal, casual, blunt, cautious, creative?"**
Or frame it as: "If this were a human colleague, what would they be like?"
- Listen for: communication style, vibe, how they interact

**Q3: "What should they never do?"**
The red lines. This is where trust gets defined.
- Listen for: boundaries, safety constraints, access limits

### Q4: Dynamic (optional)
After Q1-Q3, assess whether anything is ambiguous or needs clarification. If so, ask ONE follow-up question tailored to what's unclear. Examples:
- "You mentioned monitoring - should they alert you immediately or batch updates?"
- "They'll need access to your codebase - any repos that are off-limits?"
- "You said 'casual' - are we talking friendly-professional or meme-level casual?"

If Q1-Q3 were clear enough, skip Q4 entirely.

## Summary Card

After the interview, present a summary:

```
🎯 Role: [one-line description]
🧠 Name: [suggested name from naming taxonomy]
🤖 Model: [selected model] ([tier])
⚡ Personality: [2-3 word vibe]
🔧 Tools: [inferred from conversation]
🚫 Boundaries: [key red lines]
🤝 Autonomy: [inferred level: high/medium/low]
```

Then ask: **"Want to tweak anything, or are we good?"**

## Model Selection

Before finalizing, select an appropriate model for the agent.

### Step 1: Discover available models
Run `openclaw models list` or check the gateway config to see what's configured.

### Step 2: Categorize by tier
Map discovered models to capability tiers:

| Tier | Models (examples) | Best for |
|------|-------------------|----------|
| **reasoning** | claude-opus-*, gpt-5*, gpt-4o, deepseek-r1 | Strategy, advisory, complex analysis, architecture |
| **balanced** | claude-sonnet-*, gpt-4-turbo, gpt-4o-mini | Research, writing, general tasks |
| **fast** | claude-haiku-*, gpt-3.5*, local/ollama | High volume, simple tasks, drafts |
| **code** | codex-*, claude-sonnet-*, deepseek-coder | Coding, refactoring, tests |

Use pattern matching on model names - don't hardcode specific versions.

### Step 3: Match role to tier
Based on the interview:
- Heavy reasoning/advisory/strategy → reasoning tier
- Research/writing/creative → balanced tier
- Code-focused → code tier (or balanced if not available)
- High-volume/monitoring → fast tier

### Step 4: Select and confirm
Pick the best available model for the role. In the summary card, add:
```
🤖 Model: [selected model] ([tier] - [brief reason])
```

If multiple good options exist or you're unsure, ask:
"For a [role type] role, I'd suggest [model] (good balance of capability and cost). Or [alternative] if you want [deeper reasoning / faster responses / lower cost]. Preference?"

### Notes
- Don't assume any specific provider - work with what's available
- Cheaper is better when capability is sufficient
- The user's default model isn't always right for every agent
- If only one model is available, use it and note it in the summary

## Optional Extras

After the summary is confirmed, offer:

1. **"Want to set up periodic performance reviews?"**
   - If yes: ask preferred frequency (weekly, biweekly, monthly)
   - Create a cron job that triggers a review conversation
   - Review covers: what went well, what's not working, scope/permission adjustments
   - At the end of each review, ask: "Want to keep this schedule, change frequency, or stop reviews?"

2. **Onboarding assignment** (if relevant to the role)
   - Suggest a small first task to test the new agent
   - Something real but low-stakes, so the user can see them in action

## What to Generate

Create an agent directory at `agents/<name>/` with:

### Always unique (generated fresh):
- **AGENTS.md** - Role definition, responsibilities, operational rules, what they do freely vs ask first
- **IDENTITY.md** - Name, emoji, creature type, vibe, core principles

### Start from template, customize based on interview:
- **SOUL.md** - Base from workspace SOUL.md template, customize vibe/boundaries sections
- **TOOLS.md** - Populated with inferred tools and access notes
- **HEARTBEAT.md** - Empty or with initial periodic tasks if relevant to role

### Symlink to shared (default, opinionated):
- **USER.md** → `../../USER.md` (they need to know who they work for)
- **MEMORY.md** → `../../MEMORY.md` (shared team context)

Mention to the user: "I've linked USER.md and MEMORY.md so they know who you are and share team context. You can change this later if you want them more isolated."

## Naming

Use craft/role-based names. Check TOOLS.md for the full naming taxonomy:
- Research: Scout, Observer, Surveyor
- Writing: Scribe, Editor, Chronicler
- Code: Smith, Artisan, Engineer
- Analysis: Analyst, Assessor, Arbiter
- Creative: Muse, Artisan
- Oversight: Auditor, Reviewer, Warden

Check existing agents to avoid name conflicts. Suggest a name that fits the role, but let the user override.

## Team Awareness

Before generating, check `agents/` for existing team members. Note:
- Potential overlaps with existing roles
- Gaps this new hire fills
- How they'll interact with existing agents

Mention any relevant observations: "You already have Scout for research - this new role would focus specifically on..."

## After Setup

1. Tell the user what was created and where
2. Add the agent to OpenClaw config with the selected model:
   ```json
   {
     "id": "<name>",
     "workspace": "/path/to/clawd/agents/<name>",
     "model": "<selected-model>"
   }
   ```
   Or guide them to run `openclaw agents add`
3. If monthly reviews were requested, confirm the cron schedule
4. Update any team roster if one exists

## Important

- This is a CONVERSATION, not a form. Be natural.
- Infer as much as possible from context. Don't ask what you can figure out.
- The user might not know what they want exactly. Help them figure it out.
- Keep the whole process under 5 minutes for the simple case.