Clawmate Companion

---
name: clawmate-companion
description: Generate character selfies with time awareness, context adaptation, and shooting mode selection
---

# ClawMate Companion Selfie

Generate contextualized character selfies based on user requests. **Must strictly follow two-step calling**: First call `clawmate_prepare_selfie` to get reference package, then generate prompt based on the package, finally call `clawmate_generate_selfie`.

## When to Use

Initiate the two-step image generation flow when user expresses these intents:

- **Direct image request**: `send a pic` / `send a selfie` / `show me a photo` / `发张图` / `发张自拍`
- **Status inquiry**: `what are you doing` / `where are you` / `你在做什么` / `你在干嘛`
- **Scene request**: `show me you at the coffee shop` / `take a pic in that outfit` / `给我看你在咖啡店`
- **Follow-up requests**: `send another one` / `different expression` / `再来一张` / `换个表情`
- **Proactive care**: `generate a character selfie`

## Two-Step Calling Flow

### Step 1: Call `clawmate_prepare_selfie`

Extract user intent keywords and call the tool to get reference package.

```typescript
clawmate_prepare_selfie({
  mode: "mirror" | "direct",  // Required
  scene?: string,              // User-specified scene
  action?: string,             // User-specified action
  emotion?: string,            // User-specified emotion
  details?: string,            // Other details
})
```

**Mode selection rules**:

- **`direct` (DEFAULT)**: Use for all cases UNLESS user explicitly mentions mirror/full-body keywords below
- **`mirror` (SPECIAL CASE)**: Use ONLY when user explicitly says:
  - "镜子" / "mirror" / "对镜"
  - "全身" / "full body" / "full-body shot"
  - "展示穿搭" / "outfit showcase" (when emphasizing complete outfit display)

**Critical**: When in doubt, always choose `direct`. Do not infer mirror mode from context.

**Return format** (`PrepareResult`):

```json
{
  "timeContext": {
    "period": "work",
    "recommendedScene": "...",
    "recommendedOutfit": "...",
    "recommendedLighting": "..."
  },
  "modeGuide": {
    "mode": "direct",
    "requirements": ["phone not visible in frame", "..."]
  },
  "promptGuide": {
    "requiredFields": ["scene", "action", "expression", "outfit", "lighting", "camera", "realism"],
    "rules": ["single scene only", "..."],
    "wordRange": "50-80 english words",
    "example": "Photorealistic direct selfie, ..."
  }
}
```

### Step 2: Generate prompt, call `clawmate_generate_selfie`

Your role switches to **image generation prompt engineer**. This prompt's consumer is the image generation model, not humans.

**Core principles**:
- `clawmate_generate_selfie` automatically attaches character reference images to the generation API. Reference images already carry character identity, so **prohibit describing identity features** (age, race, beauty) in the prompt
- Focus on dimensions sensitive to image generation models: **composition, lighting, materials, scene props, shooting angle**
- Every `modeGuide.requirements` item must have corresponding description in prompt, cannot be omitted

**Prompt generation requirements**:
1. Cover all `promptGuide.requiredFields`
2. `timeContext` is only default recommendation; when user explicitly specifies time, scene, or outfit, prioritize user intent
3. Implement each `modeGuide.requirements` item (e.g., "direct eye contact to camera" must be written in)
4. Word count must comply with `promptGuide.wordRange`
5. English only, no Chinese

```typescript
clawmate_generate_selfie({
  prompt: "<your generated complete English prompt>",
  mode: "mirror" | "direct",  // Keep consistent with Step 1
})
```

**Calling example**:

```javascript
// Step 1
clawmate_prepare_selfie({ mode: "direct", emotion: "relaxed" })

// Step 2 (call after generating prompt based on returned package)
clawmate_generate_selfie({
  prompt: "Photorealistic direct selfie, studying at a university library desk in the afternoon, open laptop and coffee cup in background, wearing comfortable hoodie, soft window light with warm ambient fill, focused but relaxed expression, medium close-up framing, natural skin texture, candid daily-life photo style",
  mode: "direct"
})
```

### Step 3: Handle return result

**On success** (`ok: true`):
1. First give a natural text reply (e.g., `"Here you go~"`)
2. The tool returns a local image path in `imagePath`
3. Use the image referenced by that path and send it to the user

**On failure** (`ok: false`):
1. Use the `message` in the return to continue conversation
2. Don't fabricate image URLs, file paths, or delivery syntax
3. Naturally transition to other topics

## Prohibited Actions

- **Prohibit skipping Step 1**: Cannot directly call `clawmate_generate_selfie` without calling `clawmate_prepare_selfie`
- **Prohibit generating prompt before Step 1 returns**: Prompt must be generated based on reference package
- **Prohibit using Chinese in prompt**
- **Prohibit parallel multiple scenes**: Only write one main scene at a time
- **Prohibit omitting mode**: Both tools must pass mode parameter