# reverse_prompt (https://sellersheetai.com/docs/mcp-api/tools/reverse_prompt)

> Describe reference image(s) as a structured English prompt — step 1 of a restyle: adapt the fields to your product (subject, colors and materials stay faithful; swap scene, lightin

Access: **WRITE** · toolset sellersheet · family null

g and camera), stringify it and pass it as edit_image's prompt with your product photos as image_sources. Billed per call on the text meter, refunded on failure; synchronous, ~5–30 s. Then write the adapted prompt yourself: edit data.json_prompt so subject, details_appearance, colors and materials are YOUR product (exact geometry, proportions and label text kept; nothing of the reference product leaks in); set intended_use to the target slot (an Amazon main image is pure white, product only, no text); borrow only the look — environment_background, lighting_atmosphere, style_camera, composition, aspect_ratio; put the invariants in `constraints` (no watermark, no extra text, no new objects; 'photorealistic' for a photo; literal copy in straight quotes, brand names spelled letter by letter); one tight phrase per field, the stringified prompt under 2,500 characters. Then json.dumps it and pass it as edit_image's prompt with the product photos as image_sources and quality='high'.

## Call it

MCP tool call:

```json
{
  "name": "reverse_prompt",
  "arguments": {
    "images": []
  }
}
```

Tool API (early access — contact us):

```bash
curl -X POST https://api.sellersheetai.com/api/v1/tools/reverse_prompt \
  -H "Authorization: Bearer $SELLERSHEET_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"images":[]}'
```

## Arguments

| Name | Type | Required | Description |
|---|---|---|---|
| `images` | object[] | yes | 1+ sources — a public URL (a Drive link must be shared 'anyone with the link') or base64; for a restyle, the reference shot. Required. |
| `vision_provider` | string |  | 'auto' (OpenAI→Gemini) \| 'openai' \| 'gemini'. Default 'auto'. |

## Result

- `data.json_prompt` — field dict (subject, action_pose, environment_background, lighting_atmosphere, style_camera, colors, materials, aspect_ratio, …).
- `data.style_tags / data.analysis / data.confidence / data.vision_provider` — supporting detail and which model answered.