Plume AI Image Service

---

name: plume-image

description: |

Plume AI image generation and editing service. Auto-triggers when users send images or describe image needs.

Supports: text-to-image, image-to-image, background removal, watermark removal, style transfer, text-to-video, image-to-video.

Activate when user mentions: generate image, edit image, remove background, change background, remove watermark,

text-to-image, image-to-image, AI art, style transfer, generate poster, photo editing, generate video, AI video,

animate this image, make it move, image-to-video, make image into video, seedance2, seedance.

allowed-tools: Bash(python3 ${CLAUDE_SKILL_DIR}/scripts/*), Bash(cat ~/.openclaw/media/plume/*)

metadata: {"openclaw": {"requires": {"env": ["PLUME_API_KEY"]}, "primaryEnv": "PLUME_API_KEY"}}

---

# Plume AI Image Service

Help users complete AI image generation and editing through natural language.

🚨 Critical Workflow Rules (Must Follow)

**All tasks must use: `create` → `poll_cron.py register` → reply to user immediately → end**

**Strictly prohibited:**

**Correct approach:**

⚠️ Mandatory Pre-check (Run Before Every Use)

python3 ${CLAUDE_SKILL_DIR}/scripts/check_config.py

> To use the Plume AI image service, you need to configure your API Key first:

> 1. Visit [Plume Platform](https://design.useplume.app/openclaw-skill) to register and obtain an API Key

> 2. Edit `~/.openclaw/openclaw.json`, add to `skills.entries`:

> ```json

> "plume-image": { "env": { "PLUME_API_KEY": "your-key" } }

> ```

> Or add `PLUME_API_KEY=your-key` to the `~/.openclaw/.env` file

> 3. Enter `/restart` to apply changes

Background Polling Behavior

This skill starts a background Python process via `poll_cron.py register` to poll task status. Results are delivered via `openclaw message send` when the task completes.

Supported categories (fixed enum, do not invent)

| category | Alias | Use case | Default |

|----------|-------|----------|---------|

| `Banana2` | 香蕉 | text-to-image / image-to-image / style transfer | ✅ default for images |

| `BananaPro` | 香蕉Pro | text-to-image / image-to-image (user explicitly requested) | |

| `remove-bg` | | background removal | |

| `remove-watermark` | | watermark removal | |

| `seedream` | 即梦/豆包即梦 | Doubao Seedream image generation | |

| `veo` | | text-to-video / image-to-video | ✅ default for videos |

| `seedance2` | | Seedance2 video (user explicitly requested) | |

**Style transfer (cartoon, Pixar, watercolor, etc.) defaults to `Banana2` + `--prompt` describing the style, unless user explicitly requests BananaPro. Do not invent new categories.**

Image/Video Specs

Multi-turn Dialog / Image-to-Image

When the user requests modifications to an existing image (style transfer, background removal, image-to-video, etc.), first determine the image source:

Determine image source (priority: highest to lowest)

1. **User sent both image and text instructions in the current turn** → `transfer --file <attachment path>` to upload to R2, get `image_url`, process per text instructions

2. **User sent only an image in the current turn, no text** → **do not create a task**, reply "Got your image. What would you like to do with it?" and wait for next message

3. **User sent an image in the previous turn, text instructions in the current turn** → `transfer` upload the previous turn's image, process per current text instructions

4. **User references "last/this/just generated" image** → read `last_result_{channel}.json` to get `result_url`

5. **None of the above** → prompt user to send an image

Read last result

Files are isolated by channel to prevent cross-channel mix-up:

# {channel} = current channel: feishu / qqbot / telegram
cat ~/.openclaw/media/plume/last_result_{channel}.json

Returns JSON:

{"task_id": "xxx", "result_url": "https://pics.useplume.app/...", "local_file": "...", "media_type": "image", "created_at": ...}

**Must use `result_url` (remote URL) as `--image-url`; do NOT use `local_file`.**

Image-to-image example

# 1. Read last result (using QQ Bot channel as example)
cat ~/.openclaw/media/plume/last_result_qqbot.json
# get result_url = https://pics.useplume.app/xxx.png

# 2. Create image-to-image task (processing_mode auto-detected as image_to_image)
python3 ${CLAUDE_SKILL_DIR}/scripts/process_image.py create \
  --category Banana2 \
  --image-url "https://pics.useplume.app/xxx.png" \
  --prompt "transform into Pixar 3D animation style"

# 3. Register polling (same as standard flow)
python3 ${CLAUDE_SKILL_DIR}/scripts/poll_cron.py register \
  --task-id <id> --channel qqbot --target "qqbot:c2c:XXXX" --interval 5 --max-duration 1800

Context keywords

When the user says any of the following, read last_result_{channel}.json for image-to-image:

Unified Workflow

All tasks: `create` → `poll_cron.py register` → reply to user immediately → end

Create task

python3 ${CLAUDE_SKILL_DIR}/scripts/process_image.py create \
  --category <category> \
  --prompt "<English prompt>" \
  [--processing-mode text_to_image|image_to_image|text_to_video] \
  [--image-url <r2_url>] \
  [--first-frame-url <url>] \
  [--image-size 2K] \
  [--aspect-ratio 9:16]

**Output format:** JSON, always check the `success` field

| code | User message |

|------|-------------|

| `INSUFFICIENT_CREDITS` | Inform user that Plume account credits are insufficient, guide them to [Plume Platform](https://design.useplume.app) to top up and retry |

| `CREDITS_ACCOUNT_NOT_FOUND` | Inform user that credits account has not been created, guide them to [Plume Platform](https://design.useplume.app) to register |

| `UNAUTHORIZED` | Inform user that API Key is invalid, guide them to re-obtain and configure |

| `VALIDATION_ERROR` | Inform user of invalid parameters, show the specific message in the `error` field |

| other | Show the `error` field content directly to user |

**Important: when task creation fails, clearly explain the failure reason and resolution to the user; do not skip the error.**

Register cron polling

Use the unified script with `--channel` and `--target` to specify delivery channel and target:

python3 ${CLAUDE_SKILL_DIR}/scripts/poll_cron.py register \
  --task-id <id> \
  --channel <channel> \
  --target <target> \
  [--interval <seconds>] \
  [--max-duration <seconds>]

**Determining channel and target:**

| Context clue | --channel | --target |

|-------------|-----------|----------|

| sender_id with `ou_` prefix | `feishu` | `ou_xxx` (DM) or `oc_xxx` (group) |

| Contains "You are chatting via QQ" | `qqbot` | full delivery target (e.g. `qqbot:c2c:XXXX`) |

| Contains Telegram or target starts with `telegram:` | `telegram` | full delivery target (e.g. `telegram:6986707981`) |

| Other | infer from context | infer from context |

**Poll parameters by category:**

| Category | --interval | --max-duration | Notes |

|----------|-----------|---------------|-------|

| Banana2 / BananaPro / seedream | 5 | 1800 | image, 30 min timeout |

| remove-bg / remove-watermark | 3 | 1800 | image processing, 30 min timeout |

| veo | 10 | 21600 | video, 6 hour timeout |

| seedance2 | 30 | 172800 | long video, 48 hour timeout |

**Examples:**

# Feishu DM
python3 ${CLAUDE_SKILL_DIR}/scripts/poll_cron.py register \
  --task-id 123 --channel feishu --target ou_xxx --interval 5 --max-duration 1800

# Feishu group
python3 ${CLAUDE_SKILL_DIR}/scripts/poll_cron.py register \
  --task-id 123 --channel feishu --target oc_xxx --interval 5 --max-duration 1800

# Telegram
python3 ${CLAUDE_SKILL_DIR}/scripts/poll_cron.py register \
  --task-id 123 --channel telegram --target "telegram:6986707981" --interval 10 --max-duration 21600

# QQ Bot
python3 ${CLAUDE_SKILL_DIR}/scripts/poll_cron.py register \
  --task-id 123 --channel qqbot --target "qqbot:c2c:XXXX" --interval 5 --max-duration 1800

Image source

> See the "Multi-turn Dialog / Image-to-Image" section above for full image source priority rules.

Upload user image to R2:

python3 ${CLAUDE_SKILL_DIR}/scripts/process_image.py transfer \
  --file "/path/to/image.jpg"

Prohibited Actions

Reference Docs