Plume Infographic

Plume AI Infographic Service

Help users generate infographics through natural language. Infographics emphasize "layout design" and "information delivery", distinct from regular illustrations/posters.

Boundary with plume-image: User says "infographic/diagram/visualize/timeline/turn article into graphic" → this skill; says "generate image/poster/remove background/video" → plume-image.

Mandatory Pre-check

Must execute before each use:

python3 ${CLAUDE_SKILL_DIR}/scripts/check_config.py

• CONFIGURED — Configured, proceed with workflow
• NOT_CONFIGURED — Stop, prompt user to visit Plumehttps://design.useplume.app/openclaw-skill to get API Key and configure it in ~/.openclaw/openclaw.json under skills.entries.plume-infographic.env.PLUME_API_KEY

Template Gallery

During content planning, always include the template gallery link for users to browse styles. Append ?lang= parameter based on the user's language:

Agent auto-selects the matching link based on the language the user is currently using in conversation. Example: Browse Template Gallery

When user clicks a template and pastes it back, Agent extracts the template name to search for matching ID, then passes it via --template-id when creating the task.

If user doesn't select a template and confirms directly, Agent auto-matches style based on content (can pass via --style-hint).

Core Workflow

transfer (when reference image exists) → create (sync wait for result) → get local image path → subsequent operations (deliver/package etc.)

Before calling create, you must first send a waiting message to the user via message send, e.g. "Sure, generating your infographic now. This usually takes 1-2 minutes, please wait." This message must be a separate tool call, cannot be in the same turn as create. After message send returns success, call create in the next turn. Because create is synchronous and blocking — if the prompt text and create are in the same turn, the text may not reach the user before blocking starts.

# Upload reference image (reference mode, subcommand is transfer not upload)
python3 ${CLAUDE_SKILL_DIR}/scripts/create_infographic.py transfer --file /path/to/image.png
# Returns {"success": true, "image_url": "https://...", "width": W, "height": H}

# Create task and wait for result (default timeout 30 minutes)
python3 ${CLAUDE_SKILL_DIR}/scripts/create_infographic.py create \
  --channel <channel> --mode <article|reference> [params...]
# Returns {"success": true, "task_id": "xxx", "images": ["/abs/path/result_xxx.png", ...], "result_urls": [...]}

# After getting images, continue with operations:
# Deliver to user
openclaw message send --channel <channel> --target <target> --media /abs/path/result_xxx.png --message "Infographic generation complete"
# Or package
zip -j /tmp/infographic.zip /abs/path/result_xxx.png
openclaw message send --channel <channel> --target <target> --media /tmp/infographic.zip --message "Infographic packaged"

Prohibited: Fabricating task_id/URL, asking for API Key in chat, auto-creating tasks when user only sends image without text, creating tasks directly when user only gives a topic word or one sentence in any language (e.g. "做一张XX的信息图", "帮我做个XX图解", "Make an infographic about XX" — must guide and confirm content and style first, see "Content Planning" section), uploading local files without explicit user confirmation, deleting or modifying any JSON files under ~/.openclaw/media/plume/ (action_log, circuit_breaker, etc.).

Timeout handling: If create returns "status": "timeout", inform user the task is still processing and they can retry later.

Scenario Detection

When receiving user message, check in the following order, stop on first match:

1. Does this turn have a new image?
   ├─ Has image + has action instruction → go to [Reference Image] flow
   │   ├─ Product/physical item/character + selling points text → reference_type=product_embed
   │   ├─ Existing infographic + "use this style as reference" → reference_type=style_transfer
   │   ├─ Existing infographic + "replace content with XX" → reference_type=content_rewrite
   │   └─ Hand-drawn/sketch → reference_type=sketch
   │   Note: Even with detailed text, as long as this turn has an image, must go through reference image flow, cannot use article
   ├─ Has image + no text → Reply "Got the image, how would you like me to process it?"
   └─ No new image → continue ↓

2. Does conversation history have [media attached: /path/...] with no corresponding task record in action_log?
   ├─ Yes → **Ask user for confirmation before uploading** (e.g. "I found an image from earlier: <filename>. Shall I upload it as the reference image?")
   │       After user confirms, transfer --file <path>, determine reference_type based on user's text description, go to [Reference Image] flow
   └─ No → continue ↓

3. Is user referencing/modifying existing results? ("switch style", "try again", "regenerate", "change the look", "换个风格", "再试一次", "重新生成", "换个样式")
   ├─ Yes → Read action_log, go to [Retry] flow
   └─ No → go to [New Creation] flow (article mode)

Typical two-turn scenario: Turn 1: sends image without text → generic agent replies (skill not triggered); Turn 2: says "use this rice cooker for an infographic, selling points: ..." → skill triggers, step 2 asks user to confirm the image before uploading, then goes to product_embed.

Retry detection: When action_log records exist, step 2 is skipped, step 3 matches retry keywords → goes to retry flow.

User Intent → Mode Mapping

Important: When user requests a ratio change, must pass --aspect-ratio (e.g. 16:9 / 4:3 / 1:1 / 3:4 / 9:16), otherwise default 3:4 will be used.

For complete parameter documentation see references/modes.md, scenario examples see references/workflows.md.

Reference Image Source

1. Current turn has image + text → transfer --file to upload, pass returned width/height via --reference-image-width/--reference-image-height
2. Current turn has only image, no text → Reply "Got the image, how would you like me to process it?"
3. Previous turn had image, current turn has text → Extract [media attached: /path/...] path from history, ask user to confirm before uploading (e.g. "I'll use the image you sent earlier as the reference. OK?")
4. User references "the last generated one" → Read action_log_{channel}.json for most recent status=success entry's result_url; if empty, fallback to last_result_{channel}.json
5. None available → Prompt to send an image

result_url is a remote URL, do NOT use local_file.

Retry

1. Read action_log_{channel}.json; if empty, fallback to last_result_{channel}.json; if both empty, prompt to generate one first

2. Select base record:

   • "Try again" / "Regenerate" → Take the last entry (regardless of success/failure) and replay
   • "Switch style" / "Switch content" → Take the most recent entry with status=success, use its task_id as --last-task-id

3. Build command:

   • Batch retry must restore count: If params.count >= 2, must include --count {params.count}
   • "Try again": action=null → repeat_last_task; action!=null → replay with same action, --last-task-id from original record's last_task_id
   • switch_style must include --article (get original content from action_log), otherwise subsequent retries lose context

When retrying, no need to re-upload reference images (backend reads from original task via last_task_id).