AI Tasks Skill

When to use

• User wants to tag images with business-specific categories
• User needs controlled vocabularies (predefined value lists)
• User wants yes/no quality checks on images
• User needs to extract metadata using natural language analysis

AI Task Structure

Each AI Task contains 1-10 sub-tasks with:

1. Instruction (required): Natural language question about the image
2. Action Type (required): selecttags, selectmetadata, or yes_no
3. Vocabulary (optional): 1-30 predefined approved values

Action Types

select_tags

Adds tags from vocabulary. Supports multiple selections.

{
  "type": "select_tags",
  "instruction": "What body style is this vehicle?",
  "vocabulary": ["sedan", "suv", "hatchback"],
  "max_selections": 1
}

select_metadata

Sets custom metadata field (field must exist in DAM).

{
  "type": "select_metadata",
  "instruction": "What is the primary color?",
  "field": "primary_color",
  "vocabulary": ["red", "blue", "white", "black"],
  "max_selections": 1
}

yes_no

Binary quality check with conditional actions.

{
  "type": "yes_no",
  "instruction": "Is the product completely visible?",
  "on_yes": { "add_tags": ["framing_ok"] },
  "on_no": { "add_tags": ["needs_reshot"] }
}

Gotchas

• Instruction clarity: Be specific ("What is the collar type?" not "Describe the image")
• Vocabulary design: Use business terminology, non-overlapping, 1-30 items max
• Field requirements: For select_metadata, field must exist in DAM first
• Tag values: Cannot contain % character
• yesno tasks: Must have at least one of onyes or on_no
• Vocabulary length: Max 500 characters combined (select_tags only)
• Scope: Works on visual content only (images/videos)

Applying AI Tasks

• Via Saved Extensions: Create and apply via dashboard/API
• Via API at Upload: Include in extensions array
• Via Path Policies: Auto-apply to files in specific folders

Full examples

For complete, copy-ready ai-tasks configurations organized by industry (fashion e-commerce, travel, automotive) and detailed per-task-type parameter references, read [resources/EXAMPLES.md](resources/EXAMPLES.md).