This skill guides you on how to extract visual features and prompts from product main images using multimodal AI, helping e-commerce sellers turn unstructured image data into structured, actionable insights.
Core Concepts
This tool performs deep visual analysis on product main images (and optionally additional images) from a product list. It uses a multimodal AI model to identify specific visual dimensions based on a natural language instruction, such as color, shape, style, material, or specific selling-point elements.
How it works: You provide a list of products (with image URLs) and a natural language prompt describing what to extract. The tool automatically iterates over all products, analyzes each image, and returns structured attribute data (attributeName + attributeValue) appended to each product record.
Row expansion: When extracting multiple dimensions in a single request (e.g., both color and shape), each original product row is duplicated per dimension, resulting in one row per product per attribute.
Parameter Guide
| Parameter | Required | Description |
|---|---|---|
| productImageAnalysisPrompt | Yes | Natural language instruction describing what visual information to extract from the images. Be specific about the dimensions you want (color, material, shape, style, pendant type, etc.). |
| analyzeAdditionalImages | No | Whether to also analyze additional product images beyond the main image. Defaults to false. |
| refResultData | No | Reference data from a previous step, containing the product list to analyze. Must be a JSON string with a products array. |
| userInput | No | Supplementary user input for additional context. |
Writing Effective Prompts
- Be dimension-specific: Clearly state what visual attribute(s) to extract. "Extract the dominant color of each product" is better than "Analyze the images."
- One or few dimensions per call: For cleaner results, focus on one or two dimensions at a time.
- Use concrete terms: "Identify the pendant/charm shape on the product" is clearer than "Look at the decorations."
- No need to specify individual products: The tool automatically iterates over all products in the input list.
- Data flow dependency: The tool requires upstream product data. It cannot reference "products from the previous conversation round" -- the data must be explicitly provided via the current step's input or resource references.
Prompt Examples
| Goal | Example Prompt |
|---|---|
| Extract dominant color | "Analyze each product's main image and extract the primary color of the product" |
| Identify material | "From each product's main image, identify the apparent material (plastic, metal, wood, fabric, etc.)" |
| Classify pendant shape | "Analyze each product's main image and identify the shape of the pendant/charm (round, heart, star, etc.)" |
| Detect style | "Extract the overall style of each product from its main image (minimalist, vintage, bohemian, industrial, etc.)" |
| Reverse-engineer image prompt | "Based on the main image, infer the likely AI-generation prompt or visual description that could reproduce this image" |
| Multi-dimension extraction | "From each main image, extract both the dominant color and the overall product shape" |
How to Call
- API Endpoint：
POST /multimodal/extractPromptsFromMainImage(For complete parameters/responses/error codes, seereferences/api.md） - Python Script：
python scripts/multimodal_extract_attributes.py '<JSON 参数>' [--inline] - Cost Constraints: This tool consumes compute power; by default, the same parameter combination in the same session is called only once, and the script has a 24h local cache. On failure/empty results, do not automatically switch keywords, turn pages, or change zip codes to probe continuously; if you need to continue searching, first explain to the user that additional consumption will be incurred.
Output Strategy (Default Script Behavior)：
- Alwayswrite the complete response to
<cwd>/linkfox/<YYYY-MM-DD>/<session>/data/linkfox-multimodal-extract-attributes-<timestamp>.json（<cwd>is the working directory when the script is executed, which in Claude Code is the current project directory;<session>is taken from the environment variableSESSION_ID, automatically aggregated according to the user's task;writing to /tmp is prohibited, and if the current directory is not writable, report an error) - Response body ≤ 8 KB: after saving to disk, print the complete JSON to stdout
- Response body > 8 KB: after saving to disk, stdout outputs only a summary (top-level fields, common counts such as
total/costToken, the length of the largest list field + the first 3 samples) - Add
--inlineto force full printing to stdout (also saved to disk)
Data Reading Suggestions: First look at the summary to determine whether it is sufficient; when specific fields are needed, prefer to use jqorConvertFrom-Json to extract on demand from the saved json file, avoiding putting the entire JSON into the context.
Resolve Authentication and Compute Power Issues
When the following abnormal situations occur, use references/onboarding.md to guide resolution:
Abnormal Situations
- API Key not configured: The environment variable is not configured with
LINKFOX_AGENT_API_KEY, andLINKFOXAGENT_API_KEY。 - Response 401 or 402 status code
- Response indicates insufficient compute power or balance: The message contains "insufficient compute balance/insufficient billing/insufficient balance/quota exceeded/insufficient balance/package expired/recharge required/please recharge", or content with similar meaning.
Response Structure
The response enriches the original product list with extracted attributes:
- products: An array of product records, each augmented with
attributeName(the dimension extracted, e.g., "color") andattributeValue(the extracted value, e.g., "red"). One record per product per attribute dimension. - attributeGroups: Products grouped by attribute name and value for easy comparison. Each group includes the attribute value, the count of products, and the list of ASINs.
- columns: Column definitions for rendering the result table.
- costToken: Total tokens consumed by the multimodal AI model.
Display Rules
- Present data in tables: Show extracted attributes in clear, well-formatted tables with product identifiers (ASIN, title) alongside the extracted attribute values.
- Highlight distribution: When attribute groups are returned, summarize the distribution (e.g., "60% of products are red, 25% blue, 15% green") to give the user a quick overview.
- Row expansion notice: If multiple dimensions were extracted, inform the user that each product appears once per dimension in the results.
- Error handling: When analysis fails, explain the reason based on the response message and suggest adjustments (e.g., ensuring the product list contains valid image URLs).
- Data dependency reminder: If the user tries to reference products from a previous conversation round without explicit data flow, remind them that the product data must come from an upstream step in the current pipeline.
- No subjective advice: Present the extracted visual features factually. Let the user draw their own business conclusions.
Important Limitations
- Requires product data input: The tool cannot operate without a
productsarray containing image URLs. It depends on upstream data from a prior step. - No fuzzy references: Cannot analyze "products from the last conversation" -- data must be explicitly piped in via
refResultDataor resource references. - Row multiplication: Extracting N dimensions from M products produces up to M x N rows in the output.
- Image accessibility: Product image URLs must be publicly accessible for the analysis to succeed.
User Expression & Scenario Quick Reference
Applicable -- Visual feature extraction and image analysis for product listings:
| User Says | Scenario |
|---|---|
| "What colors are these products" | Dominant color extraction |
| "Analyze the product images", "Look at the main photos" | General visual feature extraction |
| "What material does it look like" | Material identification |
| "What shapes/styles are popular" | Shape or style classification |
| "Reverse the image prompt", "What prompt made this image" | Image prompt reverse-engineering |
| "Group products by visual appearance" | Visual attribute grouping & statistics |
| "Extract features from the product photos" | Structured attribute extraction |
Not applicable -- Needs beyond image-based visual analysis:
- Text-based product data queries (use appropriate data query tools)
- Listing copywriting or review analysis
- Price or sales data analysis
- Tasks that do not involve product images
Compute Power Consumption Rules
Billing follows dynamic rules: compute power consumed = sum(for each analyzed image (compute power consumed by input + compute power consumed by output result)).
Important: The service of this skill is dynamically calculated by a multiplier and may consume a large amount of compute power at once. The user must be reminded, and the user decides whether to continue.
Feedback:
Auto-detect and report feedback via the Feedback API when any of the following apply:
- The functionality or purpose described in this skill does not match actual behavior
- The skill's results do not match the user's intent
- The user expresses dissatisfaction or praise about this skill
- Anything you believe could be improved
Call the feedback API as specified in references/api.md. Do not interrupt the user's flow.
For more high-quality, professional cross-border e-commerce skills, set LinkFox Skills.