Stock Image Search

When to Trigger This Skill

Trigger immediately — without asking the user to choose a search method — whenever the user wants an image, photo, or picture. This covers both:

• Finding existing imagery: "find a stock image", "search for a photo of X", "get a royalty-free image", "find an ethical image", editorial or news photography
• Requests phrased ambiguously as creating or generating: "create an image of a sunset", "generate a picture of a team meeting", "make me a mountain image" — treat these as stock search requests; this skill searches a stock photo library, it does not generate images.

In all of these cases, search the stock library — this is the tool for obtaining an image. Do NOT offer alternative search methods (hybrid search, user-provided URL/path, etc.) and do NOT ask the user to pick one.

Does not apply to:

• Searching internal CMS / Data Cloud media (use experience-content-media-search)
• Requests that explicitly ask for AI-generated, synthetic, or computer-generated imagery — this skill only searches licensed stock photography and has no image-generation capability

Step 1 — Search

MCP tool: search_stock_images (server: media-management) Timeout: 30 seconds

Smart keyword extraction: If the query is longer than 20 words or returns no results, the tool automatically extracts short keyword phrases via LLM and retries. The effectiveQuery field in the response shows what was actually sent to the search service (null if the original query was used unchanged).

Response fields:

Each images[] item:

Step 2 — Present Results, then STOP

Render a numbered list with an inline thumbnail for each result, using thumbnailUrl for the image and previewUrl as the click-through link (clicking the thumbnail opens the full preview in the browser):

1. [![<title>](<thumbnailUrl>)](<previewUrl>)
   **<title>** — <artist>, <width>×<height>, <licenseModel>
2. ...

This is standard markdown image-in-link syntax ([![alt](thumbnailUrl)](previewUrl)) — it renders an actual thumbnail, not just a text link, and clicking it navigates to previewUrl. If a client doesn't render markdown images, this degrades gracefully to a linked title.

Prefix the list with Searched for: <effectiveQuery> only if effectiveQuery is non-null and differs from the user's original query.

End the turn here. Do not auto-download, do not write files, do not ask a follow-up question. Agentforce injects its own selection UI — a question or auto-select causes a duplicate "double ask". Do not pick a result on behalf of the user.

Non-interactive mode (scheduled/headless, no user present): pick images[0] from the results and note the auto-selection in the report. Then proceed directly to Step 3.

Step 3 — Download (billed)

⚠️ Each call licenses the image and spends a stock-image download credit. Only invoke after the user explicitly selects an image (or in non-interactive mode, as noted above).

Before calling download_stock_image, verify:

• [ ] User has explicitly named or numbered the image they want (or non-interactive mode is confirmed)
• [ ] This is not an automatic retry of a previous download

MCP tool: download_stock_image (server: media-management) Timeout: 60 seconds

Use largest only when the user explicitly asks for full/original resolution.

Response fields:

Step 4 — Download to Client and Preview

After a successful download_stock_image call, download the image locally and open it in VS Code.

You MUST use download-stock-image.py for this. Do not use curl, wget, or any other tool.

python3 scripts/download-stock-image.py \
  --url "<url from download response>" \
  --id "<assetId>" \
  --format "<format from download response>" \
  --preview

The script retrieves org credentials via the sf CLI (sf config get target-org and sf org auth show-access-token), downloads the image with Authorization: Bearer <accessToken> to stockimages/<assetId>.<format>, and opens it in VS Code if --preview is passed.

Pass --output-dir <path> to override the default stockimages/ directory.

Never resize or post-process the image. Use CSS (width, height, object-fit) to control display dimensions.

Vector/binary formats (eps, tiff): always pass the exact format returned by the download response — never substitute a format like jpg to work around a validation error; the file bytes won't match the extension and the file will appear corrupted. VS Code cannot render eps/tiff previews, so the script skips --preview for those formats automatically and still reports the saved path.

Step 5 — Report Back

After a successful download, tell the user:

• Image title and artist
• Local path to the saved image (printed by the script)
• managedContentBodyId
• Attribution: {artist} · {collection} (use the attribution string returned by the download response when present)

Fallbacks & Errors

Placeholder

If search_stock_images and download_stock_image are both unavailable, use this URL directly as src or CSS url():

https://cdn.scs.static.lightning.force.com/content/assets/d5222d4a11e6c2b735152d7eea824ce4/placeholder.svg

Placeholder policy: There is only one placeholder URL. Do not download it, modify it, or generate alternative placeholders using Python, ImageMagick, or any other tool. If the user asks for a placeholder of a specific size or format, tell them only this URL is available and direct them to use CSS to scale it at the point of use.