| name | blog-image-finder |
| description | Search and download candidate blog images through a local Unsplash-backed MCP server while filtering out previously used assets and preserving attribution metadata. Use when Codex needs to source a fresh image for a blog post, compare candidate photos for an article concept, download a selected image into the configured local staging folder, or check the local image-provider setup before image work. |
Blog Image Finder
Overview
Use this skill to source blog images through the local MCP server at mcp/image-provider/server.py. It resolves the active Ink profile first, searches Unsplash, skips images already recorded in that profile's local history file, downloads the selected image into that profile's configured staging folder, and returns the attribution data needed for later publishing.
Workflow
- Read
references/setup.md.
- Resolve the active Ink profile. If multiple profiles exist and none is named, ask which profile to use before image work.
- Assume the repo-local MCP registration is already loaded and call
search_images directly. Pass profile when it is known. Do not run MCP verification as a preflight step.
- Treat the returned list as candidate inputs, not final approval.
- Form
search_images.query as a short visual concept, usually 1-4 words, not as a sentence or article summary.
- Search for what can actually be photographed: a place, object, mood, or simple scene. Use metaphor or adjacent imagery when the article idea is abstract.
- Start broad, then narrow only if the first results miss the brief.
- Good queries:
empty office, quiet workspace, single desk, small team meeting, warehouse shelf.
- Bad query:
a lean technology team overseeing a dense stack of dashboards, agents, and workflows, not a crowded open office.
- Pick one image that fits the article and call
download_image with its image_id and the same profile.
- Keep the returned provider, source URL, photographer credit, and attribution text with the asset record.
- Hand the downloaded file to
../blog-image-uploader/SKILL.md when the user wants the asset uploaded to the selected profile's blog image bucket.
Rules
- Do not claim an image is rights-free. Preserve the provider and attribution metadata exactly as returned.
- Do not bypass the provider download flow with a raw image URL.
- Do not save downloaded files outside the configured
download_dir.
- Do not mix image histories across Ink profiles. The selected profile's
imageProviderConfig owns the staging folder and history file.
- Do not pass long descriptive prompts to
search_images. Use short, searchable visual concepts instead.
- Do not verify MCP setup before using the tool. Call the tool first and troubleshoot only after a concrete tool failure.
- Treat the download history as the dedupe source of truth. If the same image was already downloaded, choose another one unless the user explicitly wants the duplicate.
- Return the config error as-is when the selected profile's image provider config is missing or invalid.
- Keep the skill narrow. Use it for image sourcing and download, not for article writing or upload work.
Files
- Read
references/setup.md for config shape, tool contracts, and the MCP config snippet.
- Use
mcp/image-provider/server.py as the local MCP server entrypoint.
