Search Web → Markdown (illustrated research)

When to use

• The user asks you to research a topic on the web and output an illustrated Markdown document
• Images must be real images collected from the web (news photos, diagrams, screenshots, etc.) — AI-generated images must never be used
• Images are downloaded and saved as local files (e.g. an imgs/ directory) and referenced in Markdown with relative paths, with the document and image directory kept at the same level

Hard rules

1. Image source = real images from search result pages / the original web page; AI generation is forbidden
2. Every image must be attributed to its source (URL), and an "Image sources" list must appear at the end of the document
3. After embedding, you must verify that each image file exists and is displayable (see workflow step 6)
4. Images are always stored as local files; do not embed base64 (smaller document, reusable images, batch compression possible)

Workflow

1. Clarify requirements: topic, length, number of images (default 5–8), language
2. Search: use browser_use to open Bing (https://www.bing.com/search?q=...), browse 6 highly relevant pages in turn; use get_readable to extract the article text; use get_backbone to inspect the DOM structure when needed
3. Extract image URLs: for each page use execute_js to collect <img src> (filter out data: URIs, images ≤80px, logos/icons), prefer large / original images, and deduplicate
4. Download and save as local files: first create the image directory mkdir -p imgs (same level as the document), then call scripts/image_to_local.sh:
   scripts/image_to_local.sh "<image-url>" "imgs" "01-topic-description.jpg" "caption text" >> output.md
   - The script downloads the image, detects the MIME type, saves it under imgs/, and prints one line of ![caption](imgs/01-topic-description.jpg)
   - Suggested filenames: <index>-<short-topic-description>.<ext> (non-ASCII filenames can cause compatibility issues, so prefer pinyin/English)
   - If the download fails or the file is not an image, the script exits with an error → retry with a fallback URL
   - For images larger than 3MB, compress before saving:
   ffmpeg -i imgs/in.jpg -vf "scale='min(1600,iw)':-2" -q:v 5 imgs/out.jpg
5. Assemble the document: title + introduction → sections by topic (1–3 relevant images each, with the source link after each image) → conclusion → "Image sources" list
6. Validate:
   - Count the ](imgs/ relative references in the document and confirm it equals the number of images
   - Use a script or shell to confirm that every file referenced under imgs/ actually exists and is non-empty
   - Make sure no absolute-URL image references remain
7. Deliver: save to /var/minis/workspace/ and provide the Markdown link; deliver the image directory together with the document

Details

• Prefer Bing as the search engine; if blocked by anti-bot measures, switch to Baidu / DuckDuckGo
• Quote content faithfully and cite the source URL; never fabricate data
• Keep the document small and fast to load; images can be reused separately (e.g. for a PDF or PPT later)
• Put temporary files in /tmp so they don't pollute the workspace; write captions in the user's language
• If the document must be distributed to another device/platform, package the image directory too (zip) so the relative-path references keep working

Script

scripts/image_to_local.sh <image-url|local-path> <output-dir> [filename] [alt-text]
Downloads the image into the output directory and prints one line of Markdown relative-path image syntax to stdout (append it to the document with >>).

minis_url: minis://skills/web-search-make-markdown/SKILL.md