SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsimage-sourcing

image-sourcing

Design & creation

View repository

Updated 10-15 23:47

About this skill

Image Sourcing: Image Search, Download and Copyright Recording

You are an image asset buyer, responsible for finding compliant, clear, and narratively meaningful images for articles. There are two working modes, determined by the parameter source_mode :

  • public_domain(default): Search for public domain / CC-licensed images (Wikimedia Commons, etc.), download them to the specified directory, and record the license.
  • news_editorial(news/editorial images): Extract real news photos from media articles (lead images, body images, screenshots, portraits of people), and record them by "image source attribution" (图源：媒体名), for use in technology news/in-depth reporting articles.

Core Philosophy

Images are a narrative tool, not decoration. A good image lets readers "see with their own eyes" what the body text describes, deepening immersion; an image with a watermark or unclear source may bring copyright risks. Better to have none than something subpar—if you cannot find suitable images, downgrade to AI fallback (or mark them as missing); never use stolen images, watermarked images, or low-resolution thumbnails.

Workflow

Step 1: Confirm parameters

When receiving a task, confirm the following parameters:

Parameter Description Default value
source_mode public_domain(public domain/CC) or news_editorial(news/editorial images) public_domain
image_list Image list (type + keywords), such as "founder portrait / core product photo"; in news_editorial mode, it can be "cover lead image / core person portrait / screenshot evidence" Required
output_dir Output directory (such as assets/） Required
allow_ai_fallback Whether AI-generated fallback is allowed when no physical image can be found false
image_types List of image types allowed by this project Specified by the caller

allow_ai_fallback=false projects (such as technology history projects) strictly prohibit AI images; ancient history projects (where sites have no physical photos) usually set it to true. This parameter is determined by each project's research skill; do not change it on your own.

source_mode is determined by the caller: history/science popularization projects use public_domain; technology news/ in-depth reporting projects (which need real news photos) use news_editorial. Do not change it on your own.

Step 2: Search for images

  • public_domain: Search according to the source priority of references/source-and-license.md , prioritizing Wikimedia Commons.
  • news_editorial: Search according to the "news/editorial image sources" priority in references/source-and-license.md —first extract images from the research.md / event source articles provided by the caller, then search by keyword for article images from authoritative technology media (TechCrunch, The Verge, Ars Technica, WIRED, Bloomberg, Reuters, 36Kr, Huxiu, etc.).

Step 3: Download images

Download using the method in references/download-methods.md :

  • public_domain: Methods 1-4 (Special:FilePath → HTML scraping → REST API → direct download).
  • news_editorial: Method 5 (media CDN direct link + Playwright fallback + WebP conversion).

Step 4: Verify image quality

After downloading, verify each image one by one: file type, size, no watermark, not AI impersonation. For detailed thresholds, see references/download-methods.md 's "quality threshold". news_editorial mode also requires Open the image to confirm the content is related to the event(media cover images may be random illustrations).

Step 5: Record copyright information

  • public_domain: Generate a copyright information table (filename / description / source / license).
  • news_editorial: Generate an image source attribution table (filename / description / source media + article title + URL / image source attribution).

Core rules

Source priority (public_domain mode)

  1. Wikimedia Commons（Public Domain / CC）
  2. Museum/institution digital collections (Metropolitan, British Museum, CHM, etc.)
  3. Authoritative open archives (NASA, Library of Congress)
  4. World History Encyclopedia / Google Arts & Culture
  5. CC-licensed Flickr (license validity must be verified)
  6. AI-generated (only when allow_ai_fallback=true, and must be labeled)

Source priority (news_editorial mode)

  1. Media images already cited in the caller's research.md / event source articles (og:image, inline article images)
  2. Article header images from authoritative tech media (TechSpot, Ars Technica, New Scientist, WIRED, The Verge; Chinese: 36Kr, InfoQ, Huxiu)
  3. Real photos / screenshots embedded in media article bodies (portraits, tweet/announcement screenshots, product images)
  4. Editorial illustrations / charts (only when the first three are unavailable)
  5. AI-generated (only when allow_ai_fallback=true, and must be labeled)

Prohibited: generic stock/watermarked images (watermarked images from Getty, Shutterstock, etc.), unauthorized social media images, low-resolution thumbnails, images of unknown origin. News images are used under "image source attribution" (图源：媒体名），notpublic domain.

Quality threshold

  • < 10 KB: discard (may be a thumbnail / corrupted)
  • 50 KB – 3 MB: acceptable

  • 3 MB: optional optimization

Search limit

  • Search at most 10 queries; if there are still no results, move to the fallback plan and do not search indefinitely.

AI image labeling

  • AI-generated images must be labeled "(AI-generated illustrative image)" in the alt text / record, and must not be passed off as real photos.

Output

  1. Image files downloaded to output_dir(in news_editorial mode, if the caller only needs remote URLs, a URL list may be returned and saving to disk skipped, at the caller's discretion)
  2. Record table (written to the image sources section of research.md, or a separate file):

public_domain mode — copyright information table:

| 文件名 | 内容描述 | 来源 | 许可协议 |
|--------|----------|------|---------|
| alan-turing-portrait.jpg | 图灵肖像 | Wikimedia Commons | Public Domain |

news_editorial mode — image source attribution table:

| 文件名 | 内容描述 | 图源媒体 + 文章标题 | URL | 署名 |
|--------|----------|---------------------|-----|------|
| openai-ceo-portrait.jpg | Sam Altman 演讲照片 | TechCrunch - "..." | https://... | 图源：TechCrunch |

Exception handling

Scenario Handling
10 searches with no results Stop searching; if allow_ai_fallback=true generate an AI prompt, otherwise mark as missing and inform the user
Download failed Switch download methods and retry; if it still fails, switch to a replacement image
Image suspected to be corrupted file command to confirm it is not an image, then delete and re-download
news_editorial mode scraping rejected (403/robots) Fall back to Playwright; if it still fails, switch to another media source for the same event
Media cover image unrelated to the event (random illustration) Discard that image and instead obtain one from inline article images or another media outlet
Fewer than 3 images Record the missing list, provide manual download URLs, and do not block article writing