See

Run scripts/ask_media.sh to analyze visual media. Never call ZenMux API directly.

The script already contains built-in analysis prompts:

• Image: focus on overall content, key elements, visible text, and reusable details
• Video: focus on subtitles / spoken content first, then visuals, actions, and reusable takeaways

Do not spend time rewriting prompts unless the user has a very specific extra focus. In most cases, just pass the file path or URL directly.

Usage

# Image
scripts/ask_media.sh --image /path/to/image.png

# Multiple images
scripts/ask_media.sh --image /path/a.png --image /path/b.png

# Video (auto-compresses if >45MB)
scripts/ask_media.sh --video /path/to/video.mp4

# URL (image, video, or webpage with embedded video)
scripts/ask_media.sh "https://example.com/photo.jpg"
scripts/ask_media.sh "https://youtube.com/watch?v=xxx"

# Optional extra focus
scripts/ask_media.sh --video demo.mp4 --task "重点看界面里的操作步骤"

# Optional output name
scripts/ask_media.sh --image photo.png --name landing-page-hero

# Custom output path
scripts/ask_media.sh --image photo.png -o /tmp/result.md

On success, prints output_path=<path> to stdout. Read that file for the result.

Options

Output

By default, outputs are stored in a shared directory so Codex and Claude can both find them easily.

Each output file includes:

• A metadata header with creation time, media type, source inputs, model, and any extra focus
• The parsed markdown result

Default naming pattern:

• YYYYMMDD-HHMMSS__image__<source>.md
• YYYYMMDD-HHMMSS__images__<source>.md
• YYYYMMDD-HHMMSS__video__<source>.md

Use --name when you want a cleaner project-specific filename.

API Key

Lookup order: ZENMUX_API_KEY env var → .env.local in cwd/parents → ~/.config/see/api_key. If missing, ask user.

Dependencies

• python3: required
• ffmpeg/ffprobe: required for video (brew install ffmpeg)
• yt-dlp: optional, for webpage video extraction (brew install yt-dlp)

Workflow

Before calling the script, decide whether you actually need it:

• Local image(s) + natively multimodal agent (e.g. Claude Code): use the Read tool directly — it renders images inline. Call Read once per image for multiple files. After viewing, respond to the user directly in context — no file output needed.
• Video (any source): always use this script — Claude has no native video understanding.
• Image URL, video URL, or webpage with embedded video: always use this script — the Read tool only works with local files.

If you do need the script:

1. Run the script with file path, URL, or --image/--video flags.
2. Only add --task if the user has a specific extra focus.
3. Read the output file and continue working with the parsed text.