Minimax Video Analyzer

MiniMax Video Analyzer

Use this skill to analyze an existing local video with MiniMax-M3 through the China Mainland MiniMax API. It is optimized for software operation recordings: small programs, apps, web pages, admin panels, product demos, interaction flows, UX review, and motion-effect replication.

This skill is intentionally narrow: upload a local video, ask MiniMax-M3 to analyze the visual content, and return the text result. Do not use it for video generation, audio transcription, image generation, TTS, or music generation.

Capabilities

• Analyze local videos in mp4, avi, mov, or mkv format.
• Upload the video with purpose=video_understanding.
• Call MiniMax-M3 through the OpenAI-compatible /v1/chat/completions endpoint.
• Reference uploaded videos with mm_file://<file_id>.
• Return the model's text analysis to stdout.
• Prefer a software-analysis report when the video is an operation recording: action timeline, UI state changes, interaction feedback, UX issues, and animation/motion breakdown.
• Fall back to general visual analysis when the video is not a software interface.

Important Limits

• MiniMax-M3 video understanding does not support audio input. Treat the result as visual analysis only.
• Uploaded video_understanding files are retained by MiniMax for up to 7 days.
• The upload API supports videos up to 512 MB.
• Base64 inline video can work for small files, but this skill deliberately uses file upload by default because it is simpler and more stable.

Environment

The script reads the API key from:

MINIMAX_API_KEY

For the user's China Mainland Token Plan, the default API base is:

https://api.minimaxi.com/v1

If the current shell cannot see the key but it exists in ~/.bashrc, load it first:

source ~/.bashrc

Usage

Run the bundled script from any working directory:

python ~/.config/opencode/skills/minimax-video-analyzer/scripts/analyze_video.py \
  ~/tmp/opencode/6-真假难辨.mp4 \
  "请基于画面客观分析这个软件操作录屏，重点说明用户操作流程、界面状态变化、交互体验和可复刻的动效细节。"

If no prompt is provided, the script uses a default Chinese prompt that prioritizes software operation analysis and falls back to general video analysis when the content is not a software interface.

Default Analysis Focus

The default prompt asks MiniMax-M3 to stay objective and separate visual evidence into [观察], [推断], and [不确定]. For software videos, it focuses on:

• User operation timeline: clicks, swipes, inputs, navigation, and visible feedback.
• Page, component, and state changes: dialogs, toast messages, loading states, tabs, lists, buttons, selected/disabled states, and refresh behavior.
• UX review based on visible evidence: clarity, feedback timing, information hierarchy, hesitation, blocking, ambiguity, or possible mis-taps.
• Motion-effect replication: trigger, animated object, start/end state, changed properties, approximate duration, easing/pace, and engineering notes.
• UI structure and visual hierarchy: layout, components, color, text size, spacing, alignment, layering, and reusable component list.
• General visual fallback: camera, composition, movement, lighting, animation, transitions, effects, and editing rhythm when the video is not a software recording.

Useful options:

python ~/.config/opencode/skills/minimax-video-analyzer/scripts/analyze_video.py VIDEO_PATH [PROMPT] \
  --detail default \
  --timeout 120

Workflow For Agents

1. Confirm the user is asking to analyze an existing local video.
2. Check that MINIMAX_API_KEY is available in the current environment.
3. If the user provides a custom question, preserve it. If not, rely on the default prompt for software operation, UX, and motion-effect analysis.
4. Run scripts/analyze_video.py with the video path and the user's analysis question.
5. Report the returned text. If the user asks about spoken dialogue or sound, clearly state that MiniMax-M3 video input does not include audio understanding in this workflow.

Do Not Add

• Do not add ASR, audio extraction, or transcript merging unless the user explicitly asks for audio-aware analysis.
• Do not add caching or file deletion by default.
• Do not switch to base64 inline upload unless the user explicitly requests that path.
• Do not use the video generation skill for this task.