tracked-video-analysis（跟踪视频分析）

Tracked Video Analysis

Use this skill for long, noisy, or operationally awkward videos where trust and visibility matter as much as the final summary.

The core idea is simple:

1. Extract content first
2. Structure it second
3. Track both stages explicitly

Never claim that a background process is still running unless a live OS process or a fresh status file proves it.

Core workflow

1) Acquire the video reliably

Prefer this order:

1. direct local file
2. direct downloadable link
3. document upload
4. external file host fallback

If chat media is inaccessible, ask for a direct link instead of retrying vague media access indefinitely.

Use tmp/video_analysis/ as the working directory.

2) Prepare local tools without root

Prefer workspace-local packages over system installs.

Useful local tools:

• ffmpeg-static
• ffprobe-static
• @xenova/transformers
• wavefile

If root/elevated package install is blocked, do not stall the task—install locally in the workspace when possible.

3) Run tracked extraction

Extraction should produce:

• tmp/video_analysis/status.json
• tmp/video_analysis/progress.log
• tmp/video_analysis/transcript.jsonl
• tmp/video_analysis/analysis.md

Rules:

• Prefer chunking over one-shot whole-video ASR.
• Prefer lighter ASR first for stability.
• Update status after each chunk.
• If a run dies, resume from files when practical instead of starting from zero automatically.

4) Run tracked final structuring

Structuring should produce:

• tmp/video_analysis/final_status.json
• tmp/video_analysis/final_progress.log
• tmp/video_analysis/final_analysis.md

This stage should:

• clean filler and repeated phrases
• group related chunks
• infer categories
• normalize wording
• convert raw transcript into the user’s requested format

5) Report status honestly

Use these rules:

• Extraction running → report status.json
• Extraction complete, no final process running → say so plainly
• Final structuring running → report final_status.json
• Final result ready → read final_analysis.md and answer normally

Standard output formats

Common targets:

• Category → function → description → benefit
• Category → function → short description
• Function list + timestamps
• Clean summary with confidence caveats

For noisy ASR, prefer readable normalization over false precision.

Status discipline

Do not say “the process is running” unless at least one of these is true:

• the OS process is alive
• the relevant status file is actively updating

If extraction finished, explicitly say:

• extraction is complete
• no live extraction process remains
• only structuring remains (if true)

Read these files when needed

• Read references/pipeline.md for the canonical tracked workflow and failure handling.
• Use scripts/transcribe_tracked_light.mjs for extraction as a starting point.
• Use scripts/final_structurer.py for initial structuring as a starting point.

Delivery style

Prefer concise, readable sections.

When the user wants a polished deliverable:

1. create a clean .md file
2. keep the structure visually pleasant
3. send it as a document/file if requested

Practical note

This skill is optimized for operational reliability, not perfect transcription fidelity. If ASR is messy, produce a useful structured summary with explicit uncertainty rather than pretending the raw transcript is exact.