Analyze Video Evidence

Turn a video into a reviewable evidence bundle, then answer the user's actual question from that bundle. Prefer source-backed findings over an untraceable generic summary.

Inputs

Accept either:

• a local video plus optional SRT/VTT and yt-dlp .info.json; or
• a Bilibili/YouTube URL downloaded first with $download-web-video.

Require ffmpeg and ffprobe. Keep downloaded media, full transcripts, and frames local unless the user authorizes publication.

Prepare the first-pass evidence

Run:

python3 <skill-dir>/scripts/prepare_evidence.py \
  "./video-source/BV....mp4" \
  --output-dir "./video-evidence" \
  --subtitles "./video-source/BV....zh.srt" \
  --info-json "./video-source/BV....info.json" \
  --frame-count 24

If captions are omitted, the script prefers a nearby human Chinese caption, then automatic Chinese, English, and other SRT/VTT files. Do not claim that speech was analyzed when no captions exist.

When captions are absent:

1. Extract/transcribe audio with an available local or user-approved ASR route, such as the AX $transcribe skill.
2. Save SRT or VTT beside the video.
3. Rerun the evidence preparation command with --subtitles.

Inspect before concluding

Read transcript.md, evidence-index.md, and evidence.json. View contact-sheet.jpg, then open ambiguous individual frames at full resolution. The first pass samples the complete timeline rather than only transcript-picked moments.

Identify topic transitions, enumerated items, tables, slides, on-screen labels, and places where the creator revises or qualifies an earlier point. Add exact frames for those moments:

python3 <skill-dir>/scripts/prepare_evidence.py VIDEO \
  --output-dir "./video-evidence" \
  --subtitles CAPTIONS \
  --timestamp 00:08:42 \
  --timestamp 00:19:15

Read references/analysis-method.md when the task involves rankings, recommendations, factual verification, conflicting captions, or a publishable report.

Build the claim-evidence ledger

Before writing, map each material finding to:

• a transcript time range and short paraphrase;
• a frame path when the claim depends on visible content;
• confidence (high, medium, or low);
• a caveat when the source is context-dependent or ambiguous.

Distinguish the creator's explicit statement from your inference. Never invent card names, slide text, numbers, or tiers to fill a gap.

Produce the requested artifact

Match the user's requested format. For a table or HTML guide:

• lead with the decision-useful result;
• make tiers or priorities scannable;
• preserve conditions such as act, archetype, existing synergies, or upgrade state instead of forcing a universal order;
• attach compact timestamps to every important recommendation;
• include a methodology/source note and uncertainty labels;
• do not embed the full transcript or redistribute a large frame collection.

Use the frontend design workflow for a polished HTML artifact and visually verify responsive layout with browser screenshots.

Acceptance criteria

Require all applicable conditions:

• evidence.json, evidence-index.md, transcript.json, and transcript.md exist.
• The evidence manifest records media metadata, source SHA-256, caption hash, cue count, frame timestamps, and nearby caption context.
• A full-timeline contact sheet exists when frames were requested.
• Consequential claims have timestamped speech or visual evidence.
• Inference and source statements are labeled separately.
• The final artifact answers the requested question, opens locally, remains readable on desktop and mobile, and contains no cookies or signed media URLs.

Output bundle

<evidence-dir>/
├── evidence.json
├── evidence-index.md
├── transcript.json
├── transcript.md
├── contact-sheet.jpg
└── frames/
    ├── uniform_...
    └── targeted_...