watch-video

Watch Video

Build a timestamped account of what the video shows, says, and supports. Keep observations, spoken claims, and inferences separate.

Requirements

Local files require ffmpeg, ffprobe, Python 3.10+, and Pillow. Public URLs also require yt-dlp. Resolve SKILL_DIR to the directory containing this file. Run commands through $SKILL_DIR/scripts/run.py so they use the private .venv when present. If the runtime is missing and local dependency installation is allowed, run python3 "$SKILL_DIR/scripts/setup_runtime.py". Never sign in, load browser cookies, bypass DRM or paywalls, or work around access controls.

Workflow

1. Create a dedicated analysis directory. Never modify the source video.

2. Acquire the source and platform captions:

   python3 "$SKILL_DIR/scripts/run.py" acquire <video-path-or-url> \
     --output-dir <analysis-dir> [--language en]
   

   Local files are primary. Only accept a public URL when yt-dlp can inspect and download it directly without sign-in, cookies, paywalls, DRM, or browser automation. If direct retrieval fails, ask for an authorized local copy. For URLs, prefer manual platform captions, then automatic captions. If neither exists, record transcript unavailable and continue visually; do not generate a transcript. For local files, detect same-name .vtt, .srt, .json, or .json3 sidecars.

3. Read <analysis-dir>/source.json and <analysis-dir>/transcript.md when present. Treat captions as speech evidence, not visual evidence.

4. Build a coarse visual index:

   python3 "$SKILL_DIR/scripts/run.py" sample-frames <video-path-from-source.json> \
     --output-dir <analysis-dir>/visual-overview --mode overview --scene-threshold 0.32
   

5. Open every contact sheet listed in visual-overview/index.md. Record visible events, demonstrations, slide changes, on-screen text, and notable mismatches with the transcript.

6. Run targeted dense passes around important or ambiguous timestamps:

   python3 "$SKILL_DIR/scripts/run.py" sample-frames <video> \
     --output-dir <analysis-dir>/window-<start> \
     --mode dense --start <seconds> --end <seconds> --fps 4
   

   Use --mode every-frame only for short forensic windows. It samples with FFmpeg at the reported source frame rate, so variable-frame-rate material may contain duplicated or dropped samples. The script caps output by default. Increase the cap only when the user needs the densest available inspection.

7. Synthesize findings with timestamps. For each substantive claim, label its basis as:

   • observed: visible in sampled frames
   • spoken: present in captions or supplied transcript
   • inferred: a reasoned conclusion from observed or spoken evidence

8. Mark support as supported, conflicted, or not_found. State what evidence is missing when confidence is low.

9. For detailed work, save report.json using the schema in references/analysis-protocol.md, then validate it:

   python3 "$SKILL_DIR/scripts/run.py" validate-report <analysis-dir>/report.json \
     --manifest <analysis-dir>/visual-overview/manifest.json
   

   Repeat --manifest <path> for each dense window cited in the report.

Analysis rules

• Start coarse, then inspect important windows at higher temporal density.
• Never describe a transcript claim as something the video visibly proves.
• Quote captions sparingly; paraphrase by default.
• Attach timestamps to chapters, findings, discrepancies, and requested moments.
• Inspect frames at higher resolution when slides, code, diagrams, or small interface text matter.
• Say when captions are automatic, absent, incomplete, or out of sync.
• If the question spans many scattered facts, answer it in focused passes instead of one broad scan.
• Do not infer identity, intent, causation, or off-screen events without evidence.

Read references/analysis-protocol.md for sampling modes, the evidence schema, long-video handling, and failure cases.