Video Ingest (crv → frames ⇄ transcript → LLM Wiki)

Actually watch a recording: crv extracts deduplicated keyframes + a Whisper transcript; you then correlate what's SHOWN with what's SAID, and ingest it the vault's way — full capture in Raw/Sources/, reusable knowledge compiled into Wiki/, key screenshots in assets/. Follow AGENTS.md and the llm-wiki-ingest skill for the compile step.

Environment

• crv is the claude-real-video pip package (CLI: crv). Check crv --help; install with pip install claude-real-video if missing.
• ffmpeg/ffprobe must be reachable. If not on PATH (common on Windows), prepend for the run, e.g. $env:PATH = "C:\ffmpeg\bin;" + $env:PATH.
• On Windows PowerShell set $env:PYTHONUTF8=1 in the same command (avoids a cosmetic Unicode crash on the final print; output still lands).
• Whisper on CPU is slow (~0.5–1× real-time). Run crv in the background and wait for completion; don't poll.

Step 0 — Pick the right input

• URL? crv takes URLs directly (YouTube, TikTok, Instagram, ...): crv "https://youtube.com/watch?v=..." -o <OUTDIR> .... For login-gated sites, --cookies <netscape-file> or --cookies-from-browser <browser> (the user's own account, authorized use only).
• Audio-only file (.wav/.mp3)? No frames → no "what's shown". Screen recorders (OBS) often save a big .mkv/.mp4 and a small .wav sibling — prefer the video. Confirm streams with: ffprobe -v error -show_entries stream=codec_type,codec_name:format=duration <file>.

Step 1 — Extract (background)

$env:PYTHONUTF8=1
crv "<VIDEO-OR-URL>" -o "<SCRATCH-OUTDIR>" --lang en --whisper-model small `
    --adaptive --text-anchors --viewer --max-frames 200 `
    --why "Full examination for the vault: capture everything said and shown."

Use a scratch <OUTDIR> outside the vault. Produces transcript.txt, frames/ (frame_001.jpg… chronological), MANIFEST.txt, viewer.html. Frame selection is visual only (scene change + dedup + a time floor): crv guarantees coverage of distinct screens, not importance. Importance is judged next — by you, using the transcript.

Step 2 — Watch it: correlate frames ⇄ transcript

This is the step that makes the ingest worth anything.

1. Read transcript.txt in full. It has no timestamps and no speaker labels; Whisper often hallucinates on trailing silence — drop the nonsense tail. (Known quirk: domain terms get misheard consistently; when a word looks wrong everywhere, flag it as an open question instead of propagating it.)
2. Read the frames in order. Don't read all N blindly: do a strided overview pass (every ~8th frame) to map the arc of the recording, then read densely through the sections that matter (code, terminals, UIs, slides, whiteboards, diagrams).
3. Align said ⇄ shown by chronological position: frames and transcript both run start→end, so a frame ~40% through the frame list pairs with text ~40% through the transcript. Anchor pairs at unmistakable moments (a screen change that matches a topic change) and interpolate between anchors. Record the pairing explicitly in the Walkthrough ("What's said / What's shown").
4. Redact secrets. Recordings of dev work routinely expose tokens, keys, account IDs, URLs with credentials, passwords. Never transcribe secret values and never copy a frame that shows one into the vault. Transcribe other on-screen text/code verbatim where legible.

Step 3 — Ingest the vault's way

• Raw source → Raw/Sources/<date>-<slug>-recording.md with source frontmatter (Title, Author, Reference = the URL or filename, ContentType, Created, Processed: false, tags: [source]). Body: a provenance note (crv flags, whisper model, frame count, redactions, transcript quality), a Summary, a chronological Walkthrough (What's said / What's shown pairs), the embedded screenshots, a Key topics list of [[wikilinks]], and the full transcript appended at the end for searchability.
• Screenshots → copy only the important, secret-free frames into assets/ with descriptive names; embed with ![[name.jpg]]. Do not dump all frames or the source video into the vault (Raw/Files/ is gitignored; assets/ is committed).
• Compile → search first (python scripts/wiki_tool.py search-catalog --query "…"), then create or enrich focused Wiki/ notes, each linked back to the Raw source in sources with an accurate source_count. Prefer enriching existing notes over duplicating. Capture ambiguity as Open Questions; never invent.

Step 4 — Gate + commit

python scripts/wiki_tool.py build
python scripts/wiki_tool.py lint
python scripts/wiki_tool.py source-scan --update --accept-covered
python scripts/wiki_tool.py source-lint
python scripts/audit_public.py

Commit the Raw source, assets/ screenshots, compiled notes, and regenerated artifacts together. Add a log note for substantial ingests.

Notes

• Reference-video mode (a TikTok/YouTube the user wants studied, not archived): same Steps 0–2, then answer their question directly; only ingest to the vault if they want it kept.
• Optional crv flags: --grid (contact sheets), --keep-audio, --report, --scene 0.2 (more frames), --whisper-model medium (better transcript).
• Big videos stay outside the vault; reference them by URL/filename in the Raw source's Reference.