video-readlisted
Install: claude install-skill stanleypdp/Claude-Video-Journalist
# video-read
Produces the files a `claim-cards` consumer reads. Everything runs on the user's machine (ffmpeg, faster-whisper, rapidocr); the only paid step is **one Claude CLI call per new video**, and a video already seen (same SHA-256) costs nothing.
## What you get
For a source `X` the output goes to `OUTPUT/X/` (details in `references/output-format.md`):
```
INDEX.md one row per video: id, profile, lang, status, thesis, domain, link to the card
cards/<id>.md claim card (~600 tokens): thesis, reported numbers with `@s` locators, method, gaps
articles/<id>.md faithful narration of speech and screen + a table of data shown on slides
archive/<id>.tx.md transcript in [mm:ss] blocks (the evidence; open only to verify)
archive/<id>.screen.md de-duplicated on-screen text (OCR) with first/last time seen
archive/<id>.rejections.md what the code removed and why (diagnostics, not a source of facts)
```
## Run it
Check the environment once, then read. From this skill's `scripts/` folder:
```bash
python scripts/preflight.py # ffmpeg, whisper, claude (yt-dlp only for URLs)
python scripts/pipeline.py <file|folder|URL> # preflight -> dry run -> read -> coverage check
```
The same steps individually: `read.py <file|folder|URL> [--dry-run] [--recursive] [--output DIR] [--captions FILE] [--no-wait]`, `check.py <OUTPUT/X>`, `rebuild.py <OUTPUT/X>`, `python -m vjournalist index <OUTPUT/X>`, `tx.py <id> --at 53 --w