Reeldigest: Videos into clean, searchable transcripts

The knowledge you need is spread across dozens of tutorials and long livestreams, and you cannot search a video. Captions are missing for some languages, downloads return HTTP 403 and a six hour transcription stops halfway with no way to resume. Reeldigest turns a curated list of videos into clean transcripts and a usable digest.

What it does

• Manifest first with id, url, topic, language and duration, and an output mode chosen up front: knowledge or signals.
• Captions first with yt-dlp, then checks that a .vtt file actually exists for each id instead of trusting the exit code.
• Audio fallback for videos without captions, using a logged-in browser profile for cookies.
• Chunked, resumable transcription with mono Opus chunks of about 14 minutes, cached by video id and chunk index.
• Offset merge adds each chunk's offset to its timestamps for one transcript per video.
• Leak check greps the output for URLs, video ids, channel names and handles before commit.

How it works

1. Curate the video list and freeze it in manifest.json.
2. Fetch captions, or audio when there are none, then chunk with ffmpeg.
3. Transcribe with Whisper locally or a hosted API with a failover provider, then clean rolling-caption duplicates.
4. Digest into original articles per topic (knowledge) or verbatim quotes with timestamps in signals.json (signals), spot-checked against transcripts.

Frameworks & tools

Uses yt-dlp, ffmpeg and Whisper (whisper.cpp, faster-whisper or a hosted endpoint). Discovery of channels can use search or Playwright.

Why this beats prompting it yourself

A prompt cannot resume a stalled six hour job or notice that a caption was never written. Reeldigest verifies files on disk, caches every chunk and records an explicit failure reason for any video it could not transcribe.

Use cases

• Knowledge base from tutorials as original articles with no source credits shown.
• Signal or sentiment corpus of verbatim quotes with timestamps.
• Batch analysis through your own transcription service.

Known limitations

Not for republishing someone else's video or a single short clip. Scheduled runs need a pipeline that works without a logged-in browser session. Transcripts and audio stay gitignored.