Reel Recipes

How you’d use it

You paste a link to a cooking video and it pulls the captions (or transcribes the audio), reads the recipe out of it, and saves a clean card (ingredients, step-by-step method, timings, and tags) that you can search, filter, and favorite.

Here's what using it actually looks like:

1. 1

   Paste a video link

   Any YouTube, Shorts, Instagram, or Facebook cooking reel, just drop in the URL.

2. 2

   It reads the recipe

   It grabs the captions or transcribes the audio, then pulls out the ingredients, steps, and timings.

3. 3

   You get a clean card

   A structured recipe with a checkable ingredient list and numbered steps, no ads, no life story, no scrolling.

4. 4

   Build your collection

   Search, filter by tag or cuisine, and star your favorites so your go-to recipes are one click away.

How it’s built

The transcript is the hard part, so it degrades in tiers: it first tries the video's own captions, falls back to local Whisper transcription of the audio when there are none, and finally to the description alone, so it almost always has something to work with. Then a JSON-schema-constrained model turns messy, unpunctuated cooking narration into a strict recipe shape (quantities, units, ordered steps, a confidence rating), and the structuring engine is swappable between the Claude subscription CLI, a local Ollama model, or the Anthropic API.

• yt-dlp pulls video metadata, captions, audio, and a thumbnail; Instagram/Facebook get an automatic cookie-based retry
• Three-tier transcript fallback: video captions → local Whisper → description-only
• A JSON-schema-constrained model structures the transcript into a validated recipe (ingredients, steps, timings, tags)
• Swappable structuring backend: Claude Code CLI, local Ollama, or the Anthropic API
• Saved locally to SQLite with full-text search, tag filters, and favorites

• Next.js
• SQLite (better-sqlite3)
• yt-dlp
• Whisper
• Claude