← ClaudeAtlas

video-to-jsonlisted

Extract structured JSON that conforms to a JSON Schema from unstructured video, social-post, or article content (a recipe from a cooking video, specs and verdict from a review, steps from a tutorial, an itinerary from a travel vlog, product mentions from a TikTok). Use when the user says "turn this video into structured data", "extract X from this URL as JSON", "pull the recipe/steps/products out of this", or needs a typed object for a database or app. Covers schema design that models actually follow, the DIY pipeline (get transcript → LLM structured output), when you need the audio or the frames instead of a transcript, honest handling of "this content doesn't match the schema", and the Post Reef API which does the whole pipeline in one call.
franciscobmacedo/postreef-skills · ★ 0 · Data & Documents · score 73
Install: claude install-skill franciscobmacedo/postreef-skills
# Video / post / article → JSON conforming to a schema Two halves to this job: (1) getting the content into a form a model can read, (2) getting a model to fill a schema without inventing things. Most bad results come from skipping the design work in (2) or feeding a transcript to a schema that needed the picture. ## 1. Pick the input the schema actually needs | Schema wants | Input needed | Why | |---|---|---| | What was said (recipe steps, claims, quotes, Q&A) | Transcript | Cheapest; captions carry the words | | What people think (sentiment, questions, corrections) | Comments | Creators rarely say what's wrong; commenters do | | Tone, music, delivery, or the video has no captions | Audio | Models can listen; captions can't be listened to | | Anything only on screen: text overlays, on-screen ingredient lists, product shots, visual steps, before/after | Video frames | Cooking Shorts often show quantities as overlays and never say them | | Article/blog post | Page text + lead image | Not a video; don't run it through a video pipeline | If in doubt start with transcript + comments and only add audio/video when the result is thin. This ordering is also what keeps cost sane. ## 2. Design a schema the model will follow Full rules and a worked example in `references/schema-rules.md`. The short version: - Top-level `"type": "object"` with `properties`; each property has a `type` and a **`description` written as an instruction** ("Exact product name as shown on screen or said