Can a prompt replace Pestle?

Build a recipe extractor for sources that do not publish structured data.

The input is a URL or a block of pasted text. Try three strategies in order and record which one produced each recipe.

First, schema.org `Recipe` JSON-LD, which most large recipe sites publish and which needs no model at all. Second, a readability pass to strip the page to its main content, then a language model with a strict JSON schema — title, yield, ingredients as an array of `{quantity, unit, item, raw}`, steps as an ordered array. Third, for pasted captions and transcripts, the model directly on the text.

Model output is untrusted by construction. Show the extracted recipe beside the source text in a review screen and require a confirmation before it enters the library. Highlight anything with low confidence: an ingredient with no quantity, a step under five words, a yield the model invented. The failure mode here is not a missing recipe, it is a plausible recipe with the salt quantity wrong, and only a human catches that.

Always store the source: URL, the raw text, and the capture date. When an extraction turns out wrong six months later, the original is what lets you fix it.

Add a share target so the browser on a phone can send a page directly into the tool — a progressive web app can register as a share target and it gets you most of the way to the native capture experience.

Store recipes as Markdown with YAML front matter, indexed in SQLite.

Do not download video or transcribe it automatically. Platform terms forbid it, the pipeline breaks constantly, and pasting a caption takes four seconds.