Can a prompt replace Crouton?

Build a local recipe manager where import quality is the thing you optimise.

Import from a URL. The right approach is not scraping layouts: most recipe sites publish schema.org `Recipe` structured data as JSON-LD, and reading that gives you title, ingredients, steps, yield, times and an image, correctly, from thousands of sites at once. Fall back to a readability pass plus a language model only when the structured data is absent, and mark those imports as needing review.

Parse ingredient lines into quantity, unit and item, keeping the original string alongside. Scaling a recipe is the feature this unlocks and the parse must be conservative: a line it cannot parse stays as text and simply does not scale, rather than being scaled wrongly. "Salt to taste" multiplied by three is the failure people notice.

The cooking view is the product. Steps one at a time in large type, a wake lock so the screen stays on, a scale control at the top, and any duration mentioned in a step turned into a tappable timer. Nothing else on the screen.

Store recipes as Markdown files with YAML front matter in a folder, indexed into SQLite for search. Files first means your recipes outlive the app, sync through any folder-syncing tool you already use, and diff in git.

Add full-text search across ingredients, so "what can I make with the aubergine" is answerable.

Do not build a social feed or public recipe sharing. Export a recipe as a file or a printable page and stop there.