AI Features in an App That Never Calls an AI: The Prompt-Out, Paste-In Pattern

The goal

A meal-planning app with a strict promise: nothing leaves your phone. Users wanted to turn "a week of high-protein dinners for two" into a full plan without typing every meal. The goal was to give them a language model's help while keeping the promise exactly, with no key, no proxy, no server, and no recurring cost.

The approach

Turn the flow inside out. The app writes the prompt, the user runs it in whatever assistant they already use, and the app parses the answer defensively with a preview before anything is saved. Manual by design, and said so in the interface.

What I built

The prompt. Composed by the app with the exact schema it expects back, always in English regardless of the interface language, copied to the clipboard in one tap.

The import. A strict but forgiving parser, a preview screen, and an approve step. Model output is never trusted blindly, so a bad answer is a retry rather than a corrupted plan.

On-device intelligence where it fits. The same boundary later carried a receipt tracker: the operating system's on-device text recognition touches no network, so it is inside the rule. Line items are found by clustering text rows geometrically rather than by regex, and a reconciliation check, items minus discounts plus taxes equals the total, routes anything that does not add up to a review queue.

The result

The privacy label still says nothing is collected. There is no recurring cost, so a one-time price works. Users pick their own model, and the feature improves every time their model does, without an update from us.

Takeaways

• The boundary is about services, not intelligence. On-device frameworks are fine; hosted models are not.
• Put the schema in the prompt and a preview before the import. The human is the validator.
• Manual by design is honest. Show the seams instead of hiding them.