Can a prompt replace Rosetta Stone?

Build a picture-association drill for vocabulary, and treat image selection as the hard part rather than an implementation detail.

An item is one image, one target-language word or phrase, and optional audio. Sessions show four images and play or print one phrase; you pick the matching image. Reverse the direction on later reviews: one image, four phrases.

The design problem is ambiguity. An image meant to teach "running" that shows a person on a beach at sunset teaches four things at once. So build an authoring check into the import step: for each candidate item, the tool shows the image with the three distractors it would pair it with and asks you to confirm that only one is right. Items you cannot confirm go to a review pile instead of into the deck.

Source images from an openly licensed set — Openverse and Wikimedia Commons both allow it — and store the licence and attribution with every item. Print an attributions page from that data, because a deck you built from other people's photographs needs to carry its credits.

Schedule with FSRS over items.

For the speaking half, record yourself and play it back beside a reference recording if you have one. Do not score it. Accepting or rejecting learner pronunciation needs a model trained for exactly that, and a homemade threshold over a generic recogniser produces a number that means nothing — say so in the interface.