Translation apps were never designed to help you learn a language. A dictionary gives you some analogous words; a chatbot invents a “right” answer. But none of it sticks, because those apps don't take your situation and what you're trying to accomplish into account. The app I'm working on actually will.
An honest, live answer to one question: what does it think you want? When the read is wrong, that's your cue to add or change input. No prompts to write—the bubble is the conversation.
Each emission is a structured artifact in a known category—dictionary, grammar, vocab, OCR, usage, composition. They carry a lifecycle, so the stream stays legible.
Pinning is stronger than starring. A pinned gloss zips to the top and becomes accepted context: it keeps 特定記録 in mind even after you rewrite the composer completely.
Capture → interpret → emit → pin, again and again. The pinned set becomes the session's working answer. And if you ever wonder “where did I first see that word?”, you can play back the exact timeline of text and images you fed it to discover it.