This is exactly the gap I kept running into too. Bunpro’s cloze format is great for recall in context, but it’s a different skill from producing a full sentence with no cues at all.
Harley’s point about grading flexibility is spot on – there are usually multiple valid ways to express the same idea in Japanese, which makes automated checking really hard. AI grading helps a lot here because it can evaluate whether your sentence is natural and correct even if it’s not word-for-word identical to a model answer.
Full disclosure: I built JIVX for this specific use case – you get an English prompt, write the full Japanese sentence, and AI grades your attempt with explanations. It covers N5 through N1 with both polite and casual forms.
But even without a tool, you can practice this with any textbook: cover the Japanese, look at the English, and try to reconstruct it. I wrote about why this recognition-vs-production gap happens and a daily routine for closing it: Why You Can Read Japanese But Can't Write a Sentence (2026) | JIVX Blog