The "fix" that wasn't a fix

A few weeks ago, while running my own German writing through German Writing Coach, I noticed something odd. I'd written the phrase "wie man seine Abgabe in OpenAI sehen soll" — a textbook subordinate clause, with the conjugated verb at the end where it belongs. The tool flagged it under "Falsche Verbposition im Nebensatz" (wrong verb position in subordinate clause), explained the rule, and offered a correction: 

wie man seine Abgabe in OpenAI sehen soll → wie man seine Abgabe in OpenAI sehen soll

The "before" and the "after" were word-for-word identical. The tool was confidently telling me to fix something that wasn't broken, and the proof of the fix was the same sentence I'd submitted.

For a learner, that's worse than no feedback. If you're not sure of your own German, you might trust the tool, hunt for a difference that doesn't exist, and walk away convinced you got something wrong. The whole point of a writing coach is to be a reliable second pair of eyes — and a coach who invents mistakes isn't reliable.

Why it happened

AI models, even careful ones, sometimes pattern-match a familiar grammar rule onto a sentence that already follows it. The model sees a subordinate clause, knows there's a rule about verb position in subordinate clauses, and feels obligated to comment on it. When asked to produce a correction, it can't actually find anything to change, so it just emits the same sentence twice. It's the writing-coach equivalent of a teacher who feels obligated to mark something wrong on every paper.

What changed

This update adds three layers of protection so that the feedback you receive is feedback you can act on:

  • Smarter "before / after" examples. Each feedback item now provides its before and after as separate pieces, and the tool double-checks that the two are actually different. If a "fix" turns out to be identical to the original, the feedback item is dropped before it reaches you. As a small visual bonus, the before is now shown struck through in red and the after in green, so the change is immediately obvious at a glance.
  • Honest Clean Entries. When your German text needs no grammar fixes — a Clean Entry — the tool no longer shows hard-grammar feedback (orthography, subordinate clause structure, verb-with-preposition patterns) on top of it. If something was a real grammar mistake, it would have appeared in the Minimal Fix; if Minimal Fix is identical to your original, by definition no grammar fix was needed. Style and idiomatic upgrades still show up, because clean writing can always be sharpened.
  • Less pressure to find errors. The tool used to be asked to produce three to five feedback items per session. That target is now zero to five — only when something genuinely warrants comment. If your writing is clean, you'll see fewer items, and that's a feature, not a bug.

What you'll notice

Most of the time, you won't notice anything different — well-written text that previously got real feedback will continue to get the same real feedback. What you should stop seeing is the occasional "fix this thing that doesn't need fixing" item. Clean Entries should feel cleaner: a green badge with no contradicting grammar feedback below it.

Genuine upgrade-level feedback — style, idiomatic phrasing, register, collocations — is unaffected. The goal isn't less feedback for its own sake; it's to make sure that every piece of feedback you receive is feedback you can trust.

Why this matters for exam prep

For learners preparing for Goethe-Institut, TestDaF, or other German writing exams, confidence in your own writing is one of the most underrated skills. Knowing when you've got something right matters as much as knowing when you've got something wrong. False-positive feedback erodes that confidence. By tightening the filter, the hope is that every "this needs work" message in the tool actually means "this needs work."

Try it out

The improvements are live now on the Coach page. If you've had a session where the feedback felt off, try it again. As always, the tool is free and the source code is open. If you spot a feedback item that still doesn't pass the smell test, I'd love to hear about it.