Dutch-Language Bug: A 15k-Char Prompt Made Devs Blame the Model for Their Own Mistake

Initial_Orange2985 · reddit · 2026-11-08

A developer building a memory feature for a company AI saw French questions get Dutch answers and nearly swapped out gpt-oss-120b. The culprit was their own prompt: a language-detection instruction that counted small words — "de" twice, "en" once — and issued an absolute rule to answer in Dutch. Fix: more French cue words plus a rule forbidding language switches when é/à/ç appear. With the prompt now 15k characters of stacked instructions, the author asks: do you log the full final prompt for every request to catch this kind of bug?

Original post →

More from coding & agent

coding & agent channel →