Transcribe, Don’t Answer: A Common LLM Misfire
Sometimes, when a large language model is used to improve the text, it can get a bit confused. Depending on what you're transcribing, it sometimes thinks you're asking it a question and responds instead of just rephrasing. That's not what we're aiming for! We're just trying to get the question down. It seems to think we're directing the question to the model itself, not just telling it to transcribe. There should be a way to fix that, maybe with a custom instruction that makes it clearer, like "rephrase this text." I might try that myself too. Anyway, this is one challenge you might run into! 🤔✍️,
Log in to comment and vote
Comments3
Haraanganjot Singh
Nov 7, 2025
Hey there! 😊 So, I’m pretty sure that doesn’t happen to everyone. I tinkered a bit with your system prompt! 🤓 The one I was using basically said, “Make this text sound better by polishing up grammar and punctuation. Anyway, I made sure to specify in the prompt to just rephrase the text instead of answering it, and voila! Problem solved! 🙌 But honestly, I don’t think everyone’s experiencing this—maybe it was just my mistake for not realizing I was on a different prompt from the usual one. Oops! 😅
BlabbyAI Dev
Nov 10, 2025
Yeah, it doesn't always happen, but your prompt gets wrapped with another prompt in our backend that tells it not to answer questions and to only respond with the requested text. We should probably fix that prompt. I’m also thinking we could let advanced users like you create your own prompts without the wrapper on our servers.
What do you think? Would you want this as a feature? I’m thinking it’ll just be a small button called "Advance," and then you can enter your own prompt without any extra wrappers from our end.
Haraanganjot Singh
May 23
Sorry for the late reply, I just got around to reading your message.
I honestly think the feature of letting users create their own prompts without the backend wrapper could confuse most normal users. Many people probably would not fully understand why the feature exists.
I think the backend prompt is important, users could accidentally mess up the original behavior of the AI. Instead of rephrasing text properly and correcting grammar, the AI might start directly answering questions.
I think keeping the wrapper enabled by default is definitely the safer and more user friendly approach.
Instead of completely removing the backend wrapper, maybe the system prompt could support a special instruction from the user. For example, if the sentence ends with something like “answer this,” then the AI could switch behavior and directly answer the question instead of rephrasing it.
It would also avoid the risk of people accidentally breaking the intended behavior by editing prompts themselves.