I’m building a small LLM-based text rephrase tool, but it turned into something different than I expected.

I started thinking the main challenge would be prompt design. Instead, I kept running into cases where the model output looked correct while subtly changing meaning or missing the requested tone. That became a bigger problem than generation itself.