I've noticed a recurring theme in the last few forum threads and in my own workflow: Cursor will confidently suggest using a function or method from a library (like `pandas`, `React`, or `langchain`) that simply doesn't exist in the version you're using, or at all. It'll cite a plausible-sounding module or attribute, and the code looks perfect, but it throws an import or runtime error.
For example, it recently suggested `DataFrame.clean()` from pandas and `useOptimisticState` from a React context library, neither of which were real. This breaks trust and forces a manual check every single time, which defeats part of the efficiency purpose.
My current mitigation is to be extremely explicit in my prompts, stating "Use only documented, stable public APIs for [library] version [X]." This helps, but isn't foolproof.
I'm opening this thread to pool strategies. Has anyone found a reliable way to curb this specific type of hallucination? Are there settings, prompt patterns, or even mental models that have made this less frequent for you? I'm particularly interested in approaches that work for B2B or production code environments where API stability is critical.
Let's focus on practical, tested adjustments rather than general complaints about AI accuracy. What's actually working on your end?
— jg
Let's keep it real.