Imagine an AI assistant that not only writes code but also debugs it, even before you run it. Researchers are exploring this self-refinement concept by prompting Large Language Models (LLMs) with targeted verification questions (VQs). Think of it like an internal code review, where the AI challenges its own work. This research tackles common LLM coding errors like 'hallucinated objects' (where the AI invents non-existent functions) and 'wrong attributes' (incorrect use of object properties). The process involves converting the code into an Abstract Syntax Tree (AST), a structured representation that allows the AI to pinpoint potential problem areas. Specific VQs are then generated, targeting these nodes in the AST. For example, if the AI uses a function it hasn't defined, the VQ might be, "Is this function already defined? If not, provide an implementation." The AI then uses these VQs to revise its initial code, aiming to fix the identified issues. Experiments using the CoderEval dataset show promising results. This method significantly reduced specific error types, improving the chances of generating runnable code. While the AI still introduced some new bugs during the process, the targeted approach led to fewer errors than using general verification questions or no questions at all. This research opens exciting possibilities for more reliable and autonomous AI coding assistants. Imagine a future where AI not only generates code but also ensures its quality, saving developers valuable time and effort. However, challenges remain, such as refining the VQ generation process and ensuring the AI doesn't over-correct and introduce new errors while fixing existing ones. The next step is to expand this technique to other programming languages and error types, paving the way for truly self-improving AI programmers.