Generating code with Large Language Models (LLMs) can be like giving someone vague directions – you're never quite sure what you'll get. A new research project called "Sketch Then Generate" aims to fix this by letting users give LLMs real-time feedback as they write their prompts. Imagine sketching out the structure of your code in plain language, and having the LLM respond instantly with suggestions. This "language-oriented code sketching" uses natural language processing (NLP) to turn your prompt into an incomplete code outline, or sketch. As you type, the system maps your words to code elements, assembling them based on the relationships between the words. For example, "create a Python function to read a file" would be mapped to function definition and file I/O operations. The system then suggests relevant code snippets, and you can choose the ones that fit your intent. This interactive process not only gives you a preview of the code structure but also guides the LLM toward generating the code you actually want. The research team tested this approach with a prototype and found that it reduced the cognitive load on users, making it easier to control the LLM's output. Even when the initial sketch wasn't perfect, users appreciated the ability to see how the LLM interpreted their prompt and correct it along the way. This approach is especially helpful when you have a general idea of what you want your code to do but are unsure of the specific implementation details. While the current system is limited to Python and English, the researchers plan to expand it to other languages and explore how this "sketch-then-generate" concept could be applied to other creative fields like graphic design. This research points towards a future where interacting with LLMs feels less like guesswork and more like a collaborative partnership, bridging the gap between human intention and AI generation.