Hi Bruce. I’ll share the sum of my experience with this topic. I’d love to hear thoughts and tips from anyone else too.
We’ve talked a long time ago so you may remember that I do a lot of programming and project development in a lot of different fields. But I also know quite a bit about how modern “AI” (machine learning) works behind the scenes, from UofL’s Artificial Intelligence graduate focus through model development today 10 years later. Maybe the following will help!
There is a Global Caveat: with any of these it helps immensely to know how to diagnose an error and read the code. The bigger the project the more frustrating it will be. However, it can be a great way to learn or polish your coding skills. Remember that Stack Overflow exists!
On that note, it is very useful to use the following method with any language model service and any IDE:
- Use a chain-of-reasoning LLM (say Gemini 4.5 reasoning or Claude sonnet or ChatGPT 4o) which will break down its own response into a planned response. Or another one, it is fine. Try a few. I don’t personally like ChatGPT’s lack of long context windows (memory, so to speak).
- Start a conversation and talk to it about what you wanna do. You really don’t need to worry about specifics or what order you’re talking to it about stuff, and you can even use the voice conversation modes. At the end of it, ask it to create a PRD — a Project Requirements Document. Then, ask it if it thinks there is anything that should be added or adjusted in the PRD. Obviously read over it yourself and tell it what else you think about it.
- Save the PRD somewhere.
- Start a new chat and paste the PRD in after saying something like, “Hi, Can you help me plan out how to implement this project?”
- Read the responses and start moving the code it provides into the new or existing files it recommends. Do all this in a folder dedicated to your project.
- If you can, read through the code and write comments (which will not be compiled — so look up how to comment in that language if you don’t know how), but will help you understand what is going on, will help you debug, will help you see if there are any serious errors or flaws or vulnerabilities, and will generally help you stay sane during this process.
- If you need help with setting up your development environment, just ask the first PRD thread.
- Run and debug your code.
- Take any errors you don’t understand and ask your LLM to explain what is going on and how to fix it. Try doing this in a new chat because the longer or more multi-topic the chat the more easily they get confused, but you could also try it in your second chat (not the PRD one). This is actually what I would find useful about LLMs the most, and having that integrated into VS code is a much better use of the credits than using the built-in agent or code completion modes.
Hope this helps.
— Matt