There are quite a lot of pitfalls in using AI to help with coding. Hence, summarize and reference other users’ experiences.

Overall, as mentioned in previous blog, detailed markdown file to describe how your construct the codebase, including what classes, instance, methods in designing is the best approach to work with AI.

1. Prompt with high level of Details
2. Iterative prompts regurgitate to devise into one-shot prompt
3. Complex codebase need very simple prompts step by step, spot obvious mistakes, read every lines of codes built by AI, thought process is needed
4. Advanced reasoning models to write detailed spec, in which there are each blocks broken down to be fed to AI
5. Biggest! in build-up codebase, trying to ask AI to fix a small error could be impossible because there are memorized errors in LLM context, there are so many constraints that also overwhelms LLM. How to cope? Start a new cascade, a new prompt, start from the spec, ensure details if anything wrong, ask new conversation.

Before proceeding to a large complex codebase construction, dedicated to build small simple modules one by one to build up knowledge. Yes, with AI tool, I realize it’s still vital to build up knowledge piece by piece using AI. there is always THE RIGHT WAY to leverage AI!