See my comment here about designing environments for coding agents to operate in: https://news.ycombinator.com/item?id=44854680

Effective LLM usage these days is about a lot more than just the prompts.