This is how I've been building new things with LLMs. I start with this structure here. There is a simple Agents file that points to the main Shaping Approach. There isn't much infrastructure here — only about 500 words in total (outside of this README). I've found the less infrastructure I give to LLMs, the better they work for me.
I fill in my Idea, and then I start figuring it out. I call this process shaping — an idea I learned from Ryan Singer. Ryan's process is extremely rigorous and specific — mine is extremely lightweight and fungible. Here is an example Idea I wrote for one of my projects.
I ask the LLMs: what do you think of my idea? I have a discussion and try to flesh out the unknowns. What's possible? What have I missed? Does this make sense? If I want to take a deep dive into a question (e.g. what APIs exist for this data type?), I ask the agent to create a report.
Then I come up with some Design Priorities. What are the guiding priciples of this product? Does it need to be fast and optimized for slow internet connections? Built for an expert audience? Always have a light/dark mode? Designed to scale to hundreds of millions of users?
Then I ask the LLMs to guide me on a Stack for the product. You can tell it what your technical preferences are, and with the context of the idea and the design priorities, it can suggest a good stack.
Finally: The Plan. With an idea, design priorities, and stack, the LLM can now make a plan to build the product.
I like to start with a few prototypes, they help shape the plan in a more useful direction. LLMs are little context goblins. This document structure and the prototypes — they serve as context fuel for it to build the real thing.