@stroughtonsmith as someone else in your mentions said, it works reasonably well if you create a plan first of what you want to build, and when you would consider it finished (persisted to a Markdown file or whatever), and go back and forth with the agent, ask it for proposals, challenge its proposals.

then whenever you want to work on this project, you start with "let's work on item X from PLAN.md", and the LLM doesn't have to start from scratch again, and you've given it a more focused chunk of work to do.

you can also add pre-loaded files like AGENTS.md, and CLAUDE.md (to help you do boring admin stuff by just typing "commit and push" (and have it format, lint, cleanup, create commit messages, whatever, and give it some basic knowledge of the project and its architecture up front).

but for quick prototypes, asking for an app certainly is valid :) when you want to turn it into something real though, and want a lower slop ratio, and something roughly approximating how you would do it yourself, a plan helps.