I think this is a neat approach. When I interact with AI tooling, such as Claude Code, my general philosophy has been to maintain a strong opinion about what it is that I actually want to build. I usually have some system design done or some picture that I've drawn to make sure that I can keep it straight throughout a given session. Without that core conception of what needs to be done, it's a little too easy for an LLM to run off the rails.
This dialogue-based path is a cool way to interact with an existing codebase (and I'm a big proponent of writing and rewriting). At the very least you're made to actually think through the implications of what needs to be done and how it will play with the rest of the application.
How well do you find that this approach handles the long tail of little things that need to be corrected before finally merging? Does this approach solve the fiddly stylistic errors that need to be made on its own, or is it more that the UI / PR review approach that you've taken is more ergonomic for solving them?
hey! that's awesome to hear, thanks for the feedback.
we've tried a lot of things to make code more in-line with our paradigms (initially tried a few agents to parse out "project rules" from existing code, then used that in the system prompt), but have found that the agents tend to go off-track regardless. the highest leverage has just been changing the model (Claude writes code a certain way which we tend to prefer, vs GPT, etc) and a few strong system prompts (NEVER WRITE COMMENTS, repeated twice).
so the questions here are less about that, but more about overall functional / system requirements, and acknowledging that for stylistic things, the user will still have to review.