I like to call this way of working “Developer Mode”, it’s probably not ideal for “non-coders”.
First, I try to launch new chats more often. Unless I’m extending the functionality of something very specific, having a clean context helps Agent get it right.
Also, I have the feeling that with the first message, he does a deeper planning, and the trick I’ll discuss now, depends on this initial proposal of the plan.
The main trick is to ask Agent to include the code implementation and the SQL queries he’s going to use to develop the functionality.
Normally, he’ll ignore this part of the initial instruction, but if that happens, don’t accept the plan and insist that the code and queries are missing from the plan.
Eventually he will do it, if you’re a developer you’ll be able to evaluate if the proposal makes sense, in any case, having this, you can now accept the plan and with all the previous context: the functional definition that Agent makes of the plan and the code examples, he’ll make a better implementation.
At least that’s what I’m finding, maybe it’s just a coincidence.
Try it and tell me
Example:
