Sorry to be that guy, but you're using it wrong. The best flows right now are architect -> act -> test. First you have a session in "architect" / "plan" mode (depending on your ide/tool) where you discuss, ask questions, etc. Then, when everything is clear in "chat" mode, you ask the model to make a plan. You verify the plan, and then you tell it to start implementing it. You still get to approve tools, calls, tests, etc. You can also provide feedback on the way if you missed something (i.e. use uv instead of pip, etc).
Coding in a chat interface, and expecting the same results as with dedicated tools is ... 1-1.5 years old at this point. It might work, but your results will be subpar.
Nah it's good thanks for your input. I saw people use plan.md and todo.md and ide/commandline for this before. manus.ai demonstrates this via its chat interface as well.
Coding in a chat interface, and expecting the same results as with dedicated tools is ... 1-1.5 years old at this point. It might work, but your results will be subpar.