The "going in circles" problem several people mention seems like the key thing to watch for. In my experience it tends to happen when the codebase outgrows what the model can hold in context effectively.

What's worked better for me: treating it like onboarding a contractor. Very specific, bounded tasks with clear acceptance criteria. The moment you're spending more time explaining context than it would take to just write the code yourself, that's the signal to switch back.

I do the same. Directed tasks with smaller context and start a new "chat" when it's done what I want.