What does the next step need?
Model choice can look like a ranking problem. In practice, reviewing a difficult change and renaming a field are different jobs. The model I want for one may be unnecessarily slow for the other.
I start with the next step: read some unfamiliar code, make a small patch, explain a failure or review the result. That is a more useful brief than asking one model to carry an entire project.
Thinking and changing things
Before a broad edit, I want room to explore the code and compare approaches. Once the approach is clear, I want the edit itself to stay focused. Separating those stages makes it easier to notice when a task starts wandering.
Local models can handle plenty of the routine work. I use a hosted model when the extra capability seems likely to make a difference, then judge it by the result.
Know when to stop adding context
A growing conversation can become harder to work with. When I find myself going over the same material or correcting the same mistake, I take it as a sign to summarise what I have learned and narrow the next task.