I almost never tell it to just write me a thing (what I think of as vibe coding). (2%)
I sometimes write a pretty detailed doc or spec; have the AI draft an implementation; then review and fix it myself. I try to keep this to “reasonable PR” size, a few hundred lines (a module or two) max, and will do a few rounds per hour. (~25%)
I will often stub out modules or classes (sometimes with docstrings) and tab-complete big chunks of them. (And then turn tab completion off and rage-code the rest by hand because the AI is so far off base.) (~25%)
I will often tell the AI to write tests for stubbed methods prior to implementation. I then double check the tests before moving on to manual or AI-assisted implementation. This is usually in increments of a single AI request/response. (~35%)
I will occasionally ask the AI to change existing code and tests, usually in a single request/response. I’ve had very mixed results with this. (~10%)
I have been finding myself writing code in smaller standalone libraries and then assembling those into larger and larger composites so that each library is a size a model can more realistically reason about; and for the layers on top of it the AI wont fill its context up reading all that source instead of just the public API docs.
I sometimes write a pretty detailed doc or spec; have the AI draft an implementation; then review and fix it myself. I try to keep this to “reasonable PR” size, a few hundred lines (a module or two) max, and will do a few rounds per hour. (~25%)
I will often stub out modules or classes (sometimes with docstrings) and tab-complete big chunks of them. (And then turn tab completion off and rage-code the rest by hand because the AI is so far off base.) (~25%)
I will often tell the AI to write tests for stubbed methods prior to implementation. I then double check the tests before moving on to manual or AI-assisted implementation. This is usually in increments of a single AI request/response. (~35%)
I will occasionally ask the AI to change existing code and tests, usually in a single request/response. I’ve had very mixed results with this. (~10%)
I have been finding myself writing code in smaller standalone libraries and then assembling those into larger and larger composites so that each library is a size a model can more realistically reason about; and for the layers on top of it the AI wont fill its context up reading all that source instead of just the public API docs.