The useful change is context, not personality. A model that can see the repo can rename a symbol across files and not invent a second User type. A model that only sees the open buffer still writes the function you already had in another folder.
Boilerplate is faster: JSON types, test stubs, CSS that matches the file next door. That is real time. It is not 10× on a design review or on a production incident.
What still fails: business rules that live in a Slack thread, architecture that needs a no, and any API the training set never saw. If you cannot explain the change in a sentence, do not let the model merge it.
Treat the tool as a junior who types quickly. You still read the diff. You still run the tests. The teams that get hurt are the ones that skip those two steps because the output looks finished.