Most LLMs still play autocomplete sidekick. But seasoned devs? They get better results when the model reads and rewrites actual source files.
That gap—between how LLMs are designed to work and how pros actually use them—messes with survey data and muddies the picture on real gains in code quality and output.
