latest / Cursor and the autocomplete tax

I switched to Cursor late last year and spent a month convinced it had made me faster. Then I tried to recall what I’d actually built that month and the answer was, uh, a lot of CRUD with very confident variable names.

The thing nobody mentions about aggressive autocomplete is the tax it puts on your attention. Every suggestion is a tiny interrupt. You stop forming the thought and start evaluating someone else’s guess at the thought. Sometimes the guess is great. Sometimes it’s plausible nonsense that compiles. Either way you paid the context-switch.

The fix, for me, was turning Tab completions off by default and only flipping them on for the kinds of tasks where I genuinely want a parrot — boilerplate, test scaffolding, the fourteenth nearly-identical migration. For actual thinking work, the inline ghost text is the enemy of finishing the sentence in my own head.

The chat side I kept on. Conversational AI doesn’t interrupt you; you go to it. That asymmetry turns out to matter more than the raw capability of either mode.

None of this is an argument against the tools — they’re plainly useful. It’s an argument for treating attention as the scarce resource and configuring accordingly. If a feature is firing fifty times a minute, it had better be earning its keep fifty times a minute. Most of mine weren’t.