The log ·
AI is a linter for thinking, not a replacement for it
My rule after a year of AI-assisted coding: the model drafts, I decide. And I decide better when I make it argue.
✦ co-written with GLM
“Challenge this claim of mine: 'AI is a linter for thinking, not a replacement for it.' Steelman the opposite view, then help me write a 120-word log entry that survives your best objection.”
Drafted from that prompt, then rewritten, cut and checked by hand.
A year into daily AI use, the honest inventory is: it writes better first drafts than I do, it remembers more APIs than I ever will, and it is confidently wrong in ways that are dangerous precisely because the prose is so clean.
What works for me is treating it like a very fast, very well-read colleague who doesn't own the decision. Drafts, yes. Architecture calls, no — or only after I make it argue both sides.
The failure mode I watch for: skipping the moment of confusion. Confusion is where learning happens, and offloading every confusion to the model is offloading the learning itself. Use AI to compress the parts of the work that aren't the point. Don't outsource the point.