I don’t pretend to be a spectacular writer. Haven’t published anything substantial in my life, and probably won’t ever. Hell, I don’t even post much on here. But I always try my best, and if I may indulge in self-assessment, I’ve gotta be somewhere above the fiftieth percentile in terms of writing quality. (If not, I’m in deep trouble; “thinking begins with writing” is my second-favourite quote of all time.)
LLMs don’t write on my behalf; every word on this blog is either a direct quote or direct from my brain to your eyes (or ears, if you’re using a screen reader). This is for a couple reasons, chief among them that my writing style tends to be dense and academic, I like it that way, and on the other hand I haven’t much liked LLMs’ historical attempts in that register. Now, I use LLMs all the time for other things, including as a kind of hyper-Oblique Strategy. One of my favourite tricks is to ask Claude to bridge two completely unrelated ideas as seamlessly as possible. Sometimes I’ll do philosophy (tell me what Lauren Berlant and St. Augustine might say about corporate governance), or maybe I’ll insist that it play-act as a logician. The results aren’t always usable, but it’s fun to see what you’ll get and instructive to predict the directionality of the response. And sometimes the connexion point Claude identifies turns out to have some novel intellectual currency or orthopraxic value. But the point is (and my line has always been) that I don’t let LLMs write for me, and typically I keep my longer-form text out of the prompt window entirely.
I gave up on that tonight by feeding a mostly-finished Markdown file into Claude Sonnet 4, asking it to critique the quality of my ideas and to pick five sentences to rework in depth. Call it degenerate curiosity. The response was witheringly unflowery, thanks to the killer system prompt I’ve been using lately. But the heartbreaker was that Claude’s suggested fixes were significantly better than what I would’ve dreamed up after twenty minutes of pacing and head-scratching and backspacing. (No examples, sorry; don’t want to spoil my forthcoming work!) I didn’t end up incorporating them whole cloth into my post — still too much of an AI smell for my liking, plus my authorial voice pathologically refuses to be diluted — so I’m still the dominant creative force behind the post. But to the extent that editing sharpens writing, I can’t unthink that it’s doing a better job than I could have.
Finally, as a second-order metacognitive experiment, I ran this post into Claude too. “It models the distinction it advocates”, quoth the machine. Hey, I’ll take it.