Essays, apps, and agents — because words are magic.
I obsess over AI and work at Every, make things like Tastemaker, and spend too much time teaching machines to sound like me. Mostly I'm trying to figure out how technology changes how work feels — by sharing how it's changing how it feels to me.
I write, test, and build with AI, then report what works for me so others can decide what works for them.
My personal Substack. Longer, weirder, more confessional. The stuff that doesn't fit anywhere else.
An AI writing tool that learns what good writing is to you, then teaches it back to you.
tastemaker.botA secular Bible study — reading scripture as literature, history, and weird ancient storytelling.
Explore the atlasCurrent fixations
A running list of the questions that keep leaking out of what I'm publishing, testing, and building.
I keep turning coding agents into writing rooms, research assistants, app builders, and knowledge-work operating systems. The live question is how much access makes them genuinely useful before the doors I opened need better locks.
I can send a draft through a small fleet of AI reviewers before a human ever sees it. The live question is whether they catch problems that improve the work or merely produce an impressive pile of feedback around the same old draft.
AI returned me to a kind of intellectual aliveness I thought I'd lost, then made it possible to neglect the sleep, meals, friendships, dog walks, and scripture that keep that aliveness attached to a human body.
“I made an app where there never was an app—but I left the app unfinished.”
“AI is our era's timid scribe.”
“I hadn't realized I was following this playbook until I saw it reflected back to me.”
“It's 6PM on a Friday night. Do you know where your childfree single thirtysomething is?”
“I used to be physically unable to open my email. Not in an 'ugh, I don't want to' way—in a 'my animal brain thinks there's a predator in there waiting to kill me' way.”
I wrote about why the most useful file in my AI writing system is a list of ways I disappoint my editors.
Read on Every →Every time I thought my writing benchmark was ready, I found another methodological problem that made the results impossible to trust, and started over. Building the test was harder than testing the models.
On AI, Magnifica Humanitas, and protecting the person the machine helped bring back
Read on Substack →I gave AI agents a door into Tastemaker, then learned the door had more access than it should have. The connection is offline until its data-access design gets a thorough technical review. Turns out you check the locks before you invite the agent in.
Opus 4.8 is the most complete model Every has tested yet: a jump back to the top on senior-engineering and writing benchmarks, stronger for everyday knowledge work, and good enough to make Claude tempting again—even if the app still lags the model.
Read on Every →