I was listening to Claire Vo interview Eddie Kim this morning. Eddie is the CTO and co-founder of Gusto, and somewhere in the conversation he mentioned that the first cut of a new product got built during an unexpected layover. Claire’s read on it: this keeps happening. Someone steps out of their normal week, maybe on vacation or parental leave, and finally builds the thing they’d been meaning to build.
Eddie got stuck with a five-hour layover in London, and instead of stewing about it he opened his laptop and started prototyping an idea he’d been chewing on, in Claude Code. By the time he landed in San Francisco he had a working prototype. That prototype became Gusto Co-founder, a new product line the team shipped in about ten weeks with five people, inside an R&D organization of more than a thousand.
Everyone’s going to read this as a story about Claude Code, and the tool is real. But Claude Code was sitting on his laptop for months before London. What he didn’t have on a normal Tuesday was five uninterrupted hours and a problem that had been quietly turning over in the back of his head.
AI collapsed the cost of building the thing. It did nothing for the harder job, which is working out what the thing should be.
I get the long version of this every summer. For about five years now we’ve taken roughly two weeks off every August. I’ll check Slack here and there, but we’re mostly too busy to work, and that stretch of being disconnected is reliably when the things I’d been half-thinking about start to connect.
Where the ideas actually come from
I’ve always been interested in where ideas come from, and how you generate more of them and better ones. The study I keep coming back to is from Jonathan Schooler’s group at UC Santa Barbara. They recruited 45 physicists and 53 professional writers, people whose paychecks depend on new ideas, and emailed them every evening for two weeks with the same questions: what was your most creative idea today, and what were you doing at the time? Then they ran the whole thing again with 87 more people.
About 20% of the ideas arrived while people were doing something unrelated to work, and that held in both studies (19.9%, then 19.2%). The number that made me sit up is the split by how stuck people were. On problems where someone had been at an impasse, 26% of their ideas came during mind-wandering. On problems where they’d been making steady progress, 14% did. The second study found the same pattern, 20% against 9%. Roughly double, both times. Grinding works fine when the work is moving. When it’s stuck, the ideas start showing up in the shower.
Two honest caveats from the same paper. The mind-wandering ideas were rated no more creative or important than the ones from the desk. And when the researchers had everyone re-rate their old ideas three to six months later, the ideas that had felt like “aha!” moments lost more of their shine than the ones that hadn’t. The eureka feeling turns out to be a bad quality signal. It tells you an idea arrived sideways, and nothing about whether it’s good.
The mechanism people point to is the default mode network, the set of brain regions that ramps up exactly when you stop focusing on the outside world. My favorite fact from that literature: a hard, focused mental task raises the brain’s energy consumption by less than 5% over its resting baseline. The brain has no idle setting. Take away the task and it keeps running at nearly full draw, on whatever you left open.
Schooler’s whole apparatus, by the way, was one email a night. You could run it on yourself: one line each evening, best idea of the day and what you were doing when it hit. Two weeks of entries and you have your own base rates. I already know mine would cluster on walks.
The walking part
I try to get out and exercise a few days a week. Sometimes a three-mile run, sometimes a backyard workout, sometimes just a walk. It ends up being a good time to let things fold over in my head.
The research backs the anecdote harder than I expected. In a 2014 Stanford study, Marily Oppezzo and Daniel Schwartz measured creative output while people sat and while they walked. Walking raised it by about 60% on average, and 81% of participants scored higher on the divergent-thinking test. It held on a treadmill facing a blank wall, so it wasn’t the scenery, and the boost lingered a little after people sat back down. The catch: walking helped the generate-lots-of-options kind of thinking and did nothing for the single-right-answer kind, where only 23% improved.
There’s a boundary condition worth knowing, because it tells you which activities to pick. A 2022 study out of UVA and Minnesota had people do a creativity task, then watch either a boring video or a moderately engaging one. Mind-wandering produced more creative ideas only during the engaging video; drifting through the boring one bought people nothing extra. The activity has to occupy your hands or your feet a little without taking your whole head. A walk or the dishes qualify. Scrolling fails the test from the other side, because it takes all of your head.
Leave it half-finished on purpose
The move I’ve been trying to steal comes from Hemingway. In A Moveable Feast: “I always worked until I had something done and I always stopped when I knew what was going to happen next. That way I could be sure of going on the next day.” He quit each day mid-momentum, while he still knew the next sentence, so he never faced a blank page cold.
The second half of the trick gets quoted less and matters more: “I learned not to think about anything that I was writing from the time I stopped writing until I started again the next day. That way my subconscious would be working on it.” Stopping while you’re ahead is the easy half. Staying stopped, and trusting the background process to run, is the half he said he had to learn.
That instinct has a name in psychology, the Zeigarnik effect, after Bluma Zeigarnik, who studied why a waiter could recall the details of an unpaid order and forget them the second the bill was settled. The tidy version, that we simply remember unfinished tasks better, hasn’t held up; a 2025 meta-analysis found no reliable memory advantage. What survived is the pull to go back and finish. An open loop nags at you. Leave the loop open on purpose and the nagging is free compute.
The faster the tools get at building, the more the binding constraint slides upstream, to the part that was always hard: figuring out what’s worth building and whether you’ve got the shape right.