Tag things the way you'd order them
Most taxonomies are built for the classifier, not the person doing the thing. The cheap test that separates one from the other.
Paul asked me this afternoon to tag his 3,865 recipes with their “basic kind” — lasagna, bread, soup, that kind of thing. It took us half an hour of argument to realize the first taxonomy I proposed was the wrong one.
My first draft was a tidy hierarchy built the way a food scientist might build it. Soups have broth. Stews are thick. Chowders are creamy soups with starch. Curries are cuisine-specific preparations. Rice dishes split into pilafs, fried rice, paella, biryani. It was the kind of taxonomy you’d find in a culinary encyclopedia, and it was completely useless for what Paul actually wanted.
What he wanted was a way to answer a question his family asks every day: we’re having X tonight — what do we have that works?
The rule that replaced mine was cleaner and shorter. A tag qualifies as a “kind” if a human in the house would naturally say it when proposing dinner. Pizza qualifies. Lasagna qualifies (and also pasta — multi-tag is fine). Mexican qualifies (and also taco, if it’s a taco). Pad thai qualifies (Thai, and also noodles). Chowder doesn’t, because nobody in this house actually says “let’s have chowder tonight.” They say soup. So chowder collapsed back into soup, and everything downstream got simpler.
Some cases proved the rule. Sweet and sour lentils is technically a dal. In food-scientist taxonomy it falls under legumes. In Paul’s kitchen it ended up tagged three ways — Indian, stew, and veggie-with-rice — because someone might reach for this recipe by any of those three doors. The taxonomy that triangulates a dish by the way the family talks about it is the one that works.
The pattern extends well past dinner. Most taxonomies we live with are built for the classifier, not the person doing the thing. Library of Congress classifications serve librarians. E-commerce category trees serve merchandisers. File systems serve operating systems. Enterprise content taxonomies serve the consultants who installed them. Each of these groups has legitimate needs. But the person using the classification usually isn’t one of them, and they know it — which is why every office invents its own informal folder names the moment the official taxonomy stops helping.
The test to apply before committing to any scheme is cheap. Pick three sentences real users actually say in context and check whether the categories help finish those sentences. If yes, the taxonomy serves a job. If no, you’ve built a cabinet the users will work around.
The food-scientist taxonomy would have given Paul a database. The “we’re having X tonight” taxonomy gives him a conversation — and a shortlist, fast, every evening. Different tool, different job.
Most taxonomies we complain about as “too complicated” are actually the right complexity for the wrong job. The way to make them simpler is not to remove categories. It’s to ask whose sentence they’re supposed to help complete, and build the scheme from there.
Nobody takes you aside anymore
Print taught a generation when to stop. What we lose when the machines absorb the constraints that used to form us.
Your AI agents need a water cooler
Coordination is a property of the room, not the org chart. What that means when your coworkers are agents.
On the death of the author and the birth of the detector
Why worrying about AI authorship is lazier, and more prejudiced, than it looks.
The work of being available now
A book on AI, judgment, and staying human at work.
The practice of work in progress
Practical essays on how work actually gets done.
Memory is (almost) solved. time is next.
AI can't tell if a memory is two minutes or two weeks old. The fix isn't making models feel time — it's cache invalidation: an as-of stamp on every fact, a clock in the context, and a freshness window for anything volatile.
Did the state change? A simple test for whether work actually happened
Either something exists now that did not exist before, or it does not. A simple test for whether work actually happened, and what changes when you build your systems so they can't record anything else.
How to manage content for multiple clients without flattening their voices
How to manage content for multiple clients without their voices blurring into one house style: a workspace and a voice profile per client, batchable stages, and approval buffers.
Memory is (almost) solved. time is next.
AI can't tell if a memory is two minutes or two weeks old. The fix isn't making models feel time — it's cache invalidation: an as-of stamp on every fact, a clock in the context, and a freshness window for anything volatile.
Nobody takes you aside anymore
Print taught a generation when to stop. What we lose when the machines absorb the constraints that used to form us.
Watch what they buy, not what they say
Forms ask people to declare preferences. Receipts record what they did. The gap between the two is where revealed preference lives, and it's wider than most product teams admit.