Out of Your Head, Into the File
Working with agents means articulating your taste.
There’s a new kind of file popping up in design and PM workflows this year. They’re markdown files like SKILL.md, DESIGN.md, and SOUL.md. But while we can read them, we’re not their primary audience. What goes inside is what we used to keep in our heads and call taste.
In a piece on X called “Design is the work,” Jake Albaugh argues that design is the upstream act of intent, the work of figuring out what something should be before anyone makes it, and that AI cannot do that part. Then in the closer, he tells you he used AI to draft the essay. “The structure, the sentences, a lot of the phrasing — generated. But the argument existed before any of it.” The judgment was his; the prose wasn’t. He had already decided what “good” looked like before the model wrote a word. Albaugh’s judgment lived in his head. The next step is writing it down where something else can read it. The artifacts are appearing at every scale of design and PM work.
At its smallest, the file holds a single rule, like Emil Kowalski’s skill files for animation behavior at Linear: two transitions side by side, one scaling from scale(0) and one from scale(0.95). The right one feels right because a deflated balloon never disappears. Kowalski’s claim: “Almost every ‘taste’ decision has a logical reason if you look close enough.” Explain the reason, hand the rule to the agent.
A whole design system gets the same treatment in PJ Onori’s A/B testing harness, which measures everything from timing and lines of code to accessibility and component usage. To find out whether his components or his documentation were doing the work, he ran tests without documentation. He put it bluntly: “The documentation was clearly the heavy lifter.” His new “For agents” section in the docs is the dumpster, in his words, for getting it into the agent’s silicon head, walled off so the human reader isn’t asked to reread the same point six times.
The artifact does different work when the agent isn’t applying your judgment but standing in for it. Tommy Geoco’s agent harness is seven markdown files, an Obsidian vault, ninety days of work, and $13,100. SOUL.md holds his voice and judgment. Anyone can rent the same Claude model he uses; what makes the agent his is that file. “You can’t control the quality of the model, but you can control the quality of the system.” Anton Sten had a different concept: five Claude Code instances on a Mac mini, named after Suits characters. Harvey handles contracts and pricing. Donna drafts the follow-up emails. Mike stores what Sten would otherwise forget. Louis worries about money. Wendy reads the others’ logs and flags where they’re slipping. The architecture is markdown. Each character is a few hundred lines of plain text describing one job, written in the language Sten actually works in.
At the strategic scale, the file is something the human can’t write alone, and the agent has to extract it. Marcus Moretti’s agent-native PM guide describes an agent that interviews him to draft the strategy doc, pushing back on vague answers: “Whose situation specifically? What do they try today, and why doesn’t it work?” A PM who can’t answer ends up with a STRATEGY.md full of confident-sounding nonsense. The agent isn’t writing the taste; it’s compelling the human to put it into language. Moretti’s line for it—”The conversation is the work”—sounds glib until you watch the agent force the answer out of him.
There’s a complication in Alex Dapunt’s piece on user research. Bain consultants are trained in what they call “answer-first”: state the conclusion now, justify it later. Executives who carry that habit into research interviews produce articulate wrong answers. The clarity is what you should distrust most. A tidy STRATEGY.md can sound exactly like a true one. Someone still has to know which is which.
The homework I’m taking out of the week: write down the rules my taste operates on—the ones I reach for when reviewing a screen and telling a designer “this isn’t quite right yet”—and make sure they’re rules I can defend, not just ones I can articulate.
What I’m Consuming
Flickr: The First and Last Great Photo Platform. Brett Weinstein makes the case for Flickr Pro at $82/year. The platform’s strength is what it refuses to become: another social network chasing trends, or an AI training pipeline. SmugMug bought Flickr in 2018. Their CEO Don MacAskill says they’re trying to build it sustainably for 100+ years. Weinstein’s piece is an unabashed defense of the early web’s photo-sharing ethos, with chronological streams, robust groups, full EXIF data, and photographers keeping their copyrights. (Brett Weinstein / PetaPixel)
How Letterforms Carry Culture. Brooklyn-based designer Daniel Irizarry built three display typefaces—Borinqueneers, Piragua, Callejón—rooted in Puerto Rican and Puerto Rican-American history. The Borinqueneers face honors the 65th Infantry Regiment, where three of his mother’s uncles served. Amelia Nash’s interview pulls out the key tension: typography as cultural infrastructure rather than decoration, in an era of global sameness. (Amelia Nash / PRINT Magazine)
Tempest vs Tempest. A book-length, free PDF dive into the source code of Dave Theurer’s Tempest (1981) and Jeff Minter’s Tempest 2000 (1994). The chapters are short and tasty, single mechanics traced from gameplay down through the 6502 and 68K assembly that runs them. Recommended for anyone who wants to see how the arcade games they grew up on were built one assembler routine at a time. Adjacent coolness from the same era: The Visible Zorker: Zork 1, the full ZIL source for Zork’s Generic VERBS file, a 1983 dictionary of every command parser the trilogy could handle. (mwenge / homemade.systems)
The Future Is Shrouded in an AI Fog. Toby E. Stuart, a professor of entrepreneurship, on how AI uncertainty reshapes long-duration investment bets, from capital projects to medical school enrollment. The economic distinction he leans on is risk vs. uncertainty: risk is quantifiable, uncertainty is when the probability distribution itself is unknown. His prescription is optionality: staged investments and adaptive organizational designs that buy you the right, but not the obligation, to follow on. (Toby E. Stuart / Harvard Business Review)
Your feed is overrun with clips. Mia Sato reports on the paid clipper economy. Anonymous accounts get $150 per 100,000 views to chop podcasts and streams into algorithm bait. Clipping.net’s founder claims 62,000 clippers using his platform make $3,000/month on average. The sharper observation is in the back half: when clips become the standard for marketing, the unclipped, complete content is increasingly a means to an end. The full version exists to feed the slot machine. (Mia Sato / The Verge)