He Stopped Reading the Plans

In July I wrote about my coding dashboard: 231 agent runs, a thirty-seven-day streak, about a billion tokens. I said in One Billion Tokens Later that you still have to read the work. I did not say how much of it I had read.

Two months earlier an engineer at Anthropic had been more forthcoming. Thariq Shihipar works on Claude Code. In May he wrote that he now asks agents for HTML pages where he used to ask for Markdown. shihipar-html His last reason was the candid one. As the agent took on more, he had been "reading plans less closely", and richer pages made him feel back in the loop. I design screens for a living, and I know that feeling from the other side of the table. It is what a well-made page produces in the person approving it. He found the right problem and trusted the wrong signal.

Markdown Was Built for a Writer, and Nobody Here Is Writing

Markdown is the plain-text format agents write by default, and it got there for a dull reason. Early models had small limits, and Simon Willison recalls asking for it when GPT-4 stopped at 8,192 tokens. willison-markdown-infoq It was also easy to edit by hand, which mattered while people still edited. Shihipar says he increasingly does not. When a file needs changing, he usually prompts the agent to change it.

John Gruber designed Markdown in 2004 and was plain about its job. "HTML is a publishing format; Markdown is a writing format," his documentation says. gruber-markdown-syntax Once nobody on the team is writing, the writing format is out of work. In design terms, the author of the document has become its audience, and the document is now made by the party it describes. Shihipar passes the same pages to a second agent to build from and another to verify. A verifier checks the build against the plan. Only the person who asked can check the plan against what they meant.

Three of His Twenty Examples Hand Something Back

Shihipar published twenty sample pages with the post, and I opened all of them. shihipar-html-examples Three end with something the reader made. A board of 24 tickets has a button that copies your ordering. A panel of feature flags copies the ones you changed. A prompt editor copies your edit. Three more put a question to the reader in words. The other fourteen ask for nothing.

Designers have a name for what the fourteen do. In 1995 Kurosu and Kashimura at Hitachi had 252 students rate 26 layouts of a cash machine screen. How easy a layout looked rose with how beautiful it looked. It barely moved with most of the measures designers use to predict real ease, such as the distance between keys. kurosu-kashimura-chi95 I cited that study in Design is Attention Control to argue that a pleasant screen can hide a misdirection. Here the pleasant screen is a plan.

The Page Says Which Parts Are Safe

Design decides what a person sees first, what they see second and what they never see. On these pages the agent is the designer. Its sample code review sorts the changed files into "safe", "worth a look" and "needs attention". shihipar-html-examples Its sample pull request has a section telling the reviewer where to focus. The ticket board arrives already sorted by the agent's best guess. Three pages keep their detail folded away until you click. In each case the party under review has drawn the reviewer's route.

I have drawn that route myself. At John Lewis I led the design of search and browse, and I took two versions of the size filter to a review. One hid the sizes that had sold out. The other left them in view, with a way to be told when they returned. The room wanted a grid that looked shoppable, the first version looked finished, and I took the approval for it. There Is Perfection in Imperfection tells the rest. A woman looking for her size saw a clean grid, because I had removed the gap that would have told her it was gone.

The End of the Walkway

Step onto a moving walkway at an airport and you stride like someone much fitter. Your legs are doing their ordinary work. The floor is doing the rest, and from the inside you cannot feel the join.

You feel it at the end. The belt stops, your feet do not, and for one step the ground grabs you. That lurch is the first step you take on a plan you only rode through.

He Is the Reader Least Likely to Be Fooled

A senior engineer reading a plan for a system he built is hard to mislead. He knows which parts of the data flow are fragile, and he has a stake in the answer. The evidence favours him. A Microsoft and Carnegie Mellon survey of 319 knowledge workers found that confidence in the AI went with less critical thinking. Confidence in one's own skill went with more. lee-critical-thinking-chi2025 His real alternative is no reading at all, since he admits he rarely gets past a hundred lines of Markdown. shihipar-html A format that gets a tired expert to look has raised scrutiny from nothing. At CHI in 2006, Maryam Tohidi and colleagues found people criticised a design more freely when shown three than when shown one. tohidi-right-design-chi2006

Most of that I accept, for him. His page of four layouts, side by side with no verdict attached, is Tohidi's experiment made routine, and I would use it tomorrow. The trouble is the measure he trusts. In 2021 Zana Buçinca and colleagues at Harvard gave 199 people a food-swap task with an AI's suggestions. Some screens made them wait, ask or decide before the AI's answer appeared. Those screens cut reliance on the AI's mistakes, and they drew the least favourable ratings. bucinca-trust-or-think-2021 They were online workers on a short task, so I hold it loosely. It still separates two things he treats as one: how in the loop a page makes you feel, and how much you checked. He also says colleagues are far likelier to open a link than a file. The page goes to people who did not sit through the session that produced it.

What He Noticed

Shihipar noticed that the people who approve an agent's plans no longer write them, and he was right to be uneasy. The test of a page is what it asks before you say yes, because the party you are checking built it. I built one once, a room approved it, and a woman looking for her size paid for how finished it looked.