A short piece of fiction, written by API Evangelist. The reader is imagined; the post, and every number and quotation from it, is real: "Modeling An API By Talking To It With Structura," Kin Lane, November 3, 2026, https://apievangelist.com/2026/11/03/modeling-an-api-by-talking-to-it-with-structura/
Marisol read the post on a Tuesday evening, after the office had emptied, with a cold coffee and a whiteboard photo still open in another tab. It was a photo of her own team's design for a returns service, taken eight months ago. Nobody had touched it since, and half of it was now wrong.
She started out annoyed. The post says Pedro is the first to ask "why another UML editor?" and she wanted to answer for him: because nobody needs one. She had sat through enough modeling tools that produced beautiful diagrams and then rotted next to the code. Her eye slid past the part about talking to an assistant. A minute to get a class model of publisher, contract, author, book and language sounded like a party trick, and books are an easy domain. Returns have carriers, refund rules and three warehouses that disagree about what a pallet is.
Then she reached the sentence about the textual syntax, and she slowed down. The model can be reviewed by a human, tracked in Git, and read or written by an AI. That was the first claim she could picture in her own repository. A diagram she would ignore. A text file with a diff she could argue about in a pull request was different.
The deterministic part she believed more than she wanted to. She had watched a teammate regenerate a whole client with a language model and then spend two days finding out which endpoints had quietly changed shape. The post says the same model produces the same output every time, in about ten seconds, and that matched what she wished her own pipeline did. She was less sure about the pitch of about 80% generated, with the team owning what the site calls "the part that actually matters." Her team already owned that part. What she doubted was whether the 80% would be code she could stand to read. The post doesn't answer that, and to its credit it doesn't pretend to.
She also noticed what the author admits is missing. The main domain has nothing for an agent to find, and there is no server card. She felt a small, mean satisfaction at that. Then she remembered how rarely she had checked her own service for the same gap.
Her real discomfort came from the line that the AI drafts and explains while the architect stays responsible, described as "AI as your exoskeleton, not the driver." She liked it and distrusted it in the same breath, because responsibility only works if somebody actually reads what was drafted. At her company, nobody read drafts. They read summaries of drafts.
She did not sign up for anything that night. She closed the whiteboard photo and opened a blank text file instead. For twenty minutes she wrote down the returns domain as plain classes with typed fields: return, item, carrier, refund, warehouse. She guessed at the length limits on the strings and flagged the three places where she had no idea.
The post ends with a simple instruction: model something you know well, talk to it, and look at the generated code. She put a reminder in her calendar for Thursday afternoon. Her plan was to paste in that text file and ask whether a return can be split across warehouses, because the three of them had never agreed on that. She would look at the code last.