The walkthrough
How a data page comes together
One conversation. A reporter asks for a response-time tracker; the agent documents the dataset as it goes, and the validator refuses it three times along the way. Each exchange shows what was said and what the agent answered, with the command it ran underneath — and "under the hood" opens the full record: the output, the entry that resulted, and the files on disk.
Every command on this page was run against a scratch project when the site was built — no screenshots, no hand-written output, and no output edited after the fact. The dataset is invented: a fictional county's 311 export, chosen so no frame here could be mistaken for documentation of a real one. Its defects are not invented — each is a shape the survey found repeatedly in real government data.
Three of these are refusals
Watch for them, because they are the parts that do the work. An empty page will not validate. An issue that says "some years are affected" will not validate. And renaming an issue while its anchor still sits in the loader exits non-zero — the code is carrying an explanation nobody can look up any more.
None of that is the agent being careful. It is the format refusing, on the agent exactly as it would on you.
What is left at the end is small: one markdown file a reporter can read, a generated index, a published bundle, and a Python file with a comment in it. The validator never wrote a word of it.
See what those blocks are → · Read a real page → · Do this on your own repo →