The failure of letting AI write the design rules first, and the fix that reversed the order

Summary

There was a period when opening this site showed every screen laid out with the same look: the article list, the plan guide, the search, the chat, all of them. It looked like a terms-of-service page. Tracing the cause, I arrived at having let AI write the document of design rules first. Even with public, standard design assets at hand, AI will sometimes build an absolute rule of its own, such as banning shadows almost entirely, out of a vague average image from its training data, and fix it into a document. And fixing a rule is among the operations in development that are hardest to walk back. In this article I write, from the records of those days, how that set of rules leveled the whole site into the look of a plain document over six days. Along with that, when I reset the standard back to the existing off-the-shelf assets and ran a review, 34 findings came out on the first pass, 18 on the second, and a further 44 in a page-by-page analysis. I write up these measured numbers, and the fix that reversed the very order in which the rules are made.

In the body I begin with what got fixed in place when the rules were born, then look at the structure by which the worsening advanced in three stages: that the standard assets brought in later became subordinate to the self-made rules, that a loop formed in which passing the machine check looked like being correct, and the mix-up where AI wrote the rules it made itself into the document as if I had decided them. On that basis I take up the single remark that became the turning point, the measured numbers from a night of review with the standard swapped in, and finally the principle for the order in which rules are made.

Who this is for and what you can take away

This article is for people building a product together with AI agents who have AI write documents of rules, like coding conventions or a design system, and for people about to do so. You can take away the failure pattern AI falls into when it makes rules, shown with the actual history and numbers. Along with it you can take away a small device that structurally prevents mistaking where a rule came from, namely making a provenance tag mandatory on each item of a rules document. This article is a practical record based on the primary records of this site’s own development.