Turn One Odd Section Into a Rule for AI── Name the set that defines the shape and the attributes to match, and the next thing you add inherits the same form

01"This one section looks different" was a request about the whole site

In September 2026 the operator opened a section newly added to the site and saw that it did not look like the rest. The request that went to the AI was not to adjust that section's colours and spacing. It was to bring the section into the same shape as the other pages: a condition the whole site should satisfy.

In the same exchange the operator also narrowed the section's name. The heading claimed a wider subject than the content actually covered, so the name was cut back to match the content.

The odd thing was one section. What went to the AI was not a fix for one section. What that rewording changes is worth following through.

02An instruction about a part fixes one thing; a condition binds the rest

Ask for a bigger heading in one section and that heading gets bigger. Only the thing you pointed at changes. Ask for the section to match the other pages, and the AI has to read those pages, derive the rule they follow, and apply it.

The scope of the fix is not the only difference. Phrased as a condition, the request also changes what the AI consults. The reference stops being the numbers and words in your message and becomes the pages that already exist.

03Fixing only the part means the next section drifts the same way

Settling for a local fix costs twice. First, you give the same feedback again and again. Every new section drifts, you notice, you ask, and the work restarts from zero each time.

Second, readers carry the cost. The fourth of Jakob Nielsen's usability heuristics holds that users should not have to wonder whether different words or different placements mean the same thing. The Nielsen Norman Group's explanation of that heuristic points to internal uniformity within one product, to written design decisions that keep a team aligned, and to regular comparison against existing screens to find what has slipped.

Drift arrives one place at a time. That is why the cheaper move, at the moment you notice, is to write the condition rather than repair the instance.

04Matching means a countable set of four attributes

"Consistency" covers a lot of ground. A request lands better when the attributes to be matched are named separately.

All four can be written as conditions you can test. The W3C's accessibility guidelines set two of them as criteria to meet: components with the same function within a set of pages are identified consistently, and navigation repeated across those pages appears in the same relative order each time. The first does not demand identical text everywhere. It demands that two buttons doing the same job not carry different labels.

Some differences should stay. Line setting differs between the Japanese and English editions, and some emphasis is deliberately unlike its neighbours. Say so in the request. The pattern also needs a reference set: the very first instance of anything has nothing to be matched against.

05In material review, one comment can be rewritten as a condition on the whole set

A single product usually has several pieces of promotional material: an explanatory piece, one carrying a comparison table, a handout for a lecture. Terms diverge between them, the safety wording moves from one position to another, and sources are cited in different styles. When you find this in one piece, you can fix that piece, or you can state the condition for every piece belonging to that product. The later workload differs sharply.

The same applies when an AI does a first pass. Instead of "correct the terms in this piece", ask it to match the terminology, the position of the safety wording and the citation style to the product's existing material, and to list everything that differed. You get two things back: the corrected piece, and the list.

The scope of a name matters here too. When a heading claims more than the content covers, readers expect something that is not there. Narrowing the section's name was exactly that correction, applied to a heading. Titles of material and captions on figures drift in the same way.

06An AI can only match the material it was told to read

A language model builds its output from what you passed it and what it read. Without being told to read the existing pages, it builds the new thing the default way. That is why newly added parts drift most. Nothing was skipped in bad faith. Nothing was given to compare against.

Anthropic's published guidance for Claude Code states this directly. Rather than asking for a widget, it recommends pointing the model at how existing widgets are built and telling it to follow that pattern. The same guidance describes keeping conventions in a file named CLAUDE.md, which is read at the start of every conversation, and warns that an overlong file buries the rules that matter so they stop being followed.

Consistency therefore works in two stages: naming the reference in the request at hand, and recording the rule so it survives into the next request. With only the first, you repeat yourself forever.

07From tomorrow: name the reference, the attributes, and ask for the list of differences

There are four steps.

1. Name the reference set. Not "like the others" but the specific existing pages or documents that define the shape.
2. Name the attributes. State which of appearance, placement, wording and scope of the name should be matched.
3. Ask for the differences. Require a list of what differed, not just the corrected result.
4. Record it as a convention. Write the rule down briefly so the next thing you add inherits it.

Drop the third step and you have nothing to check against. "I have made them consistent" cannot be verified on its own. Anthropic's guidance recommends giving the model a check it can run itself, such as a test or a screenshot compared against the original. The list of differences serves that purpose when no such check exists.