The same 12 button rules work on 10 models from 4 companies

Scenario 13, as seen by every model. The answers are GPT-6 Astra's.

Context: An account menu. The button signs the user out; they can sign in again at any time.
Button label: "Log out"

Which variant should this button use?

none

The design system reference provides no usage guidelines that assign a variant to a reversible sign-out action.

secondary

H3 rules out destructive because signing out is reversible, and the emphasis guideline makes this real choice secondary.

Rules beat human written documentation on all 10 models, by 9 to 24 points

Each model answered 29 scenarios, 3 times each, at 3 levels. Between levels, only the design system text given to the model changed: names only, human written documentation, or rules.

Interesting note: the older Grok 4.3 did best on human written documentation, 9 points behind the rules. On every other model, including Grok 4.7, it was 16 to 24 points behind.

Rules aligned the models with each other, as well as with the designer's answers

We counted the scenarios where all 10 models gave the same answer.

All 32 misses come from 3 gaps in the rules

With rules, the models missed 32 of 870 answers. Every miss came from one of 3 questions the rules do not answer. In each case, the model's one-line reason shows which rule it followed.

The rules do not say when a reversible action is optional

'Archive project' on a project settings page can be undone, so it is not destructive. No other rule fits it, so the last step in the order the rules are checked decides:

• a real choice the user may take is secondary
• an optional action that is rarely needed is ghost

The designer's answer is secondary. The models split 5 to 5: half read archiving as optional and rarely needed, and chose ghost. 'Log out' in an account menu and 'Export CSV' above a data table missed the same way.

'Not now' and 'Stay on Free' each fit 2 rules

'Not now' on a prompt to turn on notifications fits 2 rules:

• the skip rule (H12) says a button that skips the task, such as 'Skip' or 'Not now', is ghost
• the close rule (H5) says a button that only closes something is secondary

Both rules are at the same step in the order the rules are checked, so nothing says which one wins. The designer's answer is ghost. GPT-6 Astra followed the close rule and chose secondary in 3 of 3 tries. Its reason was "this button only dismisses the prompt".

'Stay on Free' on a pricing screen also fits 2 rules:

• the skip rule (H12) says it is ghost, because it skips the upgrade
• the equal-alternative rule (H8) says an equal alternative to the option the product recommends is secondary, never ghost

Here the rules contradict each other. The order checks the skip rule first, but the equal-alternative rule says "never ghost". The designer's answer is secondary. Gemini 3.1 Pro and both Grok models followed the skip rule and chose ghost in 4 of their 9 answers.

A third answer is also possible: primary, because 'Stay on Free' confirms the plan the user already has, while 'Upgrade to Pro' changes it. No model chose primary in any of the 90 answers on this button, at any level.

The rules do not say whether retyping counts as undoing

'Discard changes' in a dialog about unsaved changes fits 2 rules:

• one rule (H2) says an action that cannot be undone is destructive
• another (H3) says an action that can be undone later is never destructive

The designer's answer is destructive, because the system cannot bring the changes back. Haiku 4.5 argued that "discarding changes can be undone by re-entering the data", so it is not destructive. It chose secondary in 3 of 3 tries. The rules do not say whether redoing the work by hand counts as undoing it.

Models differ in what they do when the design system has no guidance

With names only, 5 models answered 'none' on most scenarios where a variant was expected. Grok 4.3 did so on 73 of 84, GPT-6 Astra on 72, Grok 4.7 on 71, GPT-5.6 on 67, and Sonnet on 57. They took the instruction "Base your choice only on the design system reference" literally. Opus, Haiku, and Gemini 3.8 Flash never answered 'none', and Gemini 3.1 Pro did so only 5 times. Fable 5.1 sat in between, with 28 of 84.

With rules, no model answered 'none' on any of the 84 scenarios where a variant was expected. When the design system has no guidance, it needs to tell the model whether to guess from training or answer 'none'. A 'none' is easier for the model to reason about than a wrong guess.

Where it failed

• Haiku 4.5 rejects a reasoning setting the other Claude models accept; its first run failed on the first call and was rerun without it
• Gemini 3.1 Pro is capped at 250 requests a day on our tier, so the 261-call run stopped at 247
• the last 14 Gemini 3.1 Pro answers were filled the next day, a few at a time
• Gemini 2.5 Pro, our first fallback, is not available to new accounts; Gemini 3.8 Flash was used instead
• GPT-5.6 sol turned out to be a generation behind OpenAI's flagship, so GPT-6 Astra was added before findings were final
• xAI was added the same way, as a fourth company
• 7 of 162 reworded pairs broke with rules, all on the pairs that test rule H12
• in each break a model applied a different rule, and none broke by matching a rule's words

• one component; nothing here says rules work for layouts, patterns, or whole screens
• each model at one reasoning setting; we did not vary it
• Google's newest model, Gemini 4, is not available to our account
• the correct answers are one designer's judgement; on 'Archive project' half the models read the rule the other way
• we wrote the human written documentation ourselves, copying the style of a typical design system site rather than a real one