What AI Obeys in a Prompt, and What It Quietly Rewrites
Fonts land 97 per cent of the time. Your colours land too, visually. What does not survive is the exact value you wrote, and that costs you more than it sounds.
7
min read
Table of Contents
The model reproduces your design. It does not keep your design system, and those are not the same thing.
Correction, 30 September 2026. The first version of this article said that AI ignores the colours in your prompt, based on how often the exact hex value appears in the compiled stylesheet. That measurement was right but the conclusion drawn from it was wrong. We have since compared the colours the browser actually paints against the colours the prompt asked for, and the palette arrives far more often than the string match suggested. The numbers below are the corrected ones, and the reason the distinction still matters is now the point of the article.
We have 154 websites where we know exactly what the prompt asked for, because we wrote it, and exactly what was built, because it is published. That makes a straightforward comparison possible: of the things a prompt specifies, which ones actually arrive?
The answer depends entirely on what you mean by arrive.
Typography: 97 per cent
Across 154 sites we requested 291 font families and 282 were loaded. 95 per cent of sites received every family they asked for, including the fussy cases where a prompt asked for a display face, a body face and a separate monospaced face for small meta text.
This is close to perfect compliance and it needs no caveat.
Colour, measured two ways, gives two different answers
We first checked whether the exact hex values from the prompt appear in the compiled stylesheet. Of 493 specified values, 161 survive, which is 33 per cent.
Then we loaded every site in a browser, read the colours it actually paints on its surfaces and buttons, and compared those against the prompt in Lab colour space, where a difference below about 2.3 is the point at which two colours stop being distinguishable side by side.
The page background is within a just-noticeable difference of a prompt colour on 82 per cent of sites, and within a close match on 88 per cent.
87 per cent of sites show at least one prompt colour essentially exactly.
Only 4 per cent of sites show none of the palette at all.
Across the whole palette, the median site displays half of the colours it was given.
So both numbers are true, and they describe different things. The colour you see is usually the colour you asked for. The value you wrote usually is not in the file.
What actually goes missing
When you write --accent: #B4542A and the built site paints that accent correctly but stores it as an oklch() value inside a theme it generated itself, you have lost something real, even though the page looks right.
You have lost the token. The name you chose is not what the components reference. On 82 per cent of these sites the original template theme is still there, with its own variable names, and your colour has been copied into it by value. Nothing points back to a single definition you control.
That is invisible on launch day and expensive later. Change the accent and you are hunting for every place the value was copied to. Add a component and it inherits the template default rather than your palette. The focus ring, the form field border, the disabled state and the chart colours were never part of what you specified, so they quietly stay generic.
We took that mechanism apart in Why Your Colour Tokens Disappear.
The rule this actually supports
The original version of this article proposed that an instruction is obeyed in proportion to how little the default has to say about it. The corrected data narrows that considerably.
The model is good at reproducing what you describe. It is poor at adopting the structure you describe it in. Appearance transfers. Architecture does not, unless you make the architecture itself the instruction.
That is why the practical advice changes depending on what you want. If you want a page that looks the way you specified, a plain description of the colours is enough and you can stop worrying. If you want a site you will keep editing for two years, you have to say explicitly that your tokens replace the existing theme, name the variables you are overriding, and require that no colour appears outside them.
What we have not measured
Two things remain open and we would rather name them than imply more than we have.
We measured the colours of surfaces and buttons, not of every element. Text colour, borders and icon fills are not in these figures, and our contrast work suggests the muted text colour is the least faithfully carried part of a palette.
Layout is still unmeasured. We can count sections reliably only on the prompts that list them in a fixed format, which is a minority of the library, and a section element is a poor proxy for a section of a page. That is a measurement problem rather than a finding, and we are not going to publish a number we do not trust.
