"Almost" is what happens when a designer gets examples without a theory, so it copies the surface look instead of understanding the rule underneath. Three moves close the gap between "close" and "exactly".
- Principle before example. The common order is to throw twenty screenshots of designs you like at the agent, and the result is the average of them. The order that works is reversed: first write the theory (why that spacing, why one accent color, why the heading is that size), and only then hand it those same twenty examples and ask it to analyze each one against the theory. An example without a principle produces imitation, a principle with examples produces judgment.
- A checker that is not the designer. A designer reviewing its own work always passes the review. Here, every visual output goes through a separate agent whose only job is to hold the work against the rules and say pass or fail, before a human ever sees it. One community member who built exactly that kind of checker, an unusually obsessive one, reached around ninety percent fidelity against their own design system. (Ours rejects me too, roughly one illustration in three, which is precisely its job.)
- Drop what it still cannot do. Editable vectors, clean SVG, generating straight into Figma: that is where it still breaks, and every attempt to force it lands you back at "almost". The path that works today is to build in HTML, approve, and export at the end.
One more thing from the field, and it moves the needle most: when an output comes back imprecise, the correction belongs in the rules file, not in the conversation. A fix in a conversation lasts one conversation, a fix in a file lasts from there on. A community member who consolidated their whole design system into one organized file, instead of re-explaining it on every request, summed up the result as "most outputs come out almost exactly right", and that is the gap between an agent that guesses and an agent that knows.
The principle behind all of it: precision in an agent is the result of a check that runs every single time, not talent. A human designer learns your taste over a month, and an agent learns it the moment you write it down as a rule that can be checked.
A prompt, on the house
Before you design anything for me, we work in two stages.
Stage 1, theory: read my design rules file and write me
five lines on the principles behind the rules, not the rules themselves.
Stage 2, examples: here are ten examples I like. Go through each one
and write a single line: which principle it demonstrates, and where it deviates.
From now on, before you show me any output: run it against the principles,
write down what failed, fix it, and only then present it. If something
does not meet a rule and you have no solution, tell me instead of
quietly compromising.The first day with this order feels slow. After a week, "almost" becomes a rare note instead of a permanent feeling.
