Audit design system
An agent builds from what your system gives it. Where it is clear, the agent reuses your parts. Where it is not, it guesses: raw values, rebuilt components, parts nobody asked for. The audit shows where AI follows your system, where it invents, and what each gap costs.
For teams with a design system in code. No system yet: I build it from scratch.
Why now
When components and tokens are reachable, an agent writes less, reuses more and needs fewer corrections. When they are not, you pay for every guess in tokens, review and rework.
Result · my own site, measured
44% shorter: the same case-study screen is 21,229 characters written raw and 11,961 built from components.
The cost of drift
Once when it is written, again when someone catches it in review. Drift is a throughput problem, not a tidiness one.
Tokens
A component call is one line. The same part by hand is forty, and you pay for all forty every time.
Review
An off-scale value looks right and reads wrong. Catching and fixing it costs more than generating it.
Consistency
Nine near-identical greys. A button rebuilt because the real one was hard to find. Nothing fails a test; the product feels loose.
Accessibility
Contrast below the minimum, motion that ignores the reduced-motion setting. Nobody notices until someone cannot use the screen.
What the audit checks
They answer three questions: how AI reads your system, whether it follows it, and what it does when something is missing. Tested on Primer, Radix Themes, Carbon and Chakra UI: every one had findings worth fixing.
Every hard-coded value in your components, checked against your tokens: equal to a token nobody found, or on no scale. Plus unused and duplicate tokens.
Contrast on every declared pair, in every theme. Reduced motion, high contrast, visible focus, and right-to-left if you need it.
Components rebuilt by hand, parts used outside their contract, styling that changes what a component is.
Unsourced numbers and claims, and words your voice rules ban. An agent repeats what the system teaches it.
What a coding agent has to invent: decisions it made up, and names it expected but could not find.
What the audit finds
Each costs something different to fix.
Mechanical
The decision exists; someone typed the number. Find, replace, and prove with a before-and-after render that nothing moved.
Needs a decision
An optical nudge or drift: only your team can say. Grouped by where they cluster, so deciding takes minutes.
Affects people
Contrast in every theme, and motion against the reduced-motion setting. Teams fix these first.
Getting worse
A coding agent fills gaps by guessing. The audit shows what it made up and what it could not find.
The audit
From the day access arrives. Fixed scope.
What is included
What is not
Questions
Then the audit compares the design file with what ships. Tell me what you have and I will say what the report can cover.
No, it runs on top. Storybook shows each component works; the audit checks whether screens built with it follow the system, plus contrast, reduced motion and high contrast.
Yes to both. Without repository access, an export of the design system and a few product screens is enough for the drift list.
Then the report says so and names what it checked. The four public systems all scored high on tokens and still had findings worth fixing.
I do. The scan is automated. I check every finding by hand, because scans are wrong often enough to matter, then rank the fixes and write the report.
Who runs it
Nine years designing products, lately also writing the code. Based in Barcelona, on US East Coast hours. Fintech, logistics and energy, from a seed startup to one of Latin America's largest investment banks.
I built the design system for an AI platform in private capital markets, in Claude Design: 48 components, 9 templates, tokens as the single source of truth, validation that catches AI drift. The audit is that method, applied to your system.
See the workThe audit shows where AI follows your system, where it invents, and what each gap costs.
Book an audit