AI-generated prototypes fail because design systems are not built for machine consumption. Hardik Pandya, a designer at Atlassian, published a practical guide arguing that undocumented decisions, hard-coded values, and over-reliance on AI pattern recognition from mockups are the root causes of drift and inconsistency. The fix is not better prompting. It is treating every design decision as infrastructure that must be written down and kept current.
The architecture Pandya proposes has three layers. First, structured Markdown spec files encode spacing rules, color choices, component usage guidelines, and decision-making priorities so AI reads explicit instructions instead of inferring from visual mockups. Second, a token layer constrains AI to a closed set of named variables, eliminating ad-hoc value invention. Third, an audit script scans output and flags every hard-coded value before it ships. A sync routine then monitors design system updates and marks which spec files need revision. For teams auditing existing Figma files, the free FigmaLint plugin catches detached instances, missing interactive states, and unbound tokens before they become AI inputs.
The original guide is worth reading in full because Pandya does not stop at theory. He links to four live llms.txt implementations from Atlassian, IBM Carbon, the CMS Design System, and Nordhealth, so you can inspect exactly how production teams are structuring machine-readable design documentation today. The broader implication: extending code is more reliable than generating it from mockups, and the teams building that foundation now are the ones who will spend less time correcting AI output later.
[READ ORIGINAL →]