I tested 12 small AI models on 2,500 design-system questions. The model wasn’t what mattered.

Designer Arnaud Morvan found that a one-page guide for each task lifted small AI models on design-system questions far more than switching to a different model did.

Arnaud Morvan, a designer building Axolo, an AI prototyping tool, wanted design-system imports to run without any paid external AI service. He tested a dozen small open AI models on about 2,500 questions from real design systems, with 271 hard cases settled by hand. A short guide for each task, written from his own design decisions, mattered more than which model he used.

• Scripts beat the AI models on mechanical work such as sorting tokens, 97 percent against 92 percent, while the models won on meaning, such as default button text.
• With a guide of about 100 words per task, three small models rose from between 61 and 64 percent correct to between 78 and 81 percent on unseen cases.
• Importing shadcn/ui, a public component library, took 58 seconds by script; the AI model then made 1,004 suggestions and flagged 57 items for a person to confirm.
• Larger models still composed better screens, so Axolo uses the small model for import and Claude, Anthropic's AI model, for later design changes.

This is a vendor's account of his own product, and one designer settled the hard cases. A design-system team could test whether short written guides improve an AI model more than a larger model does; that is an inference.

Derived by Working Surface from the article; more in the Playbook. Source line: The skill matters more than the model.

27 October 2026: AI produces reliable work when a team writes down how its own system works and checks the output against something the AI did not write, from frozen rules to audits.