AI and design systems: what actually works (and what doesn't)

Published April 27, 2026, this video takes an honest look at what actually works when you try to use AI tools inside an existing design system. The presenter runs through real workflows covering three tool categories: AI-assisted design tools, Claude Design as a visual exploration layer, and Claude Code as a code generation layer connected to Figma via MCP.

The video is aimed at product designers and design system leads who have heard many AI promises and want a pragmatic check before committing to new tooling or workflow changes. It assumes a working design system is already in place and focuses on where AI accelerates work versus where it adds more review overhead than it saves.

Key takeaways:

1. AI generates components more reliably when a design system has clear, unambiguous token names and consistent naming between Figma and code. Inconsistency between Figma library names and code component names is identified as the single largest source of friction in AI-assisted design system work — not the AI tools themselves.

2. Claude and Figma MCP work well together for generating new screen variants within existing system constraints. The integration lets you paste a Figma frame URL and have the AI build a new layout using real components rather than placeholders, which is a meaningful improvement over earlier approaches that worked from screenshots.

3. Component documentation quality matters more than most teams anticipate. AI tools that pull from connected component libraries perform noticeably better when components have descriptions, usage guidelines, and example states. Sparse documentation leads to generic, off-system outputs.

4. The failure cases are consistent across tools: AI struggles with interaction design decisions, accessibility nuance, and anything requiring domain-specific judgment about information hierarchy. The video is direct that AI should be treated as a generative assistant, not a decision-maker, in these areas.

5. Automated design system maintenance — using AI to flag drift between Figma components and their code equivalents — is identified as the most underused practical application. The workflow is already viable with current MCP tooling and does not require a dedicated engineering resource to set up.

Worth watching if you manage a design system and are evaluating whether to invest in AI tooling for your team, or if you are a designer who wants a realistic benchmark against which to test your own tool experiments before presenting a recommendation to your team.