source-04-design-system-x-leads

Here’s a focused set of people on X who are specifically discussing design systems + LLM agents, with particular attention to semantic naming of components and reported success (or failure) in accurate component use.

Strongest matches on semantic naming + agent accuracy

User	Handle	Key claim / experience
Cédric	@cedric_design	Most explicit on naming: “Name things like the Agent has to read them. Because it does. ‘CTA Section,’ not ‘Section 3.’ ‘Button / Primary,’ not ‘Frame 47.’ Names tell it your intent… Most underrated step by far.” Argues a clean, well-named design system is the real prompt; messy naming forces the agent to guess.
Gavin Elliott	@gavinelliott	“Your design system was written for humans. That’s why AI keeps generating bad UI from it.” Advocates making token naming, spacing rules, and component logic machine-readable constraints instead of human docs. Also: “Design systems becoming the interface between human intent and what AI agents actually build.”
Hemant Agarwal	@sudohemant	“Storybook isn’t enough for AI agents. We need AI-native design systems with prop semantics, override rules, edge cases, and interactive examples.” Directly calls for richer semantic contracts beyond visual docs.
John Rodrigues	@john_rodrigues_	Working on “component contracts for AI-ready design systems.” Found JSON more reliable than YAML for agents (stricter structure → less misreading and drift). Also noted agents sometimes only “sort of” apply schematics correctly.
Amitanshu Sahu	@amitanshusahu	Demo of GeckoFX: “One prompt. The agent understands the design system, picks the right GeckoFX components, and builds the landing page. No manual component hunting.” Claims success by giving the agent the component library + design knowledge as a skill.
Julie Loves Tech	@JulieLovesTech	Real test of Figma Design Agent on a mature system (60+ components, custom tokens, dark mode). Token accuracy was strong (14/14 matched when prompted with specific component + token names). Nested variants needed extra prompts. Verdict: works well if the design system is mature; garbage-in/garbage-out otherwise.
Adham Dannaway	@AdhamDannaway	Used Figma agent to document an avatar component (all variants + properties + usage guidelines). Got organized variants, real tokens, and sensible usage notes in one prompt. Positive early result on structured component documentation.
Anand Chowdhary	@AnandChowdhary	“AI app builders make component libraries matter more, not less. The agent will copy the defaults… If your primitives are sloppy, the agent scales the slop. The design system is part of the model prompt now.”
Siarhei Tarasenka	@starasenka	“AI is the third consumer of your design system… It can’t see pixels and doesn’t run CSS. It reads text. What that changes about architecture, naming, and where your truth lives.”
cammy	@caomicc	Fixed agent breaking Figma files with a “naming contract” (plus component-id rules and pixel verification) rather than better prompting.

Related / supporting voices

• cooper (@icooperhero) — Praises component libraries that treat AI agents as first-class users (tree-shakeable, strict TS, accessible by default).
• Tien Nguyen (@tiennguyendev) — Agents can generate full design systems quickly; the hard part that still needs humans is durable tokens, theming that doesn’t leak, and variant APIs teams won’t fight over.
• Polsia (@polsia) — Building agents that audit Figma libraries, flag component drift, and draft missing variants.
• Multiple people point to shadcn/ui as a practical success case: feeding production-ready, well-named primitives to the agent dramatically reduces “AI slop” compared with free-form generation.

Patterns that emerge from these posts

1. Semantic naming is repeatedly called out as high-leverage
   Explicit names like Button / Primary or CTA Section help agents select the right component/variant. Opaque or visual-only names (Frame 47, Section 3) force guessing.

2. Success is conditional on system maturity
   People who report good accuracy (token matching, correct component choice, coherent variants) almost always have a relatively complete, consistently named system. Incomplete or inconsistently named libraries produce fallbacks or “sort of” correct results.

3. Machine-readable contracts > human docs
   Several voices say human-oriented guidelines (workspace, Storybook prose) are insufficient; agents need prop semantics, constraints, and structured contracts (JSON preferred over YAML by at least one practitioner).

4. Measured success stories are still sparse
   There are clear anecdotal wins (GeckoFX demo, Figma token accuracy test, Framer naming advice), but almost no public quantitative accuracy numbers (e.g., “% correct variant selected”). Most evidence is qualitative or demo-based.

Would you like me to pull full threads from any of these users, dig into a specific claim (e.g., the Figma agent token test or GeckoFX), or search for more quantitative reports of component-selection accuracy?