Skip to content

( By / Astreya / Product UI/UX Designer / 2026 / 5 min read )

AES Claude Design System

Successfully built the AES Claude Design System, an AI-interpretable design system grounded in the Astreya MUI design system, achieving high-quality, design-system-compliant prototype generation directly from structured briefs

The result and the decision behind it

( What it did )

Successfully built the AES Claude Design System, an AI-interpretable design system grounded in the Astreya MUI design system, achieving high-quality, design-system-compliant prototype generation directly from structured briefs

( The decision behind it )

Build a Structured Local Repository as the System Foundation

Direct Figma file provision produced hallucinations and inconsistencies. A structured, versioned repository gave Claude Design a reliable single source of truth and made the system reproducible and maintainable across the project.

Generated schematic standing in for imagery of AES Claude Design System. It depicts nothing.
Generated schematic, not a screenshot of AES Claude Design System.

Generated schematic, not a screenshot. Real imagery to follow.

( The account )

Context

Astreya's design team had an established, branded MUI design system in Figma, the source of truth for product UI across internal and client-facing tools. When Claude Design launched, the team saw an opportunity to shift from a traditional design process to an AI-driven one: generate prototypes rapidly from design briefs while keeping full consistency with established brand and component standards.

No methodology existed for doing this at the time. Claude Design was newly released, and the work required original research, hands-on experimentation, and building a documented system from scratch.

The problem

Translating a production Figma design system into Claude Design accurately was not straightforward. Initial attempts produced poor results: design tokens were missed, components were hallucinated, and generated outputs did not reflect the actual design system. There was no established playbook for structuring a Figma file, documentation, and assets so Claude Design could interpret them reliably.

A second problem emerged as the system matured: design briefs from other teams were written without design system awareness, conflicting with component rules and producing broken prototypes.

What I did

The solution was the AES Claude Design System: a structured, AI-interpretable design system ecosystem built from the ground up, taking control of every input layer feeding Claude Design, including the Figma file structure, the component documentation, the token architecture, the repository, and the design briefs themselves.

The system rests on four layers. First, a GitHub-connected repository as the single, versioned source of truth for all documentation, tokens, and assets. Second, a systematic audit methodology that verified the Figma file was genuinely interpretable by Claude Design before any generation work began. Third, layered component and screen documentation covering MUI atomic components, MUI composite components, and Astreya's custom components, each with anatomy, usage rules, and code references. Fourth, a brief alignment workflow that rewrote incoming briefs to speak the design system's language before any prototype was generated. A handful of recurring tasks across auditing, documentation, and brief rewriting were automated with custom Claude skills, but the skills were a means to operate the system consistently, not the deliverable itself.

Each layer addressed a specific failure mode identified during early experimentation. Together, they transformed an unreliable generation process into a reproducible, high-quality design system that produced compliant prototypes directly from briefs.

Outcomes

  • Established a reproducible, GitHub-connected repository architecture, spanning Figma structure, token files, and layered component documentation, that serves as an ongoing source of truth for design system maintenance
  • Proved that AI-assisted rapid prototyping is viable at Astreya's scale when the design system is correctly structured, documented, and the brief pipeline is aligned
  • Built a component glossary covering every MUI and AES custom component as the reference layer Claude Design and incoming briefs are checked against
  • Supported the system with a small set of custom Claude skills that automated recurring maintenance tasks, keeping the pipeline consistent as the design system grew

What I learned

Claude Design quality is only as good as the input structure. Investing in Figma hygiene, token correctness, naming conventions, and component documentation directly translated to output quality. There was no shortcut.

The README is the backbone of the system. Claude Design weights it heavily. Keeping it accurate, comprehensive, and rule-rich was the single most critical maintenance task for consistent prototype output.

Briefs need to speak the design system's language. Briefs written without design system awareness consistently produced broken prototypes. Brief regeneration before prototyping became a non-negotiable step in the workflow.

Iterative refinement beats one-time setup. Claude Design matured during the project. Staying adaptive and continuously testing new approaches was essential to sustained progress.

The full record / 4 decisions
My role
Product UI/UX Designer
Timeline
May 2026 – July 15, 2026
Platform
Web (Claude Design, Claude Cowork, Figma, GitHub)
Category
AI-Driven Design Systems
Status
New design, delivered

Tools Claude Design, Claude Cowork, Figma, Figma MCP, MUI (Material UI), GitHub, Claude Chat

Team Solo (cross-functional collaboration with design and engineering teams)

Decisions

Build a Structured Local Repository as the System Foundation

What
Created a Claude Cowork project housing all documentation, tokens, skills, and assets in a local GitHub-connected repository rather than working directly from the Figma file alone.
Alternatives
Providing the Figma file directly each session and iterating on outputs.
Why
Direct Figma file provision produced hallucinations and inconsistencies. A structured, versioned repository gave Claude Design a reliable single source of truth and made the system reproducible and maintainable across the project.

Build the fig-audit Skill Before Any Other Implementation Work

What
Before building documentation or providing assets to Claude Design, I built a skill specifically to audit whether a Figma file was ready for accurate AI interpretation.
Alternatives
Moving directly to documentation generation and iterating on failures.
Why
Without a systematic readiness check, iteration was blind. The `fig-audit` skill with its three-guarantee framework made failures diagnosable and fixable, not just observable.

Design System Takes Priority Over Brief Instructions in All Conflicts

What
Established a conflict resolution rule where the design system governs component selection and layout structure, and the brief retains only content and business intent.
Alternatives
Giving briefs authority over the design system when they conflicted.
Why
When brief instructions overrode the design system, prototypes broke. The design system encodes architectural constraints the brief author may not be aware of. Brief intent is preserved; structural decisions are governed by the system.

Automate the Recurring Pipeline Steps, Not the Design Work Itself

What
Recurring, repeatable steps in the pipeline, auditing a Figma file, documenting a component, centralizing the glossary, and rewriting a brief, were each automated with a small custom Claude skill, so the design system could be operated and maintained consistently as it grew.
Alternatives
Treating each phase as a standalone manual process with manual handoffs.
Why
The design system itself, not any single skill, was the deliverable. Automating the mechanical, repeatable steps kept the system's inputs consistent every time it needed auditing, documenting, or a brief rewritten, and freed my time for the design and methodology decisions that actually determined output quality.

( Tags )

  • Design Systems
  • Design Tokens
  • AI Workflow