A reverse-engineered version of Anthropic's Claude design system prompt has been open-sourced on GitHub, offering developers a comprehensive toolkit to transform any large language model into a sophisticated design assistant.

The repository, published by developer Trystan-SA, contains a complete system prompt and 14 procedural skills that explicitly reject common AI-generated design patterns in favor of accessibility-focused, professional output.

Beyond Generic AI Design Templates

The prompt addresses a persistent problem with AI design assistants: their tendency to produce generic SaaS templates featuring aggressive gradients, excessive emoji decoration, and ubiquitous Inter typography. Instead, it implements a complete design philosophy covering content discipline, visual hierarchy, and accessibility standards.

The system includes WCAG compliance, semantic HTML generation, keyboard navigation support, and proper focus ring implementation. It also enforces interaction states for hover, active, disabled, and loading conditions across all components.

The toolkit comprises two main variants: a Claude-specific version optimized for Anthropic's models, and a Codex variant adapted for OpenAI's systems. Both include the same 14 skills divided into production, system extraction, and review categories.

Production-Ready Skills Library

The skills range from discovery questions and wireframing to interactive prototype generation and accessibility audits. A typical workflow might progress from discovery questions through aesthetic direction to wireframing and prototype creation, concluding with a comprehensive polish pass.

Each skill operates as a self-contained procedure that triggers when user requests match specific criteria. The system can chain multiple skills together, enabling complex design workflows from initial concept through final implementation.

The prompt assumes an HTML-output environment similar to Claude's design tool but can be adapted for other platforms including Figma plugins or code-only assistants. The core principles translate across any design medium.

The repository has attracted 1,800 stars and 235 forks since publication, indicating strong developer interest in more sophisticated AI design tooling. The MIT license permits commercial use and modification.