Open source system prompt and skill library turns LLMs into design collaborators

A newly published open-source project provides a reverse-engineered system prompt and skill library intended to turn LLMs into opinionated, accessibility-aware design collaborators that resist common AI-generated tropes such as excessive gradients, emoji decoration, and overuse of rounded-corner cards.

The repository includes a system prompt split into 20 chapters and 14 procedural skills organized into production, system, and review categories. Production skills cover tasks like kickoff questioning, aesthetic direction, wireframing, slide deck creation, and interactive prototyping. System skills focus on extracting design tokens and components, while review skills audit accessibility, detect AI slop, evaluate visual hierarchy and rhythm, check interaction states, and perform final polish passes.

The project offers two variants: one tailored for Anthropic’s current frontier models (Fable 5 and the Opus 4.7/4.8 lineage) and a second adapted for OpenAI’s Codex. The Anthropic variant is calibrated to leverage models that follow instructions more literally, using conditions instead of quotas and explicit triggers for skills and subagents to reduce over-triggering.

The system prompt emphasizes content discipline, aesthetic discipline, visual hierarchy and rhythm, accessibility (WCAG, semantic HTML, keyboard navigation, focus rings, motion preferences), interaction and feedback states, system thinking, and respect for the medium (e.g., real CSS Grid, oklch(), text-wrap: pretty).

Users can paste the system prompt into any LLM that supports system prompts (Claude, GPT, Gemini, local models) and invoke skills by name when tasks match. Skills can be chained to support end-to-end workflows, such as a greenfield design flow from discovery to polished prototype or a brand-aware flow from token extraction to variation generation and tweaking.

The repository is MIT licensed and includes a README with usage instructions, a license file, and a directory structure that separates the system prompt, skills, and variants for different model families.