What is claude-design-system-prompt?

This repository shares a system prompt and a set of skill files that turn an AI chat assistant into a design collaborator with strong opinions about visual quality. It was built by reverse engineering the prompt behind Claude's design tool from Anthropic, then packaged so anyone can paste it into any large language model that accepts a system prompt, including Claude, GPT, Gemini, or models run locally. The core idea is to push back against generic AI generated design work: heavy gradients, emoji used as decoration, cards with a rounded corner and a colored left border, and the same one or two fonts appearing everywhere. Instead the prompt lays out a full design philosophy covering content discipline, so nothing appears unless it earns its place, visual hierarchy through size, color, weight and spacing, accessibility rules like keyboard navigation and focus states, and treating components and reusable tokens as the building blocks rather than one off pages. Alongside the main prompt sit 14 separate skill files the assistant can call on for specific jobs: asking discovery questions before starting, picking an aesthetic direction when no brand exists yet, building wireframes, making slide decks or clickable prototypes, extracting a design system or component inventory from existing sources, and running review passes for accessibility, generic AI look, visual rhythm, interaction states, and a final polish check. These skills are meant to be chained together, for example moving from discovery questions through aesthetic direction, wireframing, prototyping, and a polish pass. There are two variants included: one written for Claude Code and Claude.ai, using subagents for parallel review work, and one adapted for OpenAI's Codex that runs the same reviews in sequence instead. The project also documents how its instructions were calibrated for newer models that follow directions more literally, noting that older models may need stronger, more insistent phrasing to get the same behavior. The project is open source under the MIT license, so it can be copied, modified, and reused freely, including for commercial purposes, as long as the copyright notice is kept.