可以关注一下这个 https://github.com/Trystan-SA/claude-design-system-prompt… Claude Design System Prompt

有人逆向工程了 Claude Design 的设计系统提示词，提供了一套完整的设计哲学。

里面有主 prompt，居然有 20 个章节… 还有 14 个 skills。 还提供了针对 Codex 的版本，当然我估计这套玩意儿肯定是在 Claude 效果最佳的，肯定做了针对性优化。

大家感兴趣的可以在别的模型上试试看效果怎么样。

Trystan-SA/claude-design-system-prompt

Source: https://github.com/Trystan-SA/claude-design-system-prompt

Claude Design System Prompt

Reverse-engineered system prompt of Claude Design from Anthropic.

A system prompt and skill library that turns an LLM into an opinionated, accessibility-aware, AI-slop-resistant design collaborator.

Open source, MIT licensed. Drop the prompt into any LLM that supports system prompts (Claude, GPT, Gemini, local models) and pair with the procedural skills as needed.

What this is

Most “design assistant” prompts produce generic SaaS-template output — aggressive gradients, emoji decoration, rounded-corner-with-left-border cards, Inter-everywhere typography. This prompt explicitly rejects those patterns and replaces them with a complete design philosophy covering:

• Content discipline (no filler — every element earns its place)
• Aesthetic discipline (avoid AI tropes, commit to a palette and tone)
• Visual hierarchy and rhythm (size, color, weight, position, density, spacing scales)
• Accessibility (WCAG, semantic HTML, keyboard navigation, focus rings, motion preferences)
• Interaction and feedback (hover, active, disabled, focus, loading, validation states)
• System thinking (components and tokens over one-off pages)
• Respecting the medium (real CSS Grid, oklch(), text-wrap: pretty, real interactive prototypes)
• Quality over quantity (depth over breadth, polish every detail)

Plus 14 procedural skills the agent can invoke for production, extraction, and review work.

What’s included

claude-design-system-prompt/
├── claude/                              Claude Code / Claude.ai variant
│   ├── system-prompt.md                 Main system prompt — 20 chapters
│   └── skills/                          14 invokable skills
│       ├── discovery-questions.md       Kickoff question protocol
│       ├── frontend-aesthetic-direction.md  Commit to a look when no brand exists
│       ├── wireframe.md                 Low-fi exploration, 3+ variations
│       ├── make-a-deck.md               Slide presentations in HTML
│       ├── make-a-prototype.md          Interactive clickable prototype
│       ├── make-tweakable.md            Floating tweak panel
│       ├── generate-variations.md       3+ hi-fi variations across axes
│       ├── design-system-extract.md     Pull tokens from sources
│       ├── component-extract.md         Inventory reusable components
│       ├── accessibility-audit.md       WCAG, semantic, keyboard, motion
│       ├── ai-slop-check.md             Gradient / emoji / font / house-style trope detection
│       ├── hierarchy-rhythm-review.md   Size / weight / color + spacing scale
│       ├── interaction-states-pass.md   Hover / active / disabled / focus / loading
│       └── polish-pass.md               Umbrella final-gate review
├── codex/                               OpenAI Codex variant (single-loop, no subagents)
│   ├── AGENTS.md                        Codex auto-discovered entry point
│   ├── system-prompt.md                 Same prompt, adapted for Codex
│   └── skills/                          Same skills, sequential reviews instead of parallel agents
├── README.md                            This file
└── LICENSE                              MIT

How to use it

Use the system prompt directly

Paste the contents of system-prompt.md as the system prompt for any LLM that supports them. The agent will follow the design philosophy and reference the skills by name when tasks match.

Use the skills as procedures

Each skill in skills/ is a self-contained, phased procedure. The skill name is the trigger — when the user’s request matches a skill description, the agent loads that skill and follows it.

Skills group into three categories:

Production — build something discovery-questions · frontend-aesthetic-direction · wireframe · make-a-deck · make-a-prototype · make-tweakable · generate-variations

System — extract structure design-system-extract · component-extract

Review — audit and fix accessibility-audit · ai-slop-check · hierarchy-rhythm-review · interaction-states-pass · polish-pass

Skills can be chained. A typical greenfield flow:

discovery-questions → frontend-aesthetic-direction → wireframe → make-a-prototype → polish-pass

A brand-aware flow:

design-system-extract → generate-variations → make-tweakable → polish-pass

Adapt for your platform

The prompt assumes an HTML-output design environment (similar to Claude.ai’s design tool). If your target environment is different — a Figma plugin, a code-only assistant, a chat-only design coach — you’ll need to adjust the workflow chapters and tool references. The principles (chapters 5–16) translate to any medium.

Model calibration

The claude/ variant is calibrated for current Anthropic frontier models (Fable 5 and the Opus 4.7/4.8 lineage), which follow instructions more literally and need less aggressive prompting than earlier generations:

• Conditions instead of quotas. No “ask at least N questions”, no “CRITICAL: YOU MUST”. Current models treat quotas as literal contracts and over-trigger on them; the prompt states the conditions under which to act, plus an autonomy clause for minor decisions (pick a reasonable option and note it, rather than asking).
• Explicit triggers for skills and subagents. These models under-reach for optional capabilities by default, so every skill description states when to invoke it, and verifier delegation has an explicit trigger (“after every substantive visual change”).
• Coverage-first reviews. Review agents report everything with confidence/severity estimates and let the aggregation step filter. Current models follow “only report important issues” literally, which silently suppresses findings.
• House-style guard. The current models’ default aesthetic (cream background, serif display type, terracotta/amber accents) is detected by ai-slop-check (rule 9) and pre-empted by frontend-aesthetic-direction’s four-directions protocol. Sampling parameters (temperature) no longer exist on these models, so visual variety must come from explicit per-variation specs, not randomness.

On older models (Claude Opus/Sonnet 4.6 and earlier, or non-Anthropic models), the calmer phrasing may under-trigger — restore stronger imperative language if you see the model skipping question rounds or reviews. The codex/ variant is maintained separately and is unaffected by these notes.

Design principles, in short

The 20 chapters in system-prompt.md cover:

Contributing

Issues and PRs welcome. Particularly useful contributions:

• Additional review skills (e.g., copy review, motion review, dark-mode parity check)
• Adapted prompts for other environments (Figma, code-only, terminal-only)
• Real-world failure cases the prompt should defend against
• Translations of the prompt into other languages

Please keep the same operational tone and avoid bloating the prompt — every chapter should earn its place, the same standard the prompt holds the agent to.

License

MIT — see LICENSE.

You can use, modify, and distribute this prompt and skill library for any purpose, including commercial use. No attribution required, but appreciated.