Claude Design: The Industry-Disrupting System Prompt Leaked on GitHub

Anthropic is on an absolute tear this year.

A quick glance at Claude’s update log reveals a relentless timeline: On January 12, Claude Cowork launched—a desktop-grade Agent capable of direct file system interaction, moving beyond simple chatbots to become a true virtual colleague.

By February 5, Opus 4.6 arrived with a 1-million-token window and a record-breaking 14.5-hour task completion capacity. March saw the full release of Claude Memory and the maturation of the Cowork plugin ecosystem. Then came Claude Code Channels, allowing developers to issue commands via Telegram or Discord—send a text from your phone, and return to finished code.

The momentum continued into April with Claude Code updates focused on Agent Teams and multi-person parallel development. Just days ago, Opus 4.7 debuted, significantly boosting software engineering capabilities with the new “x-high effort” mode.

Then came April 17.

Anthropic unveiled Claude Design: a tool that generates prototypes, pitch decks, and one-pagers using natural language. Following the announcement, stock prices for design giants like Figma and Adobe took a noticeable dip.

Claude Design isn’t built for professional designers—it’s for founders and product managers who have the vision but haven’t mastered Figma. As TechCrunch put it: “Anthropic launches Claude Design, a new product for creating quick visuals.”

Less than 24 hours post-launch, security researcher Pliny the Liberator leaked the full system prompt on the GitHub “CL4R1T4S” repository. At over 3,000 words, the prompt is a masterclass in AI instruction.

01 Deconstructing the Prompt

I. The Persona: “Expert Designer” vs. “Manager”

The prompt sets the tone immediately:

“You are an expert designer working with the user as a manager.”

This isn’t just an assistant helping you with a slide deck. It’s a senior designer you’ve hired. You provide the requirements; the AI handles the delivery. While the tool uses HTML, the outputs are versatile: interactive prototypes, animations, or structured documents.

Crucially, the prompt specifies: “Avoid web design tropes and conventions unless you are making a web page.” It distinguishes “Design” as a broader discipline than mere “Web Development.”

II. The “Anti-AI Slop” Checklist

The prompt includes a rigorous “blacklist” to avoid the generic “AI-generated” look:

  • No Gradient Abuse: Stop overusing flashy backgrounds.
  • No Emojis: Prohibited unless explicitly requested by branding.
  • No “Generic AI UI”: Bans the classic “rounded corners + left-border accent color” containers that scream “AI-generated.”
  • No Fake SVGs: If an asset doesn’t exist, use a placeholder. Don’t hallucinate messy SVG icons.
  • Font Discipline: Bans overused fonts like Inter, Roboto, and Arial.

The core principle: A clean placeholder is superior to a garbage implementation.

III. A Product System Within a Prompt

This is more than a prompt; it’s an engineering framework.

  • Starter Components: It includes pre-built device frames (iPhone shells, slide decks) so the AI doesn’t draw from scratch.
  • Engineering Rigor: It locks React and Babel versions via CDN with specific integrity hashes (e.g., react@18.3.1). This ensures rendering consistency.
  • The “Tweaks” Mechanism: A built-in panel allowing users to adjust colors, fonts, and spacing in real-time, which then persists into the code.
  • Two-Stage Verification: It uses a done command to check for console errors, followed by a fork_verifier_agent—a sub-agent that takes screenshots in an independent iframe to inspect layout issues.
  • Context Management: A snip tool allows the AI to discard unnecessary history, preventing context overflow during long design sessions.

IV. “Good Design Does Not Start From Scratch”

The prompt emphasizes that high-fidelity design must be rooted in existing context (UI Kits, brand guidelines, existing code). It explicitly states:

“Mocking a full product from scratch is a LAST RESORT and will lead to poor design.”

The AI is instructed to ask at least 10 clarifying questions regarding audience, constraints, and style before starting. It is designed to produce at least three variations—from conservative to creative—for the user to mix and match.

02 The Real Moat: Methodology, Not Technology

In traditional tech companies, the workflow is linear: PM writes a PRD → Designer creates mocks → Review → Scheduling → Dev → Test → Launch. This cycle takes 2–4 weeks at minimum.

Boris Cherny, head of Claude Code, recently shared on Lenny’s Newsletter that his team operates differently. They don’t start with PRDs. Instead, they build hundreds of functional prototypes using AI and then select what’s worth releasing. Boris himself merges 20–30 PRs a day while running five Claude instances simultaneously. The entire Cowork product was reportedly built in just 10 days.

At Anthropic, internal communication likely looks very different. Employees likely use personal Agents that maintain their specific context, allowing Agent-to-Agent synchronization to handle the “alignment” that usually bogs down human teams.

In most large corporations, the most exhausting part isn’t coding; it’s the constant shifting of direction and the endless meetings to align stakeholders. Boris bypasses this by giving his team “limited budget but infinite tokens,” forcing them to use AI as a force multiplier rather than hiring more people.

The core competitiveness of Anthropic isn’t just the Claude Design feature or the Opus 4.7 model. It’s that they have discovered the product iteration methodology for the AI era.