- What: Vercel introduced
design.md, a public system and markdown file to guide AI agents in building on-brand web pages. - Why: To solve the problem of subjective AI interpretations of design language and ensure consistency across external tools.
- How: A three-part system consisting of a markdown guidance file, a public stylesheet of design tokens, and a human-in-the-loop evaluation process.
- Availability: The
design.mdfile is a public resource accessible via a single URL for any AI agent to load.
Vercel, the frontend cloud platform, has announced the implementation of design.md, a new framework designed to teach AI agents how to maintain brand-specific design standards when building web pages. The system addresses the inherent subjectivity of large language models (LLMs) by providing a centralized, public repository of design judgment, typography, and structural guidelines.
According to an announcement from the Vercel team, the initiative stems from a need to maintain the "Vercel look and feel" even when using AI agents in environments outside of the company’s internal codebases. By moving beyond simple prompting and into a structured evaluation-based system, Vercel aims to ensure that AI-generated reports, proposals, and presentations adhere to the same rigorous design standards as their primary product offerings.
The Shift from Internal Skills to Public Standards
Vercel has long utilized "coding agents" to assist in the design and construction of its digital properties. Previously, the company relied on a tool called product-design, a specialized skill that lived within individual repositories. This allowed agents to reference existing code, components, and guidelines locally.
However, Vercel identified a significant gap: AI agents are increasingly used to generate "one-off" pages, such as renewal proposals or security governance briefs, in tools that lack access to private repositories. To bridge this gap, Vercel developed design.md as a single, public URL that any agent can load, regardless of the environment in which it operates.
The development of design.md was driven by two primary requirements. First, the guidance had to be accessible via a universal public link. Second, it needed to cover the full spectrum of Vercel’s design philosophy, encompassing brand identity, layout composition, copywriting, responsiveness, and information architecture.
Overcoming the Subjectivity of AI Design
During the early stages of development, Vercel discovered that a "naive approach"—simply porting internal prompts into a public file—was insufficient. The company found that while prompts could describe visual language, every model interpreted those descriptions differently.
A central challenge identified by the team was the ambiguity of design language. For instance, instructions like "keep the layout clean" were interpreted in vastly different ways by different LLMs. Furthermore, because these agents were working outside of Vercel’s codebase, they lacked the context of real, "shipped" components that usually provide a baseline for visual style.
To solve this, Vercel shifted from descriptive prompts to a system that distills design judgment into a repeatable format. The team began writing a new file from scratch, testing every iteration against a set of seven "eval prompts" based on real-world use cases:
- Usage and performance reports
- Renewal proposals
- Benchmark reports
- Interactive planning pages
- Build-versus-buy briefs
- Security governance briefs
- Presentation decks
By keeping these prompts fixed while iterating on the design.md file, Vercel could ensure that any improvements in the output were directly attributable to the guidance provided in the markdown file.
The Three-Part System for Design Judgment
The finalized design.md framework functions as a three-part system that provides a comprehensive roadmap for AI agents.
First, the markdown file (design.md) itself provides the core guidance. It instructs agents on how to frame the reader's job, structure evidence, and choose an appropriate composition. This layer focuses on "judgment," helping agents decide whether to prioritize a quick executive summary or a detailed audit. It also mandates the use of concrete claims in copywriting, moving away from generic marketing language.
Second, Vercel utilizes a public stylesheet. This file defines a bounded vocabulary of classes and tokens. By limiting the agent’s choices to a specific set of design tokens, Vercel minimizes the risk of the model hallucinating non-standard styles or layouts.
Third, the company implemented an evaluation loop. This involves continuous human feedback that is fed back into the system to refine the guidance and create "deterministic checks." This loop ensures that as design trends or brand standards evolve, the AI agents can be updated systematically.
Comparative Results and Technical Impact
Vercel’s testing suggests that the inclusion of design.md drastically alters model behavior. In a side-by-side comparison using a renewal proposal evaluation, a model running without design.md generated a generic SaaS dashboard.
In contrast, the same model with design.md loaded produced a page with a refined hierarchy. The AI-generated page led with the renewal recommendation, organized commercial evidence into a cohesive grid, and placed peer values on a single scale for easier comparison. This test confirmed that the system influences not only the styling of a page but its fundamental structure and information hierarchy.
Impact on the AI Design Industry
The introduction of design.md has immediate implications for developers and design teams seeking to maintain brand integrity across an increasingly fragmented AI toolset. By codifying design "judgment" into a machine-readable format, Vercel provides a blueprint for how organizations can prevent "brand drift" when using generative AI.
For developers, this approach offers a way to leverage the speed of AI agents without the need for constant manual oversight or redesigning. For the broader industry, it marks a shift toward "design-as-code" where the aesthetic and structural rules of a brand are treated with the same version control and evaluation rigor as software.
What’s Next
Vercel has made design.md a public resource, allowing it to be integrated into various AI workflows. The company stated that it will continue to build out the guidance "one rule at a time," using its evaluation loop to further refine the system. No specific timeline for future iterations or expanded feature sets was provided in the announcement.