Codifying Product Judgment

Coding agents often replicate existing UI patterns without understanding the underlying reasoning. Vercel addresses this by treating product decisions as first-class code, stored in the repository alongside implementation. The product-design skill acts as a single entry point for agents, providing context on interaction patterns, copy, and interface quality. This system relies on a strict operating contract that prioritizes evidence over taste, requiring agents to identify the user's job, desired outcome, and success signals before proposing UI changes.

The Three-Tiered Governance Model

Vercel uses a tiered approach to enforce design standards based on the complexity and determinism of the rule:

  1. Deterministic Linters: Rules that can be reliably checked (e.g., preferring radio buttons over selects for 2-3 options, or enforcing accessible names) are implemented as lint rules. These provide immediate, fast feedback to developers and agents.
  2. Agent Skills: For decisions requiring product context (e.g., destructive action patterns or state management), the product-design skill provides prose-based guidance, examples, and decision records.
  3. Human-in-the-Loop Evidence: A weekly workflow aggregates Slack discussions, Figma files, and PR comments into a review packet. Humans then decide which patterns should be promoted to agent guidance, lint rules, or documented as 'coverage gaps' where standards do not yet exist.

Implementation Strategy for Teams

To build a similar system, teams should start by identifying repeated review comments—such as those regarding error states or navigation—and documenting them using a structured template that includes rationale, evidence, and 'good/bad' examples. Crucially, teams must define explicit triggers in AGENTS.md to ensure the agent loads the skill at the right time. The system should separate routing (identifying the surface) from rules (the design standard) and evidence (the source of truth), ensuring that agents cite their sources when proposing changes. By keeping guidance current through a dedicated evidence-intake process, teams can scale their design standards without relying on tribal knowledge.