Teaching agents product design at Vercel
Coding agents can produce working UI fast, but what's harder is a different shape. They can copy your product's style, match its patterns, and try to follow its conventions. What they cannot do is understand why those patterns exist. Code shows agents what shipped, not why one component, phrase, or interaction became your standard. That reasoning lives in design reviews, PR comments, Slack threads, and with the people who were in the room. For an agent, context that isn't in the codebase doesn't exist.
Vercel is an agent-native team. We treat accepted product decisions like code, keeping them in the repository, reviewing changes against them, and making them available to every agent working there.
The way we do this is through product-design. It's a system with three parts:
An agent skill that gives coding agents the context behind decisions that require product or codebase judgment.
Linters that enforce clear rules automatically.
A review loop that gathers evidence from Slack, Figma, and GitHub, then prepares guideline updates for review.
Any team can build the same structure around their own standards.
Inside the product-design skill
The skill lives inside the repository alongside the code it governs. Here's a simplified view of its structure:
repository/├── AGENTS.md├── .agents/│ └── skills/│ └── product-design/│ ├── AGENTS.md│ ├── SKILL.md│ ├── references/│ │ ├── product-judgment.md│ │ ├── interface-quality.md│ │ ├── resilience.md│ │ ├── surfaces.md│ │ ├── surfaces-{surface}.md│ │ ├── copy.md│ │ ├── rules.md│ │ ├── glossary.md│ │ ├── patterns.md│ │ └── coverage-gaps.md│ └── exemplars/│ └── pr-{name}.md└── tooling/ └── scripts/ └── evals/ ├── fixtures.json ├── rules-checklist.json └── <fixture>/ ├── before/ └── after/The product-design skill structure inside the repository.
The repository AGENTS.md tells coding agents when to load the skill. The skill-local AGENTS.md defines load order, validation, and governance. SKILL.md owns the runtime workflow.
references/ stores product-judgment, interface-quality, resilience, copy, canonical product names, interaction patterns, and surface-specific decisions.
exemplars/ documents decisions worth repeating from shipped pull requests, along with mistakes to avoid. coverage-gaps.md lists areas where we do not have a standard yet.
copywriting-eval/ tests copy and interface-language behavior. It does not evaluate the broader product-design workflow.
How the skill routes
SKILL.md resolves the request mode first: shape, implement, review, copy, or harden. This keeps audits from becoming edits and copy passes from expanding into redesigns. It skips backend-only work, telemetry, console errors, generated files, and tests with no shipped UI impact.
The skill routes to canonical sources instead of duplicating them. Component APIs, design-system rules, accessibility criteria, and interaction guidance stay with their owners.
Routing is specific to both task and surface. Material changes load product-judgment and interface-quality first. Copy, component, layout, interaction, accessibility, and resilience work each route to focused references. A modal loads destructive-action patterns and canonical verbs. A settings form loads labels, validation, progressive disclosure, and accessible-name guidance.
You can use this simplified structure as a starting point and replace the paths and standards with your own:
1---2name: product-design3description: >-4 Single entry point for product design and user-facing product implementation5 in apps/vercel-site. Use whenever work changes what a user sees, understands,6 chooses, or does: shaping requirements and flows; building or redesigning7 pages and components; reviewing URLs, screenshots, diffs, or Vercel Agent8 findings; improving product copy, information architecture, component choice,9 Geist compliance, hierarchy, layout, interaction, accessibility, responsive10 behavior, and loading, empty, error, permission, billing, or destructive11 states. Trigger on design, UX, UI, usability, flow, onboarding, settings,12 dashboard, build, improve, fix, audit, review, polish, simplify, or13 production-ready requests. Also use when backend behavior changes a14 user-visible outcome. Not for backend-only work with no user-visible effect,15 tests with no shipped UI impact, telemetry-only work, documentation, or16 marketing content.17---18
19# Vercel Product Design20
21Make the interface correct for the user, the product, and Vercel. Working code is not enough: choose the right interaction, make scope and consequences clear, cover reality beyond the happy path, and verify the rendered result.22
23## Operating Contract24
25- **Start with the job, not the pixels.** Identify who is acting, what they are trying to accomplish, the product object involved, and what the system will change.26- **Define the outcome before the output.** Establish the current user problem, desired behavior, success signal, and non-goals before choosing a surface or component.27- **Use evidence, not taste.** Trace decisions to product behavior, canonical repository guidance, an accepted design decision, or a verified adjacent pattern.28- **Separate facts from decisions.** Mark assumptions and unresolved product choices explicitly; do not hide them inside implementation details.29- **Treat shipped code as evidence, not automatic precedent.** It proves what exists, not why it is correct. Check it against current components, product behavior, and explicit guidance.30- **Choose the smallest coherent intervention.** Consider better defaults, behavior, or reuse before adding UI. Do not solve one job by creating unrelated settings or abstractions.31- **Decide before decorating.** Resolve information architecture, component semantics, interaction, and state behavior before styling or rewriting copy.32- **Design every reachable state.** Include only states the product can actually enter, but do not stop at the populated success case.33- **Verify the real surface.** Source inspection establishes behavior; a rendered interface establishes visual and interaction quality. Never claim visual verification from code alone.34- **Keep one user-facing entry point.** Invoke `product-design`; route internally to the canonical sources below.35
36## Request Modes37
38Resolve the mode from the user's verb and artifact before acting.39
40| Mode | Typical request | Required behavior |41| --------- | ---------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- |42| Shape | "Design this flow", "How should this work?", feature brief without settled UI | Frame the problem and evidence, compare material alternatives, then define the flow, states, acceptance criteria, risks, and open decisions. Do not edit unless asked. |43| Implement | "Build", "fix", "improve", "make compliant", or "run product-design on everything" | Resolve material product decisions, then implement the smallest coherent end-to-end change within scope. Do not absorb unrelated review findings. |44| Review | "Audit", "critique", "what's wrong?", code review | Inspect source and rendered evidence, then report prioritized findings. Do not edit unless asked. |45| Copy | "Fix the copy", "rewrite these errors" | Edit user-facing language, accessible names, and directly required JSX only. Report structural blockers without silently broadening scope. |46| Harden | "Polish", "production-ready", "handle edge cases" | Preserve the settled product direction while fixing state, resilience, responsive, accessibility, and finish defects. |47
48When intent is ambiguous, use the narrowest mode supported by the verb. A URL, screenshot, route, or component identifies scope; it does not by itself authorize edits.49
50A material decision changes the user's task, default, scope, consequence, navigation, interaction surface, or reachable states. Copy mechanics, token replacement, and established component substitutions usually are not material.51
52## Decision Authority53
54Resolve conflicts in this order:55
561. The user's explicit goal and constraints.572. Verified user/product evidence and system truth.583. Repository-canonical guidance: `AGENTS.md`, Geist component APIs, `packages/geist/STYLE_GUIDE.md`, and routed skills.594. Accepted product/design decisions and exemplars with stable evidence.605. Verified adjacent shipped patterns in the same product area.616. General interface heuristics.62
63## Workflow64
65### 1. Set scope and mode66
67Name the target surface and request mode in the work plan or review notes.68
69### 2. Load product context70
71Before proposing UI, read the applicable `AGENTS.md` chain, supplied briefs and designs, and the product logic that determines mutations, permissions, validation, errors, and side effects.72
73### 3. Model the product decision74
75For Shape, Implement, Harden, full Review, or any material product/flow change, read `product-judgment.md` and write a compact internal brief covering user, job, current behavior, desired outcome, success signal, non-goals, object, scope, action, consequence, reversibility, permissions, and open decisions.76
77### 4. Map the surface and states78
79Inventory entry points, visible regions, overlays, transitions, exits, and return paths. Map only reachable states including loading, empty, sparse, populated, validation, error, permission, disabled, optimistic, stale, destructive, and responsive variants.80
81### 5. Load the routed references82
83| Need | Load |84| ---- | ---- |85| Product/flow/component decision | `product-judgment.md` + `component-guide` |86| Implementation, material visual change, or full review | `interface-quality.md` |87| Copy or accessible names | `copy.md` + `surfaces.md` routing |88| Layout, typography, color, spacing, Geist APIs | `design-guidelines` + `packages/geist/STYLE_GUIDE.md` |89| Keyboard, focus, forms, touch, animation, URL state, performance | `web-interface-guidelines` |90| Overflow, localization, extreme data, network/error resilience | `resilience.md` |91
92### 6. Decide, then implement93
94For each non-mechanical change, be able to answer: what user problem does this solve, why is this component appropriate, what consequence must the interface communicate, which evidence supports the decision, and what is the smallest coherent change?95
96### 7. Verify97
981. Confirm the primary job and acceptance criteria.992. Run repository lint checks.1003. Inspect relevant compact and wide viewports.1014. Exercise every materially changed reachable state.1025. Verify keyboard order, focus movement, loading behavior, and pointer/touch targets.1036. Test long content, large values, constrained width, and localization/RTL risk.1047. Load `review-design-system` for structural visible changes.105
106## Product Design Standards107
108- Make the user's primary task and primary action unmistakable.109- Preserve the user's mental model and current context unless changing it solves a verified problem.110- Name the exact object, scope, and consequence of important actions.111- Use navigation components for navigation and action components for actions.112- Choose surface persistence to match importance.113- Prefer inline disclosure before adding a modal.114- Expose advanced controls when needed without making the default path carry their complexity.115- Prefer strong defaults and direct behavior over adding configuration the user must learn and maintain.116- Use semantic Geist components and their APIs before custom HTML or styling.117- Use hierarchy, spacing, and alignment before adding containers.118- Preserve user input through validation and recoverable errors.119- Keep loading control labels stable; use the component's loading/busy affordance.120- Make destructive actions proportional to impact and provide undo when the system can honestly support it.121- Do not add decorative novelty, motion, or copy unless it clarifies structure, state, or brand intent.122
123## Review Output124
125Lead with findings, ordered by user impact:126
127- **P0:** blocks the primary task, creates severe accessibility failure, or can cause unrecoverable user harm.128- **P1:** likely task failure, misleading consequence, missing critical state, or major responsive/accessibility defect.129- **P2:** meaningful friction, inconsistency, weak hierarchy, or recoverability issue.130- **P3:** minor craft or consistency improvement.131
132For each finding include: file/line or rendered location, verification status, canonical source, user consequence, and smallest concrete fix.133
134## Skill Integrity135
136- Add or change a rule only after current-source verification and human acceptance.137- Record scope, rationale, evidence, exceptions, and a bad/good example.138- Prefer the narrowest destination: canonical source, routed reference, exemplar, lint/eval check, or coverage gap.139- Keep deterministic checks mechanical. Keep judgment in prose with its evidence and degree of freedom.140- Never promote one screenshot, one shipped file, or one reviewer comment into a universal rule by itself.The product-design SKILL.md. Routing modes, operating contract, and governance.
Routing is only part of what makes the skill useful. The other part is how findings stay traceable once the skill produces them.
Make findings traceable
Copy rules have stable IDs and point to their canonical sources: