Claude Skill

frontend-design

Imported from paulrberg/agent-skills/skills/frontend-design.

LLM Mart · 0 points · 0 views 0 listing impressions 0 install-command copies
Virus-scanned Reviewed automatically before listing.

Full trust report

Download paulrberg-agent-skills-skills_frontend-design-913232a.zip · 3 KB
Part of paulrberg/agent-skills — 42 skills

Install

skills CLI npx skills add https://github.com/PaulRBerg/agent-skills/tree/main/skills/frontend-design
Claude Code claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install paulrberg-agent-skills@llmmart
Git git clone https://github.com/PaulRBerg/agent-skills.git

The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole paulrberg/agent-skills collection as a plugin from our marketplace. Git is the plain clone.

Skill manifest

Frontend Design

Create a working frontend with a clear, subject-specific point of view, then prove it in the rendered UI.

Contract

  • For a critique, review, or design plan, inspect the relevant product and report the direction without editing files.
  • For a build or redesign, make the requested in-scope changes and run local, non-destructive validation.
  • Treat explicit user direction, supplied references, repository instructions, existing product behavior, and the local design system as authoritative. Fidelity requests override this skill's defaults.
  • Preserve the stack and working behavior. Do not replace the framework or design system, add dependencies, invent features, or fetch or generate media unless the request requires it.
  • Make the smallest defensible assumption when the subject, audience, or page job is missing. State it before building; ask only when the answer would materially change scope or identity.

Workflow

1. Read the product before drawing the page

Inspect the affected route or component, neighboring UI, tokens, typography, assets, dependencies, and repository validation commands. Identify:

  • the concrete subject and audience;
  • the screen's single primary job and content hierarchy;
  • the existing visual language and interaction conventions;
  • the required viewports, themes, states, and accessibility constraints.

Use real product content and supplied assets wherever possible. If copy is missing, write only what the user needs to understand the interface and act.

2. Set one art direction

Before coding, define a compact direction:

  • Thesis: one sentence connecting the visual idea to the subject and screen job.
  • System: role-based color tokens, explicit type roles and scale, spacing and density, and a layout principle.
  • Signature: one memorable element drawn from the domain's objects, workflows, data shapes, environments, or language.
  • Risk: one deliberate departure from the obvious solution, with a reason it serves this brief.
  • Motion and copy: when motion helps, how the interface speaks, and where restraint matters.

Adapt this direction to existing brand constraints instead of creating a parallel design language. Keep options and process narration internal unless the user asked to choose among directions.

3. Run the anti-default critique

Apply the subject-swap test: mentally replace the subject with an unrelated one. Any major choice that still works unchanged is probably a template default; revise it or justify why it is structurally correct.

Make every visual device earn its place. Cards, pills, gradients, numbering, dividers, oversized type, split heroes, illustrations, and animation must communicate hierarchy, function, sequence, or subject—not merely decorate. Concentrate expressive force in the signature and make the surrounding system support it.

Name concrete patterns to avoid; generic "avoid an AI look" advice only swaps one default for another. Without a brief that calls for them, avoid these known model defaults: cream or off-white backgrounds, italic accent words in headlines, numbered 01/02/03 section labels, monospace labels, and pill-shaped buttons. After the first render, name any other default the result fell back on and revise it.

4. Build the actual experience

  • Follow local components, tokens, styling architecture, and dependency versions before introducing new primitives.
  • Match composition to use: operational tools favor scannable information and efficient repeated actions; editorial or marketing surfaces may use a more expressive narrative.
  • Make the first viewport establish both identity and primary purpose. For an application or tool, show the real working experience rather than wrapping it in an unrequested landing page.
  • Use structural devices only when they encode real relationships. Use cards for genuinely grouped or repeated objects, not as the default wrapper for every section, and avoid nested cards.
  • Give typography deliberate display, body, and utility roles. Use available fonts first; do not add a font dependency solely to manufacture novelty.
  • Write interface copy from the user's side: specific nouns, active verbs, consistent action names, sentence case, and useful empty and error states. A control's label must describe its result.
  • Use familiar controls and the project's icon library. Give unfamiliar icon-only controls accessible names and visible tooltips where appropriate.
  • Keep motion purposeful and concentrated. Respect reduced-motion preferences and ensure the experience remains clear without animation.
  • Build responsive constraints from content rather than device labels. Prevent overlap, clipping, unreadable wrapping, and layout shift at narrow and wide widths.
  • Preserve semantic structure, keyboard operation, visible focus, adequate contrast, and target sizes. Match the project's accessibility standard when it is stricter.
  • Keep selector specificity and component ownership predictable; do not rely on competing selectors or fragile cascade order to establish spacing and state.

5. Render, inspect, and revise

Run repository-required checks and the narrowest additional checks that prove the changed behavior; select among formatting, lint, typecheck, tests, and build according to the affected surface. Reuse passing evidence until edits or unresolved concerns invalidate it. Then do rendered verification with the chromium-browser skill when it is available in this session; otherwise use the host's DevTools/browser tool. Do not fall back to Computer Use or ad-hoc Playwright scripts while a DevTools browser tool is available.

Scale the inspection matrix to the change. Repository instructions may reduce the viewport/state matrix, including waiving multi-viewport checks; follow them. Absent such instructions, a small edit to existing UI needs verification only of the changed states at one representative viewport, while a new or substantially redesigned surface needs representative narrow and wide viewports and every changed interaction and state. Behavior belongs in the repository's automated tests where they exist; use the browser for visual, layout, and theme questions.

Check the rendered result for content hierarchy, subject specificity, asset loading, overflow, overlap, truncation, contrast, focus, hover, motion, empty/error states, and theme variants in scope. Compare it with the brief and the art direction. Fix visible defects and remove any element that does not materially serve either.

Do not claim visual verification from source inspection alone when rendering is available. If the environment cannot render the UI, state that limitation and report exactly what was verified instead.

Completion

Completion requires the requested artifact or code, a subject-specific direction reflected in the implementation, passing relevant local checks, and rendered inspection evidence when tooling permits. Report the direction in one sentence, the checks and viewports exercised, and any remaining limitation.

Finish builds with ### ✨ Built: <surface>, 🎨 Direction — <one sentence>, and ### 🧪 Verification; use a compact viewport/state/result table when several states were inspected and link screenshots or artifacts when available. Add ### ⚠️ Remaining only when non-empty. Agent-report decoration does not authorize emoji or ASCII ornament in shipped interface copy, snapshots, or source unless the brief or design system calls for it.

Files (agent-skills)
  • agents
    • openai.yaml 42 B
      policy:
        allow_implicit_invocation: true
      
  • SKILL.md 7.9 KB
    ---
    compatibility:
      Designed for Codex and Claude Code; prefer the chromium-browser skill for rendered verification when available,
      otherwise the host's DevTools/browser tool.
    name: frontend-design
    skill-dependencies:
      - chromium-browser
    description:
      Use when creating or substantially redesigning web interfaces, landing pages, dashboards, components, or other
      frontend UI where visual direction and implementation quality matter. Produces subject-specific art direction,
      accessible responsive code, and rendered visual verification.
    ---
    
    # Frontend Design
    
    Create a working frontend with a clear, subject-specific point of view, then prove it in the rendered UI.
    
    ## Contract
    
    - For a critique, review, or design plan, inspect the relevant product and report the direction without editing files.
    - For a build or redesign, make the requested in-scope changes and run local, non-destructive validation.
    - Treat explicit user direction, supplied references, repository instructions, existing product behavior, and the local
      design system as authoritative. Fidelity requests override this skill's defaults.
    - Preserve the stack and working behavior. Do not replace the framework or design system, add dependencies, invent
      features, or fetch or generate media unless the request requires it.
    - Make the smallest defensible assumption when the subject, audience, or page job is missing. State it before building;
      ask only when the answer would materially change scope or identity.
    
    ## Workflow
    
    ### 1. Read the product before drawing the page
    
    Inspect the affected route or component, neighboring UI, tokens, typography, assets, dependencies, and repository
    validation commands. Identify:
    
    - the concrete subject and audience;
    - the screen's single primary job and content hierarchy;
    - the existing visual language and interaction conventions;
    - the required viewports, themes, states, and accessibility constraints.
    
    Use real product content and supplied assets wherever possible. If copy is missing, write only what the user needs to
    understand the interface and act.
    
    ### 2. Set one art direction
    
    Before coding, define a compact direction:
    
    - **Thesis:** one sentence connecting the visual idea to the subject and screen job.
    - **System:** role-based color tokens, explicit type roles and scale, spacing and density, and a layout principle.
    - **Signature:** one memorable element drawn from the domain's objects, workflows, data shapes, environments, or
      language.
    - **Risk:** one deliberate departure from the obvious solution, with a reason it serves this brief.
    - **Motion and copy:** when motion helps, how the interface speaks, and where restraint matters.
    
    Adapt this direction to existing brand constraints instead of creating a parallel design language. Keep options and
    process narration internal unless the user asked to choose among directions.
    
    ### 3. Run the anti-default critique
    
    Apply the **subject-swap test**: mentally replace the subject with an unrelated one. Any major choice that still works
    unchanged is probably a template default; revise it or justify why it is structurally correct.
    
    Make every visual device earn its place. Cards, pills, gradients, numbering, dividers, oversized type, split heroes,
    illustrations, and animation must communicate hierarchy, function, sequence, or subject—not merely decorate. Concentrate
    expressive force in the signature and make the surrounding system support it.
    
    Name concrete patterns to avoid; generic "avoid an AI look" advice only swaps one default for another. Without a brief
    that calls for them, avoid these known model defaults: cream or off-white backgrounds, italic accent words in headlines,
    numbered `01/02/03` section labels, monospace labels, and pill-shaped buttons. After the first render, name any other
    default the result fell back on and revise it.
    
    ### 4. Build the actual experience
    
    - Follow local components, tokens, styling architecture, and dependency versions before introducing new primitives.
    - Match composition to use: operational tools favor scannable information and efficient repeated actions; editorial or
      marketing surfaces may use a more expressive narrative.
    - Make the first viewport establish both identity and primary purpose. For an application or tool, show the real working
      experience rather than wrapping it in an unrequested landing page.
    - Use structural devices only when they encode real relationships. Use cards for genuinely grouped or repeated objects,
      not as the default wrapper for every section, and avoid nested cards.
    - Give typography deliberate display, body, and utility roles. Use available fonts first; do not add a font dependency
      solely to manufacture novelty.
    - Write interface copy from the user's side: specific nouns, active verbs, consistent action names, sentence case, and
      useful empty and error states. A control's label must describe its result.
    - Use familiar controls and the project's icon library. Give unfamiliar icon-only controls accessible names and visible
      tooltips where appropriate.
    - Keep motion purposeful and concentrated. Respect reduced-motion preferences and ensure the experience remains clear
      without animation.
    - Build responsive constraints from content rather than device labels. Prevent overlap, clipping, unreadable wrapping,
      and layout shift at narrow and wide widths.
    - Preserve semantic structure, keyboard operation, visible focus, adequate contrast, and target sizes. Match the
      project's accessibility standard when it is stricter.
    - Keep selector specificity and component ownership predictable; do not rely on competing selectors or fragile cascade
      order to establish spacing and state.
    
    ### 5. Render, inspect, and revise
    
    Run repository-required checks and the narrowest additional checks that prove the changed behavior; select among
    formatting, lint, typecheck, tests, and build according to the affected surface. Reuse passing evidence until edits or
    unresolved concerns invalidate it. Then do rendered verification with the chromium-browser skill when it is available in
    this session; otherwise use the host's DevTools/browser tool. Do not fall back to Computer Use or ad-hoc Playwright
    scripts while a DevTools browser tool is available.
    
    Scale the inspection matrix to the change. Repository instructions may reduce the viewport/state matrix, including
    waiving multi-viewport checks; follow them. Absent such instructions, a small edit to existing UI needs verification
    only of the changed states at one representative viewport, while a new or substantially redesigned surface needs
    representative narrow and wide viewports and every changed interaction and state. Behavior belongs in the repository's
    automated tests where they exist; use the browser for visual, layout, and theme questions.
    
    Check the rendered result for content hierarchy, subject specificity, asset loading, overflow, overlap, truncation,
    contrast, focus, hover, motion, empty/error states, and theme variants in scope. Compare it with the brief and the art
    direction. Fix visible defects and remove any element that does not materially serve either.
    
    Do not claim visual verification from source inspection alone when rendering is available. If the environment cannot
    render the UI, state that limitation and report exactly what was verified instead.
    
    ## Completion
    
    Completion requires the requested artifact or code, a subject-specific direction reflected in the implementation,
    passing relevant local checks, and rendered inspection evidence when tooling permits. Report the direction in one
    sentence, the checks and viewports exercised, and any remaining limitation.
    
    Finish builds with `### ✨ Built: <surface>`, `🎨 Direction — <one sentence>`, and `### 🧪 Verification`; use a compact
    viewport/state/result table when several states were inspected and link screenshots or artifacts when available. Add
    `### ⚠️ Remaining` only when non-empty. Agent-report decoration does not authorize emoji or ASCII ornament in shipped
    interface copy, snapshots, or source unless the brief or design system calls for it.
    

Comments (0)

Sign in to join the conversation.

No comments yet.

Reviews (0)

No reviews yet.

Related