Webdesign

Voice Notification (REQUIRED FIRST ACTION)

curl -s -X POST http://localhost:31337/notify \
  -H "Content-Type: application/json" \
  -d '{"message": "Running the Webdesign skill", "voice_enabled": true}' > /dev/null

What It Does

Designs and integrates web interfaces via two paths. ClaudeDesign drives Anthropic's claude.ai/design product through the Interceptor skill, then folds the result into your codebase. DirectDesign writes the design inline using Anthropic's open-source frontend-design philosophy, loaded from a local reference. Default routing: short, ad-hoc, in-codebase work goes to DirectDesign; multi-page or shareable prototypes go to ClaudeDesign.

The Problem

Producing good web UI usually means either fighting a visual tool that can't touch your real codebase, or hand-coding from scratch and landing on the same generic defaults every time. The two failure modes pull in opposite directions: visual-first tools give you polish but a handoff gap, while writing code directly gives you integration but flat aesthetics. This skill gives you both routes under one roof — a visual round-trip through claude.ai/design when review matters, or inline design with a real aesthetic doctrine loaded when speed and in-codebase iteration matter — and routes to the right one based on the ask.

How It Works

Webdesign covers three paths for producing web UI. Pick the path that fits; surface options by name when intent is ambiguous. We always drive whatever version of Claude Design is live — there is no version to pin.

Path 1 — DirectDesign (default workhorse; {{DA_NAME}} writes the design inline)

{{DA_NAME}} writes the design directly with Anthropic's open-source frontend-design aesthetic doctrine loaded inline. The load-bearing prompt content is mirrored from github.com/anthropics/skills/tree/main/skills/frontend-design (MIT-licensed) into References/FrontendDesignPhilosophy.md — register list, anti-default rules, type pair recipes, motion vocabulary, color discipline. Self-contained: no runtime dependency, no browser, no auth. This is the path that has actually shipped every real design this skill has produced. Workflow: DirectDesign.

Path 2 — Native Claude Design CLI (preferred for code integration + design-system sync)

Anthropic shipped first-party /design and /design-sync commands inside Claude Code (June 2026, GA on Pro/Max/Team/Enterprise — official: support.claude.com/en/articles/14604416). /design-sync pulls a codebase's real design system into Claude Design and pushes built changes back; /design creates and edits designs from the terminal. These are the deterministic, first-party replacement for the hand-rolled Interceptor handoff-bundle apparatus — prefer them for anything code-bound. Workflow: NativeDesignSync.

Path 3 — ClaudeDesign via Interceptor (EXPERIMENTAL visual-review fallback)

⚠️ Unverified and currently non-functional. This path drives the claude.ai/design web canvas through the Interceptor skill for visual-first review. It requires an authenticated claude.ai session in the interceptor-test Chrome profile, which is not currently logged in, and it has never been run end-to-end (every real run of this skill used DirectDesign). Use only when you specifically want the visual web canvas AND have set up the login first. Workflows: CreatePrototype, ExtractDesignSystem, RefinePrototype, WebsiteToRedesign, ExportToCode, IntegrateIntoApp, DeployDesign. Tool: DriveClaudeDesign.ts.

Routing rule

When the user asks for "a nice design" / "design something" without naming a path:

• Default to DirectDesign for short, ad-hoc, in-codebase work — speed and in-context iteration.
• Use the native CLI (/design-sync) when the job is to sync a real codebase design system or do code-bound design work and the commands are available.
• Reach for ClaudeDesign (Path 3) only when visual web-canvas review is explicitly wanted, and remember it needs the one-time login and is unproven.

Integration-Aware Operation (CRITICAL)

This skill is frequently called as a sub-step of larger site work — writing a blog post, building an admin dashboard, shipping a marketing page. When invoked from a parent context, the skill:

• Accepts existing-project context as input: framework, token file, component directory, deployment target.
• Produces output as diffs / patches against the existing app, not isolated HTML files.
• Respects existing design tokens and component patterns — does NOT overwrite them unless the user requests a full redesign.
• Routes integration work through Workflows/IntegrateIntoApp.md.

When invoked standalone for a greenfield design, the skill produces a self-contained prototype and optionally scaffolds a new app.

Customization

User-specific design preferences (color palette, typography, spacing grid, animation timing, framework defaults) live at:

~/.claude/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/Webdesign/
├── PREFERENCES.md     # Design tokens, preferred frameworks
├── README.md
└── EXTEND.yaml

The skill reads PREFERENCES.md if present and passes those tokens into Claude Design's brief and any downstream handoff bundle. Without a customization layer, the skill defaults to Claude Design's own system-extraction output.

Workflow Routing

When executing a workflow, output this notification:

Running **WorkflowName** in **Webdesign**...

For code-bound work (export, integration, design-system extraction/sync), prefer Path 2 (NativeDesignSync → /design-sync). The Path 3 bundle workflows above remain as a documented fallback for when you're working from the web canvas, but they are unproven and need the test-profile login.

Prerequisites (PREFLIGHT)

Path 1 (DirectDesign) needs nothing — no browser, no auth. Path 2 (native CLI) needs only a Claude subscription that includes Claude Design and a current Claude Code (/update if the commands don't show). The checks below apply ONLY to Path 3 (ClaudeDesign via Interceptor):

1. Interceptor skill available — which interceptor returns a path. If not, instruct user to invoke Skill("Interceptor") setup first.
2. Authenticated claude.ai session — the interceptor-test Chrome profile must be logged into claude.ai. Check at run time (an unauthenticated profile hits the marketing wall, not the app); if it isn't, a one-time headed login is required before any Path 3 workflow can run.
3. Claude Design access — the Claude subscription must include Claude Design (Pro, Max, Team, or Enterprise with admin opt-in).
4. For IntegrateIntoApp: parent-project path + framework identifier (next, astro, vitepress, vite-react, vue, vanilla) passed in context.

Missing prerequisites → halt with a clear remediation step. Never silently fall back.

Gotchas

Accumulate lessons here. Information density is highest in gotchas.

• Two paths, one skill. ClaudeDesign (CreatePrototype et al.) drives claude.ai/design; DirectDesign has {{DA_NAME}} write the design inline using References/FrontendDesignPhilosophy.md. Don't conflate them. When the user's intent is ambiguous, surface both as named choices and let them pick — never silently route.
• Native-first — code and tokens are the source of truth, Figma is not a dependency. Import a design system by linking the repo (Claude Design reads real components and tokens from code) or uploading token/component files, never a .fig export. Do NOT add a Figma round-trip in either direction (design → .fig, or code → editable Figma frames). The bet is that design lives in the codebase, not an external interchange file. If a no-repo visual-review need ever comes up, solve it with a URL/screenshot share, not by coupling the skill to Figma.
• DirectDesign is self-contained. The aesthetic doctrine lives in References/FrontendDesignPhilosophy.md (mirrored from anthropics/skills MIT source). No runtime dependency on the upstream frontend-design Claude Code plugin being installed — DirectDesign works in any LifeOS environment.
• Native /design + /design-sync are the code-bound path now (CONFIRMED). Anthropic's official docs (support.claude.com/en/articles/14604416) document both commands, GA on Pro/Max/Team/Enterprise as of June 2026. /design-sync does bidirectional codebase↔design-system sync — use it instead of the Interceptor handoff-bundle apparatus for anything code-bound. Still no public REST API or MCP server; the CLI commands are the programmatic surface. If they don't show up, run /update.
• The ClaudeDesign/Interceptor path (Path 3) is unproven and currently blocked. It has never run end-to-end — every real run of this skill used DirectDesign. The interceptor-test profile is not logged into claude.ai, so DriveClaudeDesign.ts reaches the marketing wall, not the app. The tool targets controls by accessibility-tree heuristics (composer by role=textbox/contenteditable, send by /send|submit/, export by /export/), so a moved button is NOT the blocker — the missing auth and the native-CLI supersession are. Before trusting Path 3: log the test profile into claude.ai once, then do a single supervised run.
• Real Chrome required. Use the Interceptor skill — it is the only sanctioned browser automation in LifeOS. Claude Design's UI depends on claude.ai's full session state; CDP-based automation trips bot detection and drops session cookies.
• Handoff bundles are directories, not single files. A bundle contains PROMPT.md, optional tokens.json, components/, assets/, and framework-specific scaffolding. Treat the whole directory as the unit.
• frontend-design plugin auto-activates. When the handoff bundle is fed to Claude Code, the plugin (already installed in the official marketplace) picks up the frontend work automatically — do NOT manually invoke it.
• Claude Design's design-system extraction runs during onboarding. For a new codebase you want Claude Design to understand, run ExtractDesignSystem FIRST before CreatePrototype — otherwise Claude Design uses generic defaults and overrides your tokens.
• Integration ≠ overwrite. IntegrateIntoApp produces diffs on top of existing code. If the user wants a full redesign that replaces existing UI, explicitly flag this and get confirmation.
• Canva exports are editable. If the user wants a non-developer (marketer, founder) to refine the design, route through Workflows/ExportToCode.md with --format canva.
• No real-time collab. Claude Design does not support multiplayer editing like Figma. Share via URL export for async review.
• Enterprise gate. Enterprise accounts need an admin to enable Claude Design in Organization settings before the palette icon appears in claude.ai.
• Session quotas. Claude Design generation is token-heavy. As of the June 2026 update its usage shares one pool with claude.ai chat, Claude Code, and Cowork — no longer a separate quota. Pro is thin for sustained design work; Max recommended.
• Design-system-first is the token fix. The biggest token sink is re-inferring your brand on every pass and then correcting it. Run ExtractDesignSystem once so the system is a fixed reusable reference; every later generation reuses it instead of guessing. Fewer correction cycles = far fewer tokens over a project's life. This is the single highest-leverage move against quota burn.
• Output fidelity ≠ production-ready. Claude Design produces polished visuals, but hand-off code often needs a verification + a11y pass. Run Tools/VerifyDesign.ts post-integration.
• Vision doesn't guess. If the prompt doesn't specify responsive breakpoints, contrast requirements, or dark-mode behavior, Claude Design picks defaults that may not match the target app. Be explicit in the brief.

Examples

Example 1: Create a prototype from a brief

User: "Design a pricing page for an AI security startup — editorial aesthetic, dark only"
→ Invokes CreatePrototype workflow
→ Preflight: Interceptor + authenticated claude.ai session
→ Composes brief with explicit aesthetic, constraints, differentiation
→ Drives claude.ai/design via Tools/DriveClaudeDesign.ts
→ Screenshots output, verifies a11y via Tools/VerifyDesign.ts
→ Returns bundle path + preview URL

Example 2: Land a Claude Design prototype inside an existing Astro app

User: "Integrate this prototype into ~/Projects/landing — it's an Astro site"
→ Invokes IntegrateIntoApp workflow
→ Audits target project (framework, tokens, components)
→ Runs ExtractDesignSystem first to prime Claude Design with app's real tokens
→ Translates prototype to Astro conventions via frontend-design plugin
→ Produces unified diff against the working tree
→ Pauses for human review before applying
→ Applies patch on a branch, runs tests, screenshots in-context

Example 3: Redesign an existing live site

User: "Redesign example.com — modernize, keep the copy, make it brutalist"
→ Invokes WebsiteToRedesign workflow
→ Captures current state (screenshot + HTML + tokens)
→ Writes critique (what works, what's dated, what to preserve)
→ Composes rebuild brief with explicit aesthetic and preserve list
→ Drives Claude Design with critique + original screenshot as input
→ Iterates via RefinePrototype until satisfied
→ Hands off to IntegrateIntoApp or ExportToCode

File Organization

skills/Webdesign/
├── SKILL.md                          # This file — routing + gotchas
├── README.md                         # Public-facing intro
├── Workflows/
│   ├── DirectDesign.md               # Path 2 — {{DA_NAME}} writes inline (no claude.ai round-trip)
│   ├── CreatePrototype.md            # Path 1 — drives claude.ai/design
│   ├── ExtractDesignSystem.md
│   ├── RefinePrototype.md
│   ├── WebsiteToRedesign.md
│   ├── ExportToCode.md
│   ├── IntegrateIntoApp.md
│   └── DeployDesign.md
├── Tools/
│   ├── DriveClaudeDesign.ts          # Interceptor wrapper for claude.ai/design
│   ├── ProcessHandoffBundle.ts       # Parse bundle → structured brief
│   └── VerifyDesign.ts               # Screenshot + a11y probe
└── References/
    ├── FrontendDesignPhilosophy.md   # Aesthetic doctrine — load-bearing for DirectDesign (MIT-attributed mirror)
    ├── ClaudeDesignCapabilities.md   # What Claude Design does / doesn't do
    ├── InputFormats.md               # Prompt patterns, codebase prep
    ├── ExportFormats.md              # html / pdf / pptx / canva / url / bundle
    └── HandoffBundleSpec.md          # Bundle structure for Claude Code handoff

Execution Log

{"ts":"ISO8601","workflow":"CreatePrototype","brief":"one-line","outputs":["path1","path2"],"duration_s":42}

This log is read-only metadata; it is not part of the public skill distribution.