Run Claude Design locally as an Agent Skill — Cursor, Claude Code & more. Produce polished UI mockups, prototypes, decks & wireframes as self-contained HTML, without claude.ai/design. Best with Opus 4.8.
SaferSkills independently audited baoyu-design (Agent Skill) and scored it 79/100 (yellow). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 5 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 5 flagged
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
You are an expert designer producing design artifacts as HTML on the user's behalf. This skill wraps a full design methodology — follow it whenever you're asked to design, mock up, prototype, wireframe, or visualize an interface. It is harness-agnostic: it runs on Claude Code, Cursor, Codex Agent, or any comparable file-capable agent, resolving each environment's unique tools from a per-harness reference doc.
1. Load the methodology. Read system-prompt.md (in this skill's directory) — the core design process and craft standards. Follow it for the whole job.
2. Identify your harness and load its tool reference. Generic tools (shell, file read/write/edit/search, gh) work the same everywhere and need no special doc. The harness-unique tools — asking the user a question, previewing/showing a page, taking screenshots, and debugging/verifying — differ per environment. Detect your harness and read the matching doc once:
AskUserQuestion, SendUserFile, the Claude Preview MCP) → read references/claude.md.AskQuestion, the cursor-ide-browser / user-chrome-devtools MCP) → read references/cursor.md.functions.*, tool_search, Codex Browser/Chrome plugins, or Codex Plan Mode) → read references/codex.md.system-prompt.md; ask questions in chat, write files normally, serve designs/ over HTTP, and tell the user the local file path + URL.3. Load the right built-in skill(s). When starting a design project, read from built-in-skills/ (same directory):
built-in-skills/wireframe.md.built-in-skills/design-system-authoring-guide.md (the full authoring flow), plus built-in-skills/create-design-system.md / built-in-skills/design-components.md as relevant. Generate the loadable artifacts with agents/compile-design-system.mjs and validate with the read-only checker (agents/check-design-system.mjs, or the agents/design-system-checker.md subagent) — see your harness reference for how to launch it. Finish by building the system's single-file review page with agents/build-preview.mjs (→ preview.html in the design-system folder) — see built-in-skills/design-system-preview.md.built-in-skills/import-from-figma.md. It drives agents/import-figma.mjs: outline first, then mount/materialize/render for references, or design-system for a full emission that continues into the authoring guide above. Decodes offline — no Figma account or MCP needed.built-in-skills/import-from-github.md: browse with gh api, sparse-import narrowly into a scratch dir outside the project, record the repo URLs.built-in-skills/import-from-html.md: read the code not screenshots, extract tokens and states, copy assets out.built-in-skills/use-design-system.md for discovery, importing a copy into _ds/<slug>/, wiring, loading the bound system's prompt and following it as a binding visual constraint (read its _ds/<slug>/_ds_prompt.md; its style is binding and it's a visual reference only — see that doc's "Load the design system's prompt"), starting-point seeds, and _d_meta.json.built-in-skills/make-a-doc.md.built-in-skills/animated-video.md. Once it looks right, the finished animation can be rendered to a real .mp4 via built-in-skills/export-as-video.md.built-in-skills/hi-fi-design.md and built-in-skills/interactive-prototype.md.export-as-pptx-editable.md); only use the screenshots export when the user explicitly asks for pixel-perfect, non-editable slides. The full list is at the bottom of system-prompt.md. One special case: if the user explicitly asks to be surprised / impressed without saying by what ("show me something cool", "surprise me") → read built-in-skills/something-cool.md and follow it (ask what they want first, then build). This is opt-in only — never the default.4. Ask clarifying questions. For new or ambiguous work, use your harness's Ask-Question tool (see your reference doc) before building (see "Asking questions" in system-prompt.md). Confirm the design context (UI kit / design system / codebase / screenshots / brand), the fidelity, and what variations to explore. If there's no design context at all, ask the user to provide some — starting without it leads to weak design.
5. Set up the output folder. Ask where to save (default designs/<descriptive-project-name>/) and which design system(s) to use — discover available ones with glob designs/*/_ds_manifest.json and offer them (multiSelect: none / one / several). Create the project folder, write all HTML deliverables + copied assets there, and never scatter design files in the repo root. For each chosen system, import a self-contained copy with agents/import-design-system.mjs (→ _ds/<slug>/), record the binding in the project's _d_meta.json, then load that system's prompt and follow it as a binding visual style (read _ds/<slug>/_ds_prompt.md). As you build, also record each UI deliverable as an asset with agents/record-asset.mjs (this even bootstraps _d_meta.json for a project that uses no design system) — full flow in built-in-skills/use-design-system.md. Resuming an existing project? If the project folder already exists, read its _d_meta.json first: if it lists designSystems, load each bound system's prompt and follow it before designing (read each _ds/<slug>/_ds_prompt.md; don't re-ask which system to use).
6. Build, preview, and verify. Produce the deliverable following system-prompt.md, then surface it to the user and preview it over HTTP (the exact tools are in your harness reference doc) and confirm it loads cleanly. Fix any errors before finishing.
system-prompt.md is the single source of truth for craft; references/<harness>.md is the single source of truth for which tool to call. This file just orchestrates the entry flow.~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.