mrgnhnt96/ultrahand

A Claude Code plugin with research-backed UX design skills for each screen size, built as interactive, resizable Claude artifacts, plus a UX review skill and a pair of theme skills that keep every design on one consistent theme.
## Install
Add ultrahand to a Claude Code marketplace you control by listing it in that
marketplace's .claude-plugin/marketplace.json:
json
"name": "ultrahand",
"source": { "source": "github", "repo": "mrgnhnt96/ultrahand" },
"description": "Research-backed UX design skills for mobile, tablet and desktop, plus a UX review skill."
Then install it (replace with that marketplace's name):
sh
claude plugin marketplace update
claude plugin install ultrahand@ --scope user
Installs take whatever is on main.
## Skills
| Skill | Use it for |
| ultrahand:mobile | Designing phone UI screens and flows, with mobile size presets |
| ultrahand:tablet | Designing tablet UI, including split view |
| ultrahand:desktop | Designing desktop UI with free window resize and min/max bounds |
| ultrahand:review | Reviewing an existing UI against cited UX research |
| ultrahand:theme | Creating, extracting or extending a project's theme (design tokens) |
| ultrahand:theme-review | Auditing a theme, and the designs and code that use it, for consistency and component reuse |
Skills only trigger on UI design, theme or UX review requests (including corrections to a
design). They load on their own when a request matches; to pick one yourself, call it by
name: /ultrahand:mobile design a checkout screen.
Each design skill builds its screen into the shared canvas in shared/canvas/
(see its README): mobile size presets, tablet split view, and free desktop resize
within min/max bounds, with light/dark, 200% text, state and tap-target toggles.
Design claims cite their sources in each skill's references/sources.md.
Every design round, the first design and each change after it, is reviewed before you're
told it's ready, by a reviewer following the review method in its own headless claude -p
session. It starts lean, on the prompt cache, and records what each review cost. Later
rounds are reviewed by their diff, and the whole design is reviewed again when something
shared changes. The loop and its runner are in shared/review/.
The review skill doesn't review anything in your session. It splits the request into
screens and flows, writes a plan, and launches the ultrahand program, which starts a
separate lean claude process for each one. Only their merged findings come back, so a
review of a large app no longer fills the session that asked for it. The skill then judges
duplicates and ranking, checks the worst findings against the code, and writes the report.
The method the reviewers follow is in skills/review/references/method.md.
A project's theme lives in the project (design/theme.tokens.json, in the W3C design
tokens format, plus design/THEME.md for its decisions, component inventory and