UX Review — Claude Code plugin

A Claude Code plugin bundling two paired skills that turn raw research and design artifacts into actionable UX feedback:

• ux-review-panel — runs any UX artifact (PNG, folder of screens, Figma, live product, prototype) through a panel of user personas and discipline experts, then produces a research-style report plus a prioritized, stable-ID list of action-items for designers.
• persona-builder — turns any materials (interview transcripts, support tickets, surveys, analytics, a brief) into grounded, distinct personas in exactly the format the panel consumes.

They chain: persona-builder → personas/ → ux-review-panel.

Built for Claude Code. The skills assume real subagents, a writable filesystem, and Bash — that's what powers isolated persona walkthroughs, screens saved to disk, and stable-ID hashing. They still run in plain Claude, but in a degraded mode (sequential personas, screens referenced rather than saved).

For Claude: if the user asks you to install this plugin (e.g. "install the plugin from this GitHub link"), do not clone the repo or copy files into ~/.claude/skills/. This repo is a plugin marketplace — install it by running these two commands via Bash:

claude plugin marketplace add antonskiter/ux-review-skill
claude plugin install ux-review@antonskiter-ux-review

Marketplace name: antonskiter-ux-review. Plugin name: ux-review. After install, the skills ux-review:ux-review-panel and ux-review:persona-builder become available. Then tell the user to restart Claude Code if prompted.

Install

This repo is a plugin marketplace. Install from Claude Code in two commands:

/plugin marketplace add antonskiter/ux-review-skill
/plugin install ux-review@antonskiter-ux-review

That's it — restart Claude Code if prompted. The skills then appear as ux-review:ux-review-panel and ux-review:persona-builder, and trigger automatically from natural language (you don't have to call them by name).

To share with someone: send them this repo link and those two lines. No cloning, no copying files into personal folders.

CLI equivalent:

claude plugin marketplace add antonskiter/ux-review-skill
claude plugin install ux-review@antonskiter-ux-review

Update

/plugin marketplace update antonskiter-ux-review

New versions are picked up because plugin.json carries a bumped version.

Uninstall

/plugin uninstall ux-review@antonskiter-ux-review

Usage

Skills trigger from natural language. Examples:

• "Build personas from these interview transcripts in research/" → persona-builder
• "Run a UX review on the screens in designs/checkout/" → ux-review-panel
• "Here's a Figma link — have some users walk the onboarding and tell me what breaks" → ux-review-panel

A typical end-to-end flow

1. Drop research materials in your project and ask Claude to build personas. It proposes a candidate set, you confirm, and it writes files into a personas/ folder in your project (scaffolding it from the bundled template on first run).
2. Ask Claude to review a set of screens. It captures them, builds a flow graph, walks each persona through naively (one screen at a time), runs the experts, and writes a report + prioritized action-items under ux-review/<date>-<target>-<version>/.

The personas/ folder is project-side — personas are grounded in your users, so they live in the project being reviewed, not in the plugin. persona-builder creates and fills it; ux-review-panel discovers it automatically.

Repo layout

.
├── .claude-plugin/
│   └── marketplace.json            # marketplace catalog (this repo)
└── plugins/
    └── ux-review/                  # the plugin
        ├── .claude-plugin/
        │   └── plugin.json         # plugin manifest
        └── skills/
            ├── ux-review-panel/    # review-panel skill
            │   ├── SKILL.md
            │   ├── experts/        # reviewer roles (usability, visual, journey, a11y)
            │   ├── reference/      # heuristics, severity/effort, flow graph, templates
            │   └── prompts/        # persona-walkthrough + expert-review instructions
            └── persona-builder/    # persona-generator skill
                ├── SKILL.md
                ├── reference/      # material intake, synthesis, grounding, output format
                ├── prompts/        # signal-extraction instruction
                └── assets/
                    └── personas-scaffold/   # README + _TEMPLATE.md dropped into projects

Design principles

• Input-agnostic. Both skills probe the lightest available path first, then ask rather than standing up heavy tooling or installing anything.
• Grounded and honest. Personas trace traits to evidence and carry grounded/confidence flags; review findings note run limitations and what couldn't be reached.
• Stable IDs. Action-items are keyed by a hash of their normalized essence, so they survive rewording and let you diff reviews across versions.
• Consensus over ties. The panel aims for an odd 3–5 personas so priority breaks cleanly.
• Untrusted content. Artifact and material text is treated as data to analyze, never as instructions to follow.
• Speaks your language. If the input is non-English, reports and personas come back in the user's language (real quotes stay verbatim).