visual-hierarchy-reviewer — Claude Code skill: UI review
visual-hierarchy-reviewer is a Claude Code skill that reviews task hierarchy and design language from screenshots. It runs on interface reviews, including requests to reduce an AI-made look while keeping the product’s brand.
MIT ★ 529 updated 2026-11-17
npx skills add ShuaiMXu/happyhands --skill visual-hierarchy-reviewer -a claude-codeWhat visual-hierarchy-reviewer does
The skill compares one product state across desktop and mobile screenshots. It notes what works, then ties each finding to a visible element and screenshot, explains the clash with the task or brand, and suggests a small in-scope change. Its checks draw on a shared design rubric and linked methods.
Review requires original screenshots at actual viewport sizes. Gradients, fonts and card layouts alone do not establish AI authorship, and the handler does not assign an AI score. A model must review the screenshots. Findings belong in the current experience audit; planning or editing requires existing scope approval.
What it helps with
- compare interface hierarchy across desktop and mobile
- spot visual choices that compete with the main task
- anchor review findings to screenshot evidence
- preserve the product’s brand during a visual review
Example prompts for the agent
- Review these screenshots on desktop and mobile: can users tell what to do first?
- Find visual choices that clash with the product’s brand and suggest small changes.
- Add screenshot-backed interface findings to the current experience audit.
How to install visual-hierarchy-reviewer
1Skills CLI (Vercel)
npx skills add ShuaiMXu/happyhands --skill visual-hierarchy-reviewer -a claude-code2Manually: put the files in place
~/.claude/skills/visual-hierarchy-reviewer/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/visual-hierarchy-reviewer/
1Skills CLI (Vercel)
npx skills add ShuaiMXu/happyhands --skill visual-hierarchy-reviewer -a codex2Manually: put the files in place
.agents/skills/visual-hierarchy-reviewer/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add ShuaiMXu/happyhands --skill visual-hierarchy-reviewer -a cursor2Manually: put the files in place
.cursor/skills/visual-hierarchy-reviewer/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install ShuaiMXu/happyhands skills/product-design/visual-hierarchy-reviewer2Manually: put the files in place
.github/skills/visual-hierarchy-reviewer/SKILL.md1Skills CLI (Vercel)
npx skills add ShuaiMXu/happyhands --skill visual-hierarchy-reviewer -a gemini-cli2Platform CLI
gemini skills install https://github.com/ShuaiMXu/happyhands --path skills/product-design/visual-hierarchy-reviewerManually: put the files in place
.devin/skills/visual-hierarchy-reviewer/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about visual-hierarchy-reviewer
How do I install visual-hierarchy-reviewer in Claude Code?
Run npx skills add ShuaiMXu/happyhands --skill visual-hierarchy-reviewer -a claude-code. The skill file is at ~/.claude/skills/visual-hierarchy-reviewer/SKILL.md.
What does visual-hierarchy-reviewer do?
It reviews task hierarchy and design language from screenshots, links findings to visible evidence, and suggests scoped changes.
What does the interface review need?
It needs original desktop and mobile screenshots at actual viewport sizes. Claims about interaction or disclosure need interaction evidence.
Source text
SKILL.md · MIT
--- name: visual-hierarchy-reviewer description: Review screenshot evidence for task hierarchy and generic or incoherent design language, including requests to remove visual AI feel while preserving the product's brand. --- # Visual Hierarchy and Design Language Reviewer Use the critical task, intended audience and supplied design foundation to evaluate the rendered interface. Read [rubric.json](rubric.json) for the shared design-language checks and counterexamples. Read [references/sources.md](references/sources.md) when evaluating or updating the adopted methods. Compare the same product state on desktop and mobile. Identify what already works before proposing a change. For every finding cite an inspected screenshot and visible element, explain the task or brand conflict, and suggest the smallest remedy within scope. Record missing evidence instead of inferring hidden behavior or inventing facts. Treat generic composition, decorative competition, inconsistent type rhythm and irrelevant imagery as questions to investigate, not proof of AI authorship. Familiar patterns, gradients, common fonts and card layouts can serve an intentional brand. Do not replace the user's direction with a house aesthetic or manufacture novelty. The bounded handler prepares the versioned rubric, task, foundation context and screenshot references. Its heading/action counts are structural review targets only. It returns `needs-evidence` until screenshot-aware review occurs; it never assigns an AI score. Submit baseline visual decisions through the existing `visual-hierarchy-reviewer` attestation. Findings still require the existing scope confirmation before planning or editing. Independent after-change QA uses the same rubric. ## Bundled OJO methodology Use the [Harness OJO adapter](../../vendor/ojo/ADAPTER.md), then its linked original Skill and relevant references. The runtime records a verified guidance snapshot in this result; design execution and independent QA receive the same pinned source with mode-specific reading routes. Preserve the existing evidence and approval boundaries. ## Complete Taste and Impeccable methodology Use the full [Taste adapter](../../vendor/taste/ADAPTER.md) and [Impeccable adapter](../../vendor/impeccable/ADAPTER.md) with their complete upstream references. The runtime stages both beside OJO with immutable provenance. Audit/Critique inform read-only review; Polish/Layout/Animate are implementation methods behind existing scope approval. Rule-handler output proves only guidance availability until screenshot-aware model review occurs. Do not replace the project's approved Foundation, apply every stylistic variant together, or infer a script pass or human acceptance from installation. ## Spacing and contextual tips Use the `spacing-relationships` and `contextual-information` checks in the shared rubric. Group spacing by task relationship and the existing Foundation, not a universal pixel scale. Review what is visible, contextual or disclosed as well as empty space. Decision-changing limits, blockers and required instructions remain visible at the decision/control; raw evidence and optional explanation remain available in full. Inspect normal desktop and mobile first, preserve natural labels, and request interaction evidence before asserting disclosure or tooltip usability. These are screenshot-aware review questions, not new deterministic findings or automatic approval. ## HappyHands UI profile For HappyHands-owned interfaces or explicitly adopting projects, read the complete `happyhands-ui/SKILL.md` in the staged design-guidance bundle. It contains the eleven user-approved copy, Radix color, hierarchy and interaction rules. The bundle hashes and preserves this profile alongside the complete vendor skills; do not infer adoption by an external target project. ## Execution contract and admission The bounded handler supplies: Heading/action-count candidates and a visual review brief. The professional method still requires: Inspect original screenshots at actual viewport sizes; demonstrate the task or reading-order consequence. Large type, many actions or a style preference alone is insufficient. Deliver into the current Run's existing experience audit or decision; do not create a competing report. Cite the frozen Skill path and exact rule, original state and source, observed failure, task consequence and the reasoning connecting them. Record preservation/counterevidence. With no demonstrated defect, keep the candidate unadmitted and make no change. `complete` describes this handler's bounded output, not professional review or human acceptance.
© ShuaiMXu · MIT · ShuaiMXu/happyhands