HTML Review Skill
mengzhou0125/html-review-skill
Renderiza documentos em HTML review lado-a-lado com anotações e feedback
Ver no GitHubSkill para Claude que renderiza documentos, especificações, mudanças de código ou propostas de design em HTML lado-a-lado com anotações e feedback. Possui sistema de design plugável, permitindo usar o Evidence Poet (padrão) ou specs customizadas, com suporte para múltiplas línguas incluindo CJK.
Instalação
npx convez-ai mengzhou0125/html-review-skill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Revisar especificações técnicas com anotações lado-a-lado
- Analisar pull requests e mudanças de código em formato visual
- Criar documentos de design review com feedback estruturado
- Gerar relatórios de análise com layout customizável
Veredito

Ferramenta sólida e bem-pensada para review visual de documentos técnicos. O sistema plugável de specs e suporte a CJK demonstram maturidade; a interface Evidence Poet é refinada mas requer instalação local. Ideal para equipes que fazem reviews estruturados e documentação.
README
A tradução para o português deste README ainda está na fila.
html-review-skill
A Claude Code skill that renders a document, spec, code change, or design proposal into a side-by-side review HTML — content + AI rationale annotations + per-section feedback collector.
Pluggable design system. Bundles the Evidence Poet spec as the default
visual style. Accept any external spec following the simple spec-interface.md contract.
Evidence Poet default character: academic journal × architecture magazine — mono labels for information order, serif headlines for narrative weight, gold lines for moments worth pausing. Sharp corners. CJK-safe font stacks (Microsoft YaHei first). Restrained, rational, never cold.
Install
git clone https://github.com/mengzhou0125/html-review-skill.git
cd html-review-skill
./install.sh # macOS / Linux / Git Bash
.\install.ps1 # Windows PowerShell
Idempotent — re-run to update. Copies the skill to ~/.claude/skills/html-review-skill/.
Trigger it
render this doc as a review HTML
build a review HTML for [spec / case study / PR / blog draft]
review this in HTML · side-by-side with annotations
用 HTML 出 review · review 这篇 [文档名]
/html-review
How it works
- Step 0 · Resolve a spec —
--spec=<path>arg · or auto-detect<project>/.claude/design.md· or fall back to bundledspecs/default-spec.md - Asks one bootstrap question (what + how long) and infers:
- Layout archetype A (right-rail · sticky annotation rail beside content) vs B (stacked · annotations inline under content)
- Tag profile A (editorial · 3 status × 4 layer) vs B (technical · 3 axes × 3 variants = 9 colors)
- Review roles to apply (Layer 0 baseline always on + Layer 1+ role-specific lenses)
- Generates a single
.htmlfile:<html lang="zh-Hans-CN">+ spec-driven fonts (CJK fallback chain stays YaHei-first)- Left ToC + per-section content + annotations + per-section
.user-feedbackblock - Fixed "📋 Copy all feedback" button — serializes every feedback block to clipboard on demand · the page saves no data
- For re-reviews: applies DIFF mode (section-level state tags · changed-section gold tint + changebar · unchanged-section fold + dim · rail-trim)
For independent design-system compliance audit (separate from the skill's structural
validation): use a spec-driven auditor. The
evidence-poet-auditor covers
Evidence Poet; for other specs, bring your own.
Pluggable design system
Three ways the skill resolves which spec to use, in priority order:
| Priority | Source | When |
|---|---|---|
| 1 | --spec=<path> arg or user prompt names a path | explicit |
| 2 | <project>/.claude/design.md | the Evidence Poet Design System installer (EPDS) has been run · or any compatible |
| 3 | bundled specs/default-spec.md | default · Evidence Poet |
To bring your own spec, write a markdown file with a §0 JSON block following the contract
in references/spec-interface.md. The contract is narrow
(~20 tokens) — colors · fonts · spacing · font sizes · line widths · corner convention.
Tool's visual identity stays constant — the --review-* and --audit-* extension
tokens (tag profile colors) are tool-defined and don't depend on the loaded spec. The
spec provides the base palette; the tool builds its UI on top. Fork the skill if you
need different tag colors.
What it does NOT do
- Does not install a design system into your project — for Evidence Poet, use evidence-poet-installer
- Does not edit the source document being reviewed (the rail proposes; the human decides; the edit is a separate step)
- Does not save user feedback (the page is ephemeral — collector serializes to clipboard on demand)
- Does not audit its own output — use a spec-driven auditor (see "Verification" above)
Companion tools
| Tool | Repo | Role |
|---|---|---|
evidence-poet-design-system | evidence-poet-design-system | The Evidence Poet design system: installer + builder + auditor. The visual style bundled here as specs/default-spec.md. |
svg-diagram-skill | svg-diagram-skill | Sibling tool · same pluggable-spec pattern · for standalone SVG diagrams |
Files
SKILL.md Claude's manifest
README.md this file
LICENSE MIT
install.sh / install.ps1 copies the skill to ~/.claude/skills/
references/
spec-interface.md 🆕 the design-system spec contract
tokens.css base + review-extension CSS variables (YaHei-first stacks)
components.css all class rules · 2 layout archetypes · tag styles · feedback block
feedback-collector.js self-contained IIFE · "Copy all feedback" button
example.html double-clickable demo of both archetypes
tag-profiles.md Profile A editorial vs Profile B technical
review-roles.md Layer 0 baseline + Layer 1+ role taxonomy
diff-mode.md incremental re-review mode
specs/
default-spec.md 🆕 bundled default · Evidence Poet spec (mirror of the EPDS canonical)
License
MIT. See LICENSE.