Web Experience Auditor
Vesias/web-experience-auditor
Auditoria web em 7 fases com subagentes especializados e roadmap ICE-priorizado
Ver no GitHubSistema multi-agente auto-otimizável para auditar websites e aplicações web em 7 fases especializadas. Detecta automaticamente o contexto do projeto, roteia tarefas para subagentes especializados, benchmarca contra práticas 2026 e sintetiza um roadmap de melhorias priorizado por ICE (Impact, Confidence, Effort).
Instalação
npx convez-ai Vesias/web-experience-auditor1 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
- Auditar websites de marketing/landing pages contra best practices atuais
- Analisar dashboards e aplicações admin por densidade de dados e padrões de UX/UI
- Gerar roadmaps de melhoria em layout, KPIs, Core Web Vitals e estrutura de dados
- Validar conformidade com padrões WCAG 2.2 e grid/layout responsivo
- Benchmarking contínuo contra padrões de produto (Linear, Stripe, Notion)
Veredito

Skill robusto e ambicioso para auditorias web profundas. O pipeline de 7 fases, roteamento dinâmico por memória e Knowledge Gate (que valida APIs contra docs oficiais) trazem rigor técnico real. Ideal para construtores que querem automação inteligente de QA; requer familiaridade com shells e arquitetura multi-agente.
README
A tradução para o português deste README ainda está na fila.
web-experience-auditor
A self-optimizing, multi-subagent audit system for live websites and web apps — packaged as a Claude Code / Claude Agent SDK skill. It splits an audit into independent phases, dispatches a specialized subagent per phase (each loading the right domain skills), benchmarks the target against current best practices, and synthesizes a prioritized, evidence-backed enhancement roadmap.
Conductor, not soloist: the orchestrator stays thin (scope → dispatch → synthesize); each subagent goes deep on one lens (layout, UX/UI, KPIs, data, research) in its own clean context.
Features
- 7-phase pipeline — Scope → Browser Capture → Grid/Layout → Section UX/UI → KPI/Metrics → Data/Structured-Data → 2026 Research → Synthesis (ICE-prioritized roadmap).
- Two modes — website (marketing/landing/content) and application (dashboards/admin tools), the latter adding Data-Density, KPI-Intelligence, and Product-Pattern (Linear/Stripe/Notion/…) lenses.
- Meta-orchestrator loop —
DETECT → KNOWLEDGE-GATE → ROUTE → EXPLORE → ANALYZE → BENCHMARK → PROPOSE → VALIDATE → LEARN, repeatable across sessions. - Memory-driven dynamic routing — skills are chosen per detected project phase, not hard-coded; prior preferences/learnings are read from the user's file-memory and reused.
- Knowledge Gate — never trusts training knowledge for version-specific APIs: checks installed versions, then verifies against Context7 → official docs before any framework/SDK decision.
- Continuous learning — accepted patterns are persisted and re-applied on the next audit.
- Hybrid prompting — English prompt scaffolding, findings in
{{OUTPUT_LANGUAGE}}(default Deutsch).
Repository layout
SKILL.md # the orchestrator (entry point, loaded by Claude)
references/
phase-prompts.md # copy-paste subagent prompt per phase
skill-phase-map.md # skill → phase map, agent types, parallelization, fallbacks
project-context-routing.md # DETECT/ROUTE — memory-driven dynamic routing
knowledge-acquisition.md # KNOWLEDGE-GATE — version check + Context7/official-docs research
multi-agent-execution.md # parallel agent topology + optional Workflow script
continuous-learning.md # persist & reuse learned patterns via memory
pattern-library-2026.md # app-mode product-pattern benchmark
data-density-kpi-intelligence.md # data-density analyzer + KPI gap/redundancy intelligence
kpi-taxonomy.md # KPI frameworks, event taxonomy, Core Web Vitals targets
ux-ui-grid-checklist.md # grid/layout + Nielsen/Krug + WCAG 2.2 checklist
data-enhancements.md # schema.org, GEO/llms.txt, instrumentation, data-viz
research-2026.md # 2026 trend checklist + WebSearch query bank + sources
scripts/
detect_project.sh [path] # project/stack/mode/phase/routes detection → YAML
check_versions.sh [path] # declared vs resolved versions of key libs (Knowledge Gate step 3)
scaffold_audit_repo.sh <dir> # generate a neutral open-source audit-skill repo skeleton
Install
Clone (or symlink) into your Claude skills directory:
git clone <this-repo-url> ~/.claude/skills/web-experience-auditor
# or, if developing in place, it already lives there
The skill auto-loads; it triggers on phrases like "analysiere die Seite", "audit this app", "UX/UI review", "KPI analysis", "find data enhancements", "2026 web research".
Usage
Analysiere https://example.com mit dem web-experience-auditor, depth=deep, Ziel=Lead, Output Deutsch.
Or drive the scripts directly:
scripts/detect_project.sh ./my-app # context profile
scripts/check_versions.sh ./my-app # installed framework versions
For full parallel execution, see references/multi-agent-execution.md (includes a ready-to-run
Workflow script). Heavy multi-agent runs are opt-in.
Requirements
- Claude Code or the Claude Agent SDK (skills + subagents + memory).
- Optional but recommended: Context7 MCP (for the Knowledge Gate), a Playwright/Chromium-capable environment (for live Browser Capture). Falls back to static fetch when unavailable.
License
MIT — see LICENSE.