Claude Code Frontend
TarasTsavolyk/claude-code-frontend
Setup Claude Code pronto para produção com agentes e workflows frontend
Ver no GitHubConfiguração production-ready do Claude Code para projetos frontend com Vue 3, Vite, Tailwind CSS e Pinia. Inclui 12 agentes especializados, 11 regras (8 path-scoped) e 7 skills de workflow para acelerar desenvolvimento frontend. Funciona como template starter que se adapta ao seu stack.
Instalação
npx convez-ai TarasTsavolyk/claude-code-frontend1 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
- Acelerar desenvolvimento de componentes Vue 3 com padrões consistentes
- Automatizar auditorias de acessibilidade, performance e segurança
- Scaffolding de features, testes E2E e refatorações com workflows
- Estabelecer regras path-scoped de arquitetura, estilo e testes no time
Veredito

Uma configuração bem estruturada e madura para times Vue/frontend que querem Claude Code com governança local. Requer setup inicial cuidadoso (user scope + project scope), mas compensa com automação robusta de tasks repetitivas.
README
A tradução para o português deste README ainda está na fila.
Claude Code Configuration for Frontend Projects
A starter template, not an installable package: you copy
CLAUDE.md+.claude/into your own repo (or your~/.claude/) and adapt it. See Setup.
A production-ready Claude Code setup for Vue 3 + Vite + Tailwind CSS 4 + Pinia + Vitest + Playwright projects, with TypeScript optional (its conventions apply only when the project actually uses TS). It bundles 12 specialized agents, 11 rules (8 path-scoped), 7 workflow skills, and an agent pipeline — focused on what a frontend engineer actually needs.
Swap the stack assumptions (package manager, styling engine, framework specifics) to match your repos —
rules/styling.md is written tech-neutrally (Tailwind 4 default; Sass/SCSS, CSS Modules, scoped <style> documented).
Contents
CLAUDE.md # always-loaded project memory (the template)
.claude/
settings.json # permissions + agent-teams flag
rules/ # conventions, loaded per matching file (8 path-scoped + 3 global)
architecture.md code-style.md styling.md testing.md forms.md
accessibility.md performance.md i18n.md
principles.md git-operations.md workflow.md
agents/ # 12 subagents, least-privilege tools
planner devil frontend-developer ui-reviewer accessibility-auditor
test-engineer performance-auditor refactoring-expert debugger
security-scanner ci-cd-engineer docs-writer
skills/ # 7 invokable workflows
scaffold-component scaffold-feature add-e2e-test
debug-frontend a11y-audit perf-audit release
Two scopes — set this up for ALL your projects
Don't copy everything into every repo. Split it:
| What | Where | Why |
|---|---|---|
| Your personal defaults (code-style, git, your workflow, your agents) | User scope ~/.claude/ | Auto-applies in every project, not committed. |
| Project stack specifics + path-scoped rules | Project scope <repo>/.claude/ + CLAUDE.md | Committed, shared with the team, path-scoping works reliably here. |
Known gotcha: path-scoped rules (
paths:frontmatter) in user scope~/.claude/rules/are currently ignored by the parser. So in~/.claude/keep rules global (nopaths:), and put anything path-scoped in the project's.claude/. Verify what loaded with/memoryinside Claude Code. If a path-scoped rule still won't load, try swappingpaths:→globs:(a parser quirk some versions prefer).
Step-by-step setup
Step 1 — Prerequisites
- Claude Code CLI installed and authenticated (
claude). - Node.js 22+ (LTS — Node 18 and 20 are end-of-life).
- A frontend project (ideally with the stack above; adjust otherwise).
Step 2 — Install your personal (user-scope) defaults — once per machine
These apply to every project you open.
mkdir -p ~/.claude/rules ~/.claude/agents ~/.claude/skills
# Global rules that aren't path-specific (your personal way of working)
cp .claude/rules/principles.md ~/.claude/rules/
cp .claude/rules/git-operations.md ~/.claude/rules/
cp .claude/rules/workflow.md ~/.claude/rules/
# Agents you always want available
cp .claude/agents/*.md ~/.claude/agents/
# Workflow skills you always want
cp -r .claude/skills/* ~/.claude/skills/
Open any project and run /agents and /memory to confirm they're picked up.
Step 3 — Install the project-scope config — per repo
From the root of a frontend repo:
# from this package's directory:
cp -r .claude /path/to/your-project/.claude
cp CLAUDE.md /path/to/your-project/CLAUDE.md
# drop personal, machine-local overrides — cp -r copies them, but they shouldn't be shared
rm -f /path/to/your-project/.claude/settings.local.json
This brings in the path-scoped rules (architecture/code-style/styling/testing/forms/accessibility/performance/i18n)
that only load when Claude touches matching files — keeping context lean. Then add .claude/settings.local.json and
.claude/worktrees/ to your project's .gitignore — this repo ignores them, but cp -r doesn't carry that over.
Step 4 — Customize CLAUDE.md for the repo
- Set
<PROJECT_NAME>and adjust the stack list if it differs. - Set the Language flag (TypeScript or JavaScript). TS conventions and the
typecheckstep apply only to TS projects; JS projects get runtime prop validation + JSDoc instead. Detection is automatic fromtsconfig.json/lang="ts", but stating it keeps it unambiguous. - Confirm the script names in the Commands block match this repo's
package.json. The package manager is auto-detected from the lockfile (npm / pnpm / yarn) — you don't hardcode it; only the script names matter. - Adjust the project-structure block to your real folder layout.
- Trim or extend the Core principles for this project.
Step 5 — Tune .claude/settings.json
- The
permissions.allowlist pre-approves your safe commands;git pushandgit commitare inask; destructive commands (force-push, hard reset,rm -rfand its variants) and.env/.pemreads are denied. Edit to match your tooling. Matching is prefix-based, so treatdenyas defense-in-depth behind theaskgates, not a guarantee against every flag ordering. - It pre-approves npm, pnpm, and yarn (install / run / exec) so it works whatever the repo uses. If you only ever use one, trim the others for a tighter allow-list.
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1enables running the Quality Gate agents in parallel (experimental). Remove it if your Claude Code version doesn't support agent teams.
Step 6 — Commit and verify
git add CLAUDE.md .claude
git commit -m "chore: add Claude Code frontend configuration"
Inside Claude Code:
/memory— shows which CLAUDE.md/rules loaded. Open a.vueand a.test.tsfile to confirm the right rules activate per path./agents— lists the 12 agents./scaffold-component(or describe a component) — confirms skills are discoverable.
Step 7 — Drive it via the pipeline
- New feature: ask for a plan first → it runs
planner(anddevilfor non-trivial UX). Then implement →frontend-developer. Then the Quality Gate (ui-reviewer+accessibility-auditor+test-engineer+performance-auditor+security-scanner) →docs-writer. Details in.claude/rules/workflow.md. - Bug: ask to debug →
debuggerfinds root cause → fix → verify.
Optional community add-ons
These are third-party and stack-agnostic — install only if you want them:
- Superpowers — structured workflows (brainstorming, planning, TDD, debugging, code review):
/plugin marketplace add obra/superpowers-marketplacethen/plugin install superpowers@superpowers-marketplace - Claude HUD — statusline showing model, context usage, running agents:
/plugin marketplace add jarrodwatts/claude-hudthen/plugin install claude-hudthen/claude-hud:setup - skills.sh — community skill registry:
npx skills add <owner/repo> - Mytets — Ukrainian theatrical style à la Подерв'янський (just for fun):
/plugin marketplace add rrader/mytetsthen/plugin install mytets
(Verify each is current before relying on it — the plugin ecosystem moves fast.)
Design notes
- Path-scoped rules — each rule loads only for the files it matches (via
paths:globs), so the model's context stays lean instead of carrying every convention all the time. - Skills are procedures — scaffold/debug/audit workflows (numbered, actionable steps), not knowledge dumps. Where
a skill overlaps a same-named agent (
a11y-audit,perf-audit,debug-frontend), it's deliberate: the skill runs inline in your session for quick/solo use; the agent (e.g.accessibility-auditor) is the isolated, read-only specialist the pipeline delegates to in the Quality Gate. - Least-privilege agents — every subagent declares an explicit
tools:list. Review agents are read-only (ui-reviewergets only Read/Glob/Grep,deviladds SendMessage to deliver its critique); the auditors that need a shell for axe/build/auditget a narrowBashbut still can't edit files. - A pipeline, not a free-for-all — planning → build → quality gate (review + a11y + tests + perf + security, in
parallel) → docs, with a read-only
deviladvocate challenging plans before code is written. - Frontend-native concerns are first-class — accessibility, performance, and styling each get a dedicated rule and (for the first two) a dedicated auditor agent.
- Bilingual triggers — every agent has EN + UA trigger words; add more languages by extending
description. - Release automation — a CHANGELOG-driven workflow ships in
.github/workflows/release.yml;docs/release-automation.mddocuments it plus a Changesets alternative for adopters.
Contributing
PRs welcome — new rules, agents, and skills especially. See CONTRIBUTING.md for conventions (keep it generic, TypeScript optional, package-manager-agnostic, least-privilege agent tools) and the templates for issues and PRs.
License
MIT — use, modify, and redistribute freely.