Bench — Conventions para Claude Code em Laravel + Vue/React
PDX-Apps/bench
Adapte Claude Code às convenções do seu projeto Laravel + Vue/React
Ver no GitHubKit de skills e agents que ensina ao Claude Code as convenções específicas do seu projeto Laravel + Vue/React, gerando código que se encaixa na primeira vez. Detecta stack, captura padrões do projeto e cria plugins personalizados por projeto.
Instalação
npx convez-ai PDX-Apps/bench1 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
- Gerar código backend/frontend que respeita padrões do projeto
- Onboarding de assistentes de IA em projetos existentes com regras específicas
- Automatizar criação de features mantendo coerência arquitetural
Veredito

Ferramenta madura (beta v0.8) que resolve um problema real: IA gerando código genérico que não se encaixa em projetos com convenções estabelecidas. Approach inteligente via detecção de stack + plugins por-projeto + skills versionadas. Ideal para equipes com Laravel/Vue/React que usam Claude Code e querem reduzir retrabalho de review.
README
A tradução para o português deste README ainda está na fila.
Bench
Teach AI coding assistants the conventions of your Laravel + Vue/React project — once — so the code they generate lands in the right place, in the right shape, the first time.
Status: beta (v0.8.x). Works end-to-end; not yet battle-tested across many projects. API soft-stable until v1.0.0 — see CHANGELOG.md.
Built for Claude Code today; designed to extend to other AI CLIs.
The problem
AI coding assistants generate generic-best-practices code. But real projects have a specific framework version, a test runner, an auth strategy, a response shape, a folder layout, a UI library. Generic code fights all of that — so you spend the review fixing placement and idioms.
Bench is a pluggable kit of slash commands + worker agents + versioned framework knowledge that adapts to your project. You describe a feature; Bench routes it to a worker that reads only the relevant conventions and produces code that fits.
Contents: Install · Using Bench · How it works · Addons · CLI reference · Docs · License
Install
Clone Bench once; then build it into each project you want to use it in.
# Once, globally — clone Bench somewhere stable:
git clone git@github.com:PDX-Apps/bench.git ~/tools/bench
# Then, once per project — from the project root:
cd ~/my-app
~/tools/bench/bin/bench build
(Tip: symlink ~/tools/bench/bin/bench onto your PATH so it's just bench build everywhere — see scripts/install-cli.sh.)
bench build detects your Laravel / PHP / Vue / React versions, builds a copy of the plugin for this project into .claude/plugins/bench/ (patterns resolved for your versions, your .bench/ overrides + addons merged), and registers it in .claude/settings.json.
Bench is per-project, not a global plugin. Each project gets its own built copy — there is no single build that's correct for every version + override set. Your project must be its own git root.
Monorepo or non-standard layout? Auto-detect reads
composer.json/package.jsonat the repo root and scansapps/,packages/,services/one or two levels down. If your Laravel or Vue/React app lives elsewhere — or the build printsBuilt BACKEND-ONLY/frontend=nonewhen you do have a frontend — set the stack explicitly so the right skills are generated:bench build --frontend=vue --vue=3 --laravel=13 --php=8.5 # or --frontend=reactThe frontend is fixed at build time.
/bench-initandbench rebuildreplay whatever the build set — they do not re-detect or change it. So if Vue/React files are missing, re-runbench buildwith--frontend=…, notbench-init.
Now open Claude Code in the project and head to Using Bench → start with /bench-init.
Using Bench
You work inside Claude Code with slash commands. Onboard the project first, then generate code, and tailor Bench as you discover the project's quirks.
1. Onboard — /bench-init
The first thing to run in a new project. It walks the concerns — a guided interview over the essentials (auth strategy, test framework, permissions model, response shape, layout), where each concern knows exactly which patterns it owns and updates all of them — then offers to scan for anything else non-standard. Everything it captures lands in a committed .bench/ folder that travels with your repo. Re-run a single concern anytime with /bench-configure <name>.
Bench never writes your CLAUDE.md — your project context lives in .bench/.
2. Generate code
Describe what you want; Bench routes it to the right worker(s):
- One artifact —
/laravel create an endpoint to list a user's orders·/vue-component OrderCard - A whole feature —
/bench implement <feature, PRD, or ticket>spawns the agents in dependency order (migration → model → controller → … → tests). - Frontend —
/frontend <request>routes to your Vue or React stack (fixed at build, no guessing). /helplists everything available in this project.
Want plain "build me X" to route through Bench automatically? Paste
examples/CLAUDE.mdinto your project'sCLAUDE.md— it tells Claude to default to the/laraveland/frontendrouters for any build request, so "I need an endpoint" generates to your conventions without you typing a command. (Works in both thestandardandcompactprofiles.)
3. Tailor to your conventions
When generated code doesn't match how your team does it, teach Bench — no core edits. Both commands write to your committed .bench/ folder; run bench rebuild afterward to re-resolve.
/bench-override— change how an existing artifact is generated. Describe it in plain language — "we don't usetoDto, we return aMapper", "controllers extendApiController", "tests are Pest, co-located" — and Bench writes a targeted override of the affected pattern(s), layered onto core so a Bench upgrade won't clobber it./bench-slice— teach Bench a new artifact type unique to your project. Point it at one of your own domains (/bench-slice app/Reports) and it scaffolds a full Bench-grade slice — a/reportskill, a worker agent, and the pattern(s) describing how your reports are built — so every future report generates the same way.
How overrides and addons actually layer onto core — append, replace, and the rest — is covered in Layering.
4. Backend guardrail (optional)
bench guardrail install adds a just-in-time hook that nudges you to route Laravel artifacts (Actions,
Controllers, FormRequests, …) through Bench's skills instead of hand-writing them — Bench's worker
agents are never nudged. Modes: warn (default), block, off via .bench/guardrails.yaml; bench-init
offers it during setup.
How it works
Three layers. The main conversation stays at the feature level — you say what you want; the implementation details and pattern files stay out of your context window.
| Layer | What it is | Example |
|---|---|---|
| Skills | Slash commands — thin routers. Parse your request, inspect the project, resolve ambiguity, then delegate. | /bench, /laravel, /frontend, /vue-component |
| Agents | Subagents that do the work in isolated context — read only the patterns a task needs, scaffold the artifact, return a summary. | controller, vue-component, migration |
| Patterns | Version-aware markdown describing how to write each artifact. The shared knowledge base. | CONTROLLER-001, STORE-001, MIGRATION-001 |
you → /laravel "add an Orders API"
└─ skill routes → controller · request · resource · route agents
└─ each reads its pattern(s) → writes code that matches your project
Patterns are the only version-aware layer. base/ targets the latest idioms — currently Laravel 13 + PHP 8.5 (and Vue 3 / current React) — while overrides/laravel-12/, overrides/php-8.4/ carry rollbacks for older versions. Skills and agents are single-copy and name no version-specific syntax — so the same agent emits L13 or L12 code depending on what your project resolved at build.
Out of the box: 4 routers (/bench, /laravel, /frontend, /help) + 27 Laravel and 10 Vue / 10 React component skills, backed by version-aware patterns for both stacks.
Bench follows a per-artifact test strategy (TEST-000): logic-bearing artifacts (Actions, Services,
Listeners, Jobs, domain methods) get unit tests; HTTP/glue (Controllers, FormRequests, Policies) get
feature tests; declarative artifacts (Events, Resources, DTOs) are asserted within the feature test
rather than given their own. implement emits the right test for each artifact it generates.
Addons
Addons are opt-in plugins that extend — or replace — what core ships: adding commands, agents, and patterns, or swapping a default (the data layer, the router, the styling system) for an alternative. They layer on the same way as your own overrides (see Layering), and can declare depends_on.addons so installing one pulls in what it needs (e.g. livewire pulls in laravel-blade to inherit its page/layout/route ownership).
bench addon add tailwind # by bundled name
bench addon add /path/to/my-addon # or a path
One addon — bench-manager — is bundled and loaded by default; it provides the /bench-* commands above. Everything else is opt-in. Bench ships 44 addons across:
Setup & workflow · Laravel packages · Laravel UI · Frontend styling · Component libraries · Data & routing · Meta-frameworks · Testing · Docs
Full catalog + how to author your own: docs/addons.md.
CLI reference
bench build [--copy|--symlink] [--addon=NAME ...] # build + register Bench for this project
bench rebuild [--addon=NAME ...] # re-resolve after editing .bench/ or upgrading a dependency
bench addon list | add <name|path> | remove <name> # manage addons
bench profile [show | set <compact|standard>] # compact = router/help skills only (agents always ship)
bench status # detected versions, addons, profile, install path
Version flags — build and rebuild auto-detect versions from composer.json / package.json; pass any of --laravel=13 --php=8.5 --frontend=vue|react|none --vue=3 to override that detection. Run bench help for the complete, always-current flag reference.
Docs
- docs/architecture.md — the layers + build pipeline in depth
- docs/addons.md — addon catalog + how to author one
- docs/layering.md — override/addon modes in full
- CHANGELOG.md
License
MIT — see LICENSE.