convez.ai
SkillDesenvolvimento WebAvançado
anilcancakir

claude-code: Parceiro de Desenvolvimento com Planejamento Prévio

anilcancakir/claude-code

Planejamento e execução de features com revisão adversarial e roteamento de modelos.

Ver no GitHub
BomPluginScore · de 0 a 100

Skill para Claude Code que transforma requisitos de features em planos revisados e executados com roteamento inteligente entre modelos (haiku/sonnet/opus). Realiza entrevistas de intenção, audita reuso e qualidade, e executa com verificação adversarial e testes reais antes de aprovação.

Instalação

$npx convez-ai anilcancakir/claude-code

1 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

Estrelas
4
Forks
0
Acompanhando
0
Saúde
71%
Último commit
há 2 dias
Versão
v0.45.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Mudanças que propagam entre múltiplos arquivos e callers distantes
  • Refatorações e migrações que precisam de plano e auditoria de impacto
  • Investigações de bugs com reprodutor e fix verificado antes da execução

Veredito

Excelente para fluxos de desenvolvimento estruturados em projetos complexos, com ênfase em planejamento antes da execução. Overhead significativo para tarefas triviais; melhor aproveitado em mudanças cross-layer e refatorações que exigem disciplina e verificação rigorosa.

README

A tradução para o português deste README ainda está na fila.

ac

Plan-first development partner for Claude Code: interview-driven plans, tier-routed agents (haiku/sonnet/opus), and an adversarial review chain that verifies through real usage, not a green typecheck.

License CI Version Claude Code

Overview

ac is a Claude Code plugin marketplace whose headline plugin turns a feature request into a reviewed, tier-routed execution plan. It interviews you for intent, drafts a wave-and-step plan, and routes each step to the cheapest model that can do it well: haiku for mechanical edits, sonnet for pattern work, opus for cross-layer changes. Every plan passes an adversarial reviewer before execution, and every step is verified four ways, including hands-on runtime QA rather than a passing typecheck alone.

Why / Why not

Use it for:

  • Cross-file features where the change in one shape propagates to distant callers.
  • Refactors and migrations that need a plan and a caller-impact check before the first edit.
  • Bug investigations that need discipline: a reproducer first, then a verified fix.

Skip it for:

  • Trivial one-line edits or a single rename. The planning overhead is not worth it.
  • Throwaway exploration where you do not want a plan artifact on disk.

Quickstart

Run these as slash commands inside Claude Code:

/plugin marketplace add anilcancakir/claude-code
/plugin install ac@ac
/ac:install

/ac:install runs an interactive post-install setup: it creates the my-coding and my-language user skills from short style interviews, merges a portable delegation section into your global CLAUDE.md, and idempotently merges plugin parity into settings.json. After install, start with /ac:plan for a new piece of work.

Commands

CommandWhat it does
/ac:planInteractive planner: parallel research, an intent interview, a reuse/quality/efficiency audit, then a tier-assigned wave-and-step plan written to .ac/plans/<slug>/plan.md. (Skill.)
/ac:executeRuns an approved plan wave by wave on the main thread, spawning tier-routed workers, with per-step 4-layer verification and a final code review. (Skill.)
/ac:installInteractive post-install setup: my-coding / my-language skills, global CLAUDE.md delegation, and settings.json parity.
/ac:init-projectDeep project initialization: parallel ac:explore agents, a scored subdirectory-rule matrix, and drafted CLAUDE.md / CLAUDE.local.md / .claude/rules/*.md.
/ac:commitAtomic commits with style detection from recent history, multi-file splitting, test pairing, optional preflight, and push when an upstream exists.

Skills

Two skills are user-invocable and back the planning workflow:

  • ac:plan: the planner behind /ac:plan.
  • ac:execute: the executor behind /ac:execute.

Six further skills are internal authoring tools the plan chain calls on its own: ac:skill-creator, ac:command-creator, ac:agent-creator, ac:claude-md-rules-creator, ac:prompt-writer, and ac:git-master.

Agents

Ten subagents back the workflow. Advisory agents answer questions; plan-chain workers execute steps; reviewers gate the plan and the implementation.

AgentModelRole
ac:explorehaikuDeep, parallel-friendly codebase research; returns file:line citations with LSP/AST-grep precision.
ac:librariansonnetExternal documentation and OSS research; returns URL/permalink citations with code-snippet evidence.
ac:oracleopusStrategic advisor for architecture, debugging stalls, and reuse-vs-build trade-offs; advises, never executes.
ac:plan-worker-quickhaikuMechanical single-file step executor: config edits, renames, scaffolds, doc-block additions.
ac:plan-worker-juniorsonnetStandard step executor: 1-3 file changes, business logic, pattern and framework-idiom application.
ac:plan-worker-senioropusSenior step executor: cross-layer changes, migrations, and complex edges with caller-impact checks.
ac:plan-reviewersonnetIndependent second-eye reviewer for standard plans; returns OKAY or REJECT with up to 3 blockers.
ac:plan-reviewer-deepopusAdversarial two-pass reviewer for complex plans; stress-tests across seven dimensions.
ac:plan-code-reviewsonnet4-stage post-implementation reviewer for standard plans; returns APPROVED or BLOCKED.
ac:plan-code-deep-reviewopus6-stage post-implementation reviewer for complex plans, including cross-layer integration and Reuse Map enforcement.

Plugin structure

plugins/ac/
  .claude-plugin/
    plugin.json          Plugin manifest.
  .mcp.json              MCP entrypoint (node cli/ac.js mcp).
  agents/                10 subagents (advisory, workers, reviewers).
  commands/              /ac:install, /ac:init-project, /ac:commit.
  skills/                8 skills (ac:plan, ac:execute + 6 creators).
  cli/                   Bundled MCP runtime (ac.js, built from cli/ac/).
  bin/                   CLI launcher.
  references/            Bundled style/CLAUDE.md templates for /ac:install.
  README.md
ComponentLives inLoaded as
Commandsplugins/ac/commands//ac:<name>
Skillsplugins/ac/skills/<name>/SKILL.mdac:<name>
Agentsplugins/ac/agents/<name>.mdac:<name> (subagents)
CLI / MCPplugins/ac/cli/ac.jsMCP server via .mcp.json

The CLI source lives at cli/ac/ (TypeScript on Bun) and builds to plugins/ac/cli/ac.js (Node-targeted ESM) that Claude Code loads through plugins/ac/.mcp.json.

Requirements

  • A recent version of Claude Code with plugin and marketplace support.
  • Node.js >= 20 for the bundled MCP runtime that Claude Code loads at plugins/ac/cli/ac.js.
  • Bun for CLI development and builds (cd cli/ac && bun run build).

Contributing

Contributions are welcome. See CONTRIBUTING.md for the dev setup, the build/typecheck/test pipeline, and the PR checklist.

License

MIT. See LICENSE.

Ver o original no GitHub

Bom para

Relacionados