convez.ai
AgenteAutomaçãoIntermediário
tenchan000517

Sistema Organizacional para Agentes de IA

tenchan000517/claude-org-system

Estrutura líder-secretária-workers para agentes de IA escaláveis e organizados

Ver no GitHub
InicialPluginScore · de 0 a 100

Um sistema operacional organizacional minimalista e reutilizável que estrutura agentes de codificação IA (como Claude Code) em papéis claros: líder, secretária e workers. Resolve o problema de contexto espalhado e memória bagunçada ao segregar projetos em diretórios próprios, mantendo orquestração centralizada e descoberta fácil.

Instalação

$npx convez-ai tenchan000517/claude-org-system

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
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Organizar múltiplos projetos de codificação com agentes IA sem misturar contexto
  • Implementar governança e papéis em equipes de desenvolvimento assistidas por IA
  • Criar sistemas multi-domínio com departamentos especializados e orquestração central

Veredito

Uma proposta metodológica sólida para estruturar fluxos de trabalho com agentes IA, com documentação completa e templates prontos. Ideal para equipes que enfrentam caos organizacional em projetos múltiplos, mas exige comprometimento com a estrutura para gerar valor real.

README

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

claude-org-system

English: this page / 日本語: ja/README.md

A minimal, reusable organizational operating system for AI coding agents (Claude Code and similar). It turns a single overloaded workspace into a clean leader → secretary → workers structure, so context stays scoped, "where is what?" is always answerable, and the system never silently breaks.

This is an original methodology. It is published for reference. See LICENSE (CC BY-NC-ND 4.0): you may read and learn from it with attribution; commercial use and redistribution of modified versions are not permitted.


The problem it solves

When one workspace (one folder) does everything — orchestration and the actual work for many unrelated projects — three things rot:

  1. Memory becomes a pile. Every project's notes accumulate in one place. You can't find anything.
  2. No router. There's no map of "what exists / where it lives / what state it's in."
  3. Drift & confusion. A bloated, mixed context makes the agent slower and more likely to act on the wrong thing.

The shape of the fix

Three roles, mapped onto how AI coding agents already work (context scoped per directory):

RoleWhat it isJob
LeaderThe orchestration folderHolds the map, dispatches to workers, does not hoard project work
SecretaryOne inquiry desk (a KB)Answers "where is X / what's the setting / what's the status"
WorkersEach project's own folderDoes the real work; keeps only its own project's memory

The enabling mechanism: most agent runtimes scope their instructions and memory by working directory. So simply doing each project's work in its own directory splits the pile automatically.

What's here

  • CREDO.md — the foundational document; layers, roles, conduct principles, cadence (everything derives from this)
  • PHILOSOPHY.md — the principles the whole system rests on
  • GETTING-STARTED.md — the minimal adoption path (start here)
  • framework/ — the full model:
    • 01-roles · 02-orchestration · 03-memory-model · 04-safety · 05-session-lifecycle (the core)
    • 06-judgment · 07-departments · 08-governance (the org layers, for multi-domain setups)
    • 09-intake · 10-monitoring · 11-operating-rhythm (the operating rituals that make the loop run)
    • 12-operations (the runtime layer — running it on a real machine: tmux/sessions, background processes, logging, notifications)
  • templates/ — fill-in skeletons for every artifact: leader / worker / department CLAUDE.md, INDEX, HANDOFF, secretary KB (single-file + structured + RAG + ingest), inbox (intake & triage), monitor (watchlist), handoff-finalize (verified session close-out), governance (resolution log, strategy review), reports
  • examples/ — worked, fictional examples: demo-org/ (solo) and multi-department/ (org scale)

Core vs complete

The core (roles 01–05 + the basic templates) is all a solo operator or single team needs, and works on its own. The org layers (06–08: judgment framework, departments, governance) and the heavier templates (department credo, resolution log, monthly report) are for multi-domain organizations. Take only what earns its place — the core never depends on the org layers.

How to adopt (one sentence)

Stand up the map first (an INDEX + a leader role note), then add pieces only when a real pain appears — never bolt on the whole apparatus at once. Details in GETTING-STARTED.md.

Scale honestly

A solo operator needs the leader + a light secretary + per-project folders. A large org may want the heavier layers (monthly reporting, ingest pipeline, formal KB). Take only what earns its place. The heavy templates here are optional — the core works without them.

Ver o original no GitHub

Bom para

Relacionados