Sapiens-DTJ: Orquestrador de IA com Controle Humano
simpscal/sapiens-dtj
GitHub como fonte de verdade — agentes executam, humanos validam cada passo
Ver no GitHubWorkflow GitHub-nativo que integra Claude Code com controle humano em cada etapa. O repositório atua como orquestrador central, mantendo requisitos, histórias, designs e bugs em issues, enquanto agentes de IA executam o trabalho nos repos de código e sapiens (humanos) fazem validação de requisitos, reviews e releases.
Instalação
npx convez-ai simpscal/sapiens-dtj1 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
- Estruturar workflows de IA consistentes em equipes
- Centralizar artefatos (requirements, designs, código) em um único lugar
- Automatizar ciclos de desenvolvimento com revisão humana em gates críticos
- Integrar Storybook como fonte de designs em vez de Figma isolado
- Orquestrar trabalho entre repos backend, frontend e infraestrutura
Veredito

Abordagem robusta para quem precisa escalar IA em equipe sem perder rastreabilidade. GitHub como SSOT é escolha forte, mas exige disciplina e CLI experience. Ideal para times que já usam GitHub Projects — reduz fragmentação de ferramentas.
README
A tradução para o português deste README ainda está na fila.
Sapiens-DTJ
Sapiens Do The Job.
A GitHub-native, human-gated AI workflow for Claude Code — AI agents execute the work; sapiens gate every step: requirements, stories, designs, code review, release.
🤔 Why This Exists
AI can 10x a team — but only if everyone uses it the same way. Without that, you get:
- Inconsistent output. Same feature, three architectures. Each team member prompts differently, gets different results. Productivity gains cancel out in review and rework.
- Context scattered everywhere. Requirements in Notion, designs in Figma, decisions in Slack, state in someone's head. AI can't read any of it. Humans waste time bridging tools.
- No single source of truth. When artifacts live across five platforms, nobody knows what's current. Work gets duplicated, decisions get reversed, context gets lost between sessions.
🧭 How It Works
- Orchestrator, not a codebase. Stories, sprints, designs, TDDs, and bugs live here as issues. Agents check out your real code repos (API, web, infra), registered once via
/setup. - GitHub is the source of truth. No Jira, no Notion — every artifact is an issue; lifecycle state lives on a GitHub Project board (Status field), not in chat.
- Design lives in code, not Figma. Mockups are Storybook surface stories committed to the frontend repo — built from your real components, reviewed as PRs, and read by implementation agents as the pixel-level reference. No design-tool drift: the mockup and the app share one component library.
flowchart LR
subgraph ORCH["This Repo · Orchestrator"]
direction TB
ISSUES["GitHub Issues · Labels · Project Board<br/>requirements · stories · TDDs<br/>designs · bugs · backlog"]
FLOWS["Workflow Commands<br/>/brainstorm · /backlog · /feature · /bugfix · /refactor"]
end
subgraph TARGETS["Your Code Repos"]
direction TB
BE["Backend repo<br/>API"]
FE["Frontend repo<br/>Web"]
IN["Infrastructure repo<br/>Terraform · SAM · etc."]
end
ORCH ==>|"branch · implement · commit · open PR · merge"| TARGETS
TARGETS -.->|"read codebase context"| ORCH
🚀 Quick Start
Prerequisites: Claude Code CLI · GitHub CLI (
gh) · the code repo(s) you want agents to work on.
1. Get the workflow
Option A — dedicated orchestrator repo (recommended). Click Use this template → create your orchestrator repo (e.g. myproduct-hub) — its issues become your requirements, stories, and bugs. Clone it next to your code repos:
ls # myproduct-api/ myproduct-web/
gh repo clone <you>/myproduct-hub
cd myproduct-hub && claude
Option B — embed in an existing repo. Copy .claude/ into your project:
git clone https://github.com/simpscal/sapiens-dtj.git
cp -r sapiens-dtj/.claude /path/to/your/project/
2. Run setup
Inside Claude Code, run /setup and pick All. It generates:
| Artifact | Purpose |
|---|---|
.claude/skills/project-config/SKILL.md | Registers codebases, detects tech stack, configures migration detection |
PRODUCT.md | Captures vision, value proposition, business model, goals, strategic direction |
| GitHub labels | requirement, user-story, bug, etc. — safe to re-run |
| GitHub Project board | Status / Type / Sprint fields — the workflow's tracking surface; safe to re-run |
[!TIP] Give every registered codebase its own
CLAUDE.md. The implementation agents read it to learn each repo's build/test commands, folder structure, naming, and patterns — the richer it is, the more the generated code matches your conventions. Skip it and agents fall back to inference, which is slower and less faithful.
3. Kick off your first sprint
Each workflow has one natural-language entry command — just say what you want:
/feature "Next feature on the backlog"
/feature, /bugfix, and /refactor read your plain-English intent and run the right phase. To drive an entire workflow end-to-end (requirement → stories → design → TDD → implement) — auto-approving routine gates, clarifying the need with you up front, then pausing before pre-release — use /auto:
/auto "Next feature on the backlog"
[!TIP] Not sure which workflow? Just describe the work to
/feature,/bugfix, or/refactor— each routes to the right step. Or/autoto run it end-to-end.
[!NOTE] Non-disruptive. Only
.claude/andPRODUCT.mdare added. Existing issues, PRs, and branches stay untouched.
🔄 Workflows
Five workflows, one shape: AI does the work, you approve at each gate — scope, plan, code review, release.
| Workflow | Entry | Covers |
|---|---|---|
| 🆕 Feature | /feature | Sprint cycle — requirement → stories → design → TDD → implement → release; plus in-sprint bugs & refactors that ride the sprint branch |
| 🐛 Bugfix | /bugfix | Production bug lifecycle — report → fix → release via main |
| 🧹 Refactor | /refactor | Standalone tech-debt cleanup, no user-visible change — spec → implement → release via main |
| 💡 Brainstorm | /brainstorm | Collaborative ideation — riff features with you against PRODUCT.md, score on strategic fit, drop the keepers into the backlog as drafts |
| 🗂 Backlog | /backlog | Capture ideas, bugs, and refactors as drafts, then promote them into a workflow |
Every phase pauses for human review before the next begins:
flowchart LR
SCOPE[Scope] -->|approve| PLAN[Plan<br/>stories · design · TDD]
PLAN -->|approve| BUILD[Implement]
BUILD -->|review| VERIFY[Verify]
VERIFY -->|approve| SHIP[Release]
Run /auto to flow through every phase automatically — it clarifies the need with you up front, then pauses before pre-release.
⚡ Commands
Say what you want — each workflow has one plain-English entry command that runs the right step:
| Command | What it does |
|---|---|
/feature <text> | Feature work — e.g. start checkout revamp, write the stories, implement 42, fix bug in 42, refactor 42, ship sprint 3 |
/bugfix <text> | Bug work — e.g. login returns 500, fix 26, close 26 |
/refactor <text> | Tech-debt — e.g. tidy the payment module, implement 51 |
/brainstorm <text> | Brainstorm feature ideas with you, score them, drop keepers into the backlog — e.g. ideas to boost conversion |
/backlog <text> | Capture an idea or bug, or promote a draft into a workflow |
/auto <text> | Run a whole workflow end-to-end; clarifies the need up front, then pauses before pre-release |
/setup | One-off project setup — codebases, product doc, labels, board |
📋 Tracking Model
Labels classify what an issue is; the GitHub Project board tracks where it sits in the pipeline.
📦 Artifact Types (labels)
What the issue is. Set once on creation.
| Label | Artifact | |
|---|---|---|
![]() | requirement | PO requirement, via /feature |
![]() | user-story | Story, via /feature |
![]() | bug | Bug, via /bugfix; source from title prefix ([Bug] production, [Dev Bug] development) |
![]() | refactoring | Refactor, via /refactor |
![]() | requirement-updated | Informational marker — requirement changed mid-sprint |
🔁 Board Status (Project field)
Where an item sits in the pipeline. Commands move it as work progresses.
| Status | Meaning | What happens next |
|---|---|---|
Backlog | Captured draft via /backlog — not yet planned | /backlog |
Todo | Issue filed, not started | /feature · /bugfix · /refactor |
In Progress | Dev is currently implementing | — |
Implemented | PRs open / merged to staging, awaiting verification | Human merges branch → sprint, or amend ACs then re-implement |
Done | Released and closed | — |




