studio

6SNT-Studio

A gated, 10-agent software studio for Claude Code (agent-teams).

English · Español

Ten specialized AI agents build software in a disciplined pipeline. Each agent owns exactly one deliverable, and nothing advances without explicit human approval at its gate. The result is AI-built software with the predictability of a real engineering process — not a one-shot prompt.

At a glance: 10 agents · 6 human gates · 11 principles · 8 execution rules · 3-tier models (opus/sonnet/haiku) · 2 entry paths (greenfield + brownfield) · local-first.

The studio is not a project — it is the infrastructure that serves every project. Projects come and go; the studio stays.

— Built and operated by CA6SNT (Valdivia, Chile). Method-first, low-bureaucracy: the procedure is the product.

---

🚀 Install (Claude Code)

In a Claude Code session:

/plugin marketplace add 6SNT-RADIO/6SNT-Studio
/plugin install studio@6SNT-Studio

That installs the 10 agents, the skills, the hooks, the /startsnt and /adopt commands, and the project scaffolder. Private repo? run gh auth login once first.

Prerequisite — enable agent-teams. The studio runs on Claude Code's experimental agent-teams. A plugin can't enable this for you, so add the flag to your ~/.claude/settings.json and restart Claude Code:

{ "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" } }

(Optional: OpenTelemetry env vars for local observability — see docs/PIPELINE.md.)

Then, in a session rooted in your project folder: new idea? run /startsnt (greenfield, guided intake). Existing project? run /adopt (brownfield on-ramp).

---

🧭 The idea in one picture

flowchart TD
  A01([01 · Strategist]):::opus --> G01{{G01 · BRIEF}}:::gate
  G01 --> A02([02 · Researcher]):::sonnet --> G02{{G02 · RESEARCH}}:::gate
  G02 --> A03([03 · UX/UI]):::sonnet --> G03{{G03 · IDENTITY}}:::gate
  G03 --> A04([04 · Architect]):::opus --> G04{{G04 · ARCHITECTURE}}:::gate
  G04 --> A05([05 · Data Modeler]):::opus
  A05 --> A06([06 · Frontend]):::sonnet
  A05 --> A07([07 · Backend]):::sonnet
  A06 --> A08([08 · QA]):::sonnet
  A07 --> A08
  A08 --> G08{{G08 · QA}}:::gate
  G08 --> A09([09 · Security]):::opus --> A10([10 · Tech Writer]):::haiku --> G10{{G10 · RELEASE}}:::gate

  classDef opus fill:#8957e5,stroke:#5a2ea6,color:#fff;
  classDef sonnet fill:#2ea44f,stroke:#1b6e34,color:#fff;
  classDef haiku fill:#d97757,stroke:#a64b32,color:#fff;
  classDef gate fill:#f0b429,stroke:#b35900,color:#1a1300;

Boxes are colored by model tier (opus / sonnet / haiku); the amber hexagons are the human gates (PO approval). 06 ∥ 07 run in parallel.

A human Product Owner (PO) approves every gate. A Lead orchestrates (creates tasks, assigns owners, closes gates) but never writes code or deliverables. The ten agents do the work — one artifact each.

Two ways in

Greenfield — an idea → /startsnt → the pipeline above. Brownfield — an existing codebase (built elsewhere/by hand) → /adopt: an on-ramp that grounds the real code, reconstructs its specs (marking inferences [TO VERIFY]), passes one human gate, then rejoins the pipeline at the project's real maturity. Not a second pipeline — an on-ramp.

flowchart LR
  P([existing project]) --> GR[ground + secret scan]:::g --> SM[smoke: does it run?]:::g --> RC[reconstruct specs·TO VERIFY]:::g --> GA{{G-ADOPT}}:::gate --> RJ[rejoin pipeline at real maturity]:::g
  classDef g fill:#2ea44f,stroke:#1b6e34,color:#fff;
  classDef gate fill:#f0b429,stroke:#b35900,color:#1a1300;

---

🧩 Three pieces

Piece	Role
Lead / Orchestrator	Coordinates: classifies work size, builds the task graph, routes escalations, closes gates. Produces no deliverables.
Agents 01–10	Each executes its stage when activated and delivers a single artifact (see AGENTS).
Product Owner (human)	Approves deliverables at each gate and makes the calls agents can't. No agent replaces the PO.

---

✅ Why it works