Pipeline-driven framework for autonomous game development using AI agents.
Mission: Quality over speed. Ship it right, not just fast.
🌐 Public site: https://brott-studio.github.io/
(This studio-framework repo is internal pipeline infrastructure. The old /studio-framework/ dashboard has been retired.)
Projects powered by this framework. Each project's own README has a live status block (sprint, backlog, PRs, audits).
Autobattler roguelike — current active project
Details: README status · open PRs · backlog · audits
Framework repo: brott-studio/studio-framework · Audits: brott-studio/studio-audits
Last updated: 2026-04-18 05:21 UTC · update workflow
If you're a fresh agent session spawned against this repo, read the framework in this order:
- FRAMEWORK.md — What the studio is. Principles. Leadership. Core rules inline.
- PIPELINE.md — How a sprint runs. Phases. Gates. Communication.
- agents/<your-role>.md — Your specific profile.
- Any policy your profile cross-references — see below.
| File | Purpose |
|---|---|
| FRAMEWORK.md | Studio operating manual — principles, leadership, pipeline, core rules |
| PIPELINE.md | Sprint flow, phases, orchestration, sprint Specc gate |
| ARC_BRIEF.md | Arc brief pattern — what HCD writes to direct an arc |
| SPAWN_PROTOCOL.md | Per-agent spawn templates, preamble, credential handling |
| ORCHESTRATION_PATTERNS.md | Multi-agent patterns (pipeline, fan-out, spike sprint) |
| SUBAGENT_PLAYBOOK.md | Spawn knobs (thinking, timeout), incremental-write protocol |
| ESCALATION.md | 🟢🟡🔴🚨 autonomy tiers, reversibility principle, two-approvals-unlock |
| COMMS.md | Channel-not-DM rule, mention discipline, subagent summary format |
| SECRETS.md | PAT location, credential helper, never-in-prompts rule, rotation |
| REPO_MAP.md | Three-repo architecture (framework / project / audits), who writes where |
| CONVENTIONS.md | Branch names, commit messages, PR titles, task IDs, file naming |
All under agents/. Each profile starts with a Core Rules inline block (autonomy / comms / secrets / framework-first) so load-bearing rules are visible at decision time without link-chasing.
- agents/riv.md — Lead Orchestrator
- agents/ett.md — Technical PM
- agents/gizmo.md — Game Designer
- agents/nutts.md — Developer
- agents/boltz.md — Lead Dev / Reviewer
- agents/optic.md — Verifier
- agents/specc.md — Inspector (independent)
- agents/patch.md — DevOps (on-demand)
- Pipeline-driven — sprints follow a defined flow: Design → Plan → Build → Review → Verify → Deploy → Audit
- Riv orchestrates — Lead Orchestrator spawned by The Bott per sprint; runs the pipeline
- Structural > compliance (where possible) — every rule tagged [Structural] or [Compliance-reliant]
- Verification over trust — Playwright visual testing, headless sims, mocked gameplay
- Learning extraction — Specc reads agent transcripts and writes KB entries
- Reversibility trumps permission — reversible decisions are made, not asked; see ESCALATION.md
- State lives in files, not memory — every agent is replaceable at any moment
This is v2, evolved from v1 (archived).
See v1's KEY_LEARNINGS.md for the 11 foundational lessons. Notable evolutions:
- Rivett → Riv + Ett. The original PM+orchestrator role was initially retired due to orchestration failures (later root-caused to a
maxSpawnDepthconfig issue, not a design flaw). Brought back split into Riv (orchestrator) and Ett (project manager). Both are active canon. - Glytch retired into Optic. One strong Verify stage beats two weak separate stages.
- Dashboard as framework infrastructure, not project deliverable.
Maintained by The Bott (Executive Producer). Designed for AI agents. Learned from 16 sprints of v1.