AI Team Orchestration
When to Use
- Starting a new project that needs planning, development, testing, and deployment
- Setting up parallel AI agent teams (dev, QA, DevOps)
- Writing brainstorm prompts that produce real debate (not generic output)
- Creating sprint plans with cross-session context survival
- Recovering from context overflow mid-sprint
Team Roles
| Agent |
Name |
Role |
Focus |
| Producer |
Remy |
Sprint planning, coordination, merging PRs |
Scope control, handoffs, issue triage |
| Product Designer |
Kira |
UX, mechanics, user experience |
Fun factor, user flows, feature design |
| Visual/Art Director |
Milo |
CSS, animations, visual identity |
Design system, polish, accessibility |
| Frontend Engineer |
Nova |
UI framework, state management, components |
React/Vue/Svelte, client-side logic |
| Backend Engineer |
Sage |
API, database, auth, security |
Server-side logic, infrastructure |
| DevOps Engineer |
Dash |
CI/CD, cloud deployment, pipelines |
GitHub Actions, Azure/AWS/GCP |
| QA Engineer |
Ivy |
E2E tests, automation, playtesting |
Playwright/Cypress, bug filing, sign-off |
Customize names and roles for your project. Not every project needs all roles.
Session Architecture
The human (CEO) is the message bus between parallel Claude Code sessions:
┌────────────────────────────────────────┐
│ @ai-team-producer — Plans, merges │
│ Session 1 — NEVER writes code │
└────────────────┬───────────────────────┘
│ Human carries messages
┌──────────┼──────────┐
▼ ▼ ▼
┌──────────┐ ┌────────┐ ┌────────┐
│ Session 2│ │Session │ │DevOps │
│@ai-team │ │3 │ │Session │
│-dev │ │@ai-team│ │(on │
│ │ │-qa │ │demand) │
│ Nova │ │ Ivy │ │ │
│ Sage │ │ │ │ │
│ Milo │ │feature/│ │feature/│
│ │ │qa-N │ │devops-N│
│ feature/ │ └────────┘ └────────┘
│ sprint-N │
└──────────┘
Each team works in a separate Claude Code session with its own repo clone:
git clone <repo> project-dev # Dev session
git clone <repo> project-qa # QA session
git clone <repo> project-devops # DevOps session (only when needed)
Open each clone in a separate terminal window or tmux pane and start Claude Code there.
Project Bootstrap
1. Create PROJECT_BRIEF.md
The single source of truth across all sessions. See references/project-brief-template.md.
Required sections (do not abbreviate):
- Project Overview
- Concept / Product Description
- Tech Stack
- Architecture (ASCII diagram)
- Key Files Map
- Team Roles
- Sprint Status (updated every sprint)
- Current State (rewritten every sprint)
- Security Rules
- How to Run Locally
- How to Deploy
- Cross-Session Handoff Protocol — how context survives between sessions
- Bug & Fix Tracking — GitHub Issues as single source of truth
- Multi-Repo Setup — separate clones, branch strategy, merge rules
2. Run a Brainstorm
See references/brainstorm-format.md. Key: name each agent explicitly with distinct personality and perspective. Require at least 2 genuine disagreements to prevent groupthink.
3. Create Sprint Plans
See references/sprint-plan-template.md. Every sprint gets:
docs/sprint-N/plan.md — prioritized tasks, success criteriadocs/sprint-N/progress.md — live tracker, enables recoverydocs/sprint-N/done.md — handoff doc written at sprint end
4. Execute Sprints
Read PROJECT_BRIEF.md, then read docs/sprint-N/plan.md. Execute Sprint N.
First: git pull origin main && git checkout -b feature/sprint-N
Close GitHub Issues in commits: "fix: description (Fixes #NN)"
Update docs/sprint-N/progress.md after each phase.
When done, push and create PR: git push origin feature/sprint-N
Follow Sections 12-14 of PROJECT_BRIEF.md.
5. QA Sign-off
After dev merges, QA does a full playthrough:
Read PROJECT_BRIEF.md. You are Ivy (QA).
Sprint N is merged to main. Do full playthrough.
File bugs as GitHub Issues. Write docs/qa/sprint-N-signoff.md.
Context Recovery
When a session gets long (>100 messages), save state and start fresh:
Before closing:
- Update
docs/sprint-N/progress.md with current status
- Update
PROJECT_BRIEF.md sections 7+8
- Write
docs/sprint-N/done.md
Cold start prompt:
Read PROJECT_BRIEF.md and docs/sprint-N/progress.md.
Continue from where it left off.
Anti-Patterns
See references/anti-patterns.md for the full list. Top 5:
| Don't |
Do Instead |
| Rebase feature branches |
Merge (rebase loses commits) |
| Producer writes code |
Producer only plans, merges, files issues |
| Batch "fix everything" commits |
One commit per fix with issue reference |
| Vague brainstorm prompts |
Name each agent with distinct perspective |
| Keep bugs only in chat |
File GitHub Issues (session context dies) |
Tips for Better Results
- "Take your time, do it right" in prompts produces better output than rushing
- Test before merge — you playtest, file issues, dev fixes, then merge
- Run team consiliums before major sprints — each agent reviews the plan from their perspective
- Save lessons to memory after every milestone
Claude Code Notes
Adapted from: skills/ai-team-orchestration/SKILL.md (awesome-copilot community)
by Denis Evdokimov
Key adaptations:
- "VS Code window" → "Claude Code session" / "terminal window" throughout — the
multi-session architecture is identical; the tool is different
runSubagent references removed — this skill's parallel-team model uses separate
Claude Code sessions, not nested Task calls. The human carries messages between them.- GitHub Issue commands updated to
gh CLI syntax
references/ subdirectory content (project-brief-template.md, brainstorm-format.md,
sprint-plan-template.md, anti-patterns.md) preserved as-is — they're plain markdown
with no Copilot-specific dependencies
1---2name: ai-team-orchestration3description: Bootstrap and run a multi-agent AI development team. Use when: starting a new software project with AI agents, setting up parallel dev/QA teams, creating sprint plans, writing brainstorm prompts with distinct agent voices, recovering a project workflow, or planning sprints. Each agent runs in a separate Claude Code session; the human acts as the message bus between sessions.4---56# AI Team Orchestration78## When to Use9- Starting a new project that needs planning, development, testing, and deployment10- Setting up parallel AI agent teams (dev, QA, DevOps)11- Writing brainstorm prompts that produce real debate (not generic output)12- Creating sprint plans with cross-session context survival13- Recovering from context overflow mid-sprint1415## Team Roles1617| Agent | Name | Role | Focus |18|-------|------|------|-------|19| Producer | **Remy** | Sprint planning, coordination, merging PRs | Scope control, handoffs, issue triage |20| Product Designer | **Kira** | UX, mechanics, user experience | Fun factor, user flows, feature design |21| Visual/Art Director | **Milo** | CSS, animations, visual identity | Design system, polish, accessibility |22| Frontend Engineer | **Nova** | UI framework, state management, components | React/Vue/Svelte, client-side logic |23| Backend Engineer | **Sage** | API, database, auth, security | Server-side logic, infrastructure |24| DevOps Engineer | **Dash** | CI/CD, cloud deployment, pipelines | GitHub Actions, Azure/AWS/GCP |25| QA Engineer | **Ivy** | E2E tests, automation, playtesting | Playwright/Cypress, bug filing, sign-off |2627Customize names and roles for your project. Not every project needs all roles.2829## Session Architecture3031The human (CEO) is the message bus between parallel Claude Code sessions:3233```34┌────────────────────────────────────────┐35│ @ai-team-producer — Plans, merges │36│ Session 1 — NEVER writes code │37└────────────────┬───────────────────────┘38 │ Human carries messages39 ┌──────────┼──────────┐40 ▼ ▼ ▼41┌──────────┐ ┌────────┐ ┌────────┐42│ Session 2│ │Session │ │DevOps │43│@ai-team │ │3 │ │Session │44│-dev │ │@ai-team│ │(on │45│ │ │-qa │ │demand) │46│ Nova │ │ Ivy │ │ │47│ Sage │ │ │ │ │48│ Milo │ │feature/│ │feature/│49│ │ │qa-N │ │devops-N│50│ feature/ │ └────────┘ └────────┘51│ sprint-N │52└──────────┘53```5455Each team works in a **separate Claude Code session** with its own repo clone:56```bash57git clone <repo> project-dev # Dev session58git clone <repo> project-qa # QA session59git clone <repo> project-devops # DevOps session (only when needed)60```6162Open each clone in a separate terminal window or tmux pane and start Claude Code there.6364## Project Bootstrap6566### 1. Create PROJECT_BRIEF.md6768The single source of truth across all sessions. See `references/project-brief-template.md`.6970**Required sections (do not abbreviate):**711. Project Overview722. Concept / Product Description733. Tech Stack744. Architecture (ASCII diagram)755. Key Files Map766. Team Roles777. Sprint Status (updated every sprint)788. Current State (rewritten every sprint)799. Security Rules8010. How to Run Locally8111. How to Deploy8212. **Cross-Session Handoff Protocol** — how context survives between sessions8313. **Bug & Fix Tracking** — GitHub Issues as single source of truth8414. **Multi-Repo Setup** — separate clones, branch strategy, merge rules8586### 2. Run a Brainstorm8788See `references/brainstorm-format.md`. Key: name each agent explicitly with distinct personality and perspective. Require at least 2 genuine disagreements to prevent groupthink.8990### 3. Create Sprint Plans9192See `references/sprint-plan-template.md`. Every sprint gets:93- `docs/sprint-N/plan.md` — prioritized tasks, success criteria94- `docs/sprint-N/progress.md` — live tracker, enables recovery95- `docs/sprint-N/done.md` — handoff doc written at sprint end9697### 4. Execute Sprints9899```100Read PROJECT_BRIEF.md, then read docs/sprint-N/plan.md. Execute Sprint N.101102First: git pull origin main && git checkout -b feature/sprint-N103104Close GitHub Issues in commits: "fix: description (Fixes #NN)"105Update docs/sprint-N/progress.md after each phase.106When done, push and create PR: git push origin feature/sprint-N107Follow Sections 12-14 of PROJECT_BRIEF.md.108```109110### 5. QA Sign-off111112After dev merges, QA does a full playthrough:113```114Read PROJECT_BRIEF.md. You are Ivy (QA).115Sprint N is merged to main. Do full playthrough.116File bugs as GitHub Issues. Write docs/qa/sprint-N-signoff.md.117```118119## Context Recovery120121When a session gets long (>100 messages), save state and start fresh:122123**Before closing:**1241. Update `docs/sprint-N/progress.md` with current status1252. Update `PROJECT_BRIEF.md` sections 7+81263. Write `docs/sprint-N/done.md`127128**Cold start prompt:**129```130Read PROJECT_BRIEF.md and docs/sprint-N/progress.md.131Continue from where it left off.132```133134## Anti-Patterns135136See `references/anti-patterns.md` for the full list. Top 5:137138| Don't | Do Instead |139|-------|------------|140| Rebase feature branches | Merge (rebase loses commits) |141| Producer writes code | Producer only plans, merges, files issues |142| Batch "fix everything" commits | One commit per fix with issue reference |143| Vague brainstorm prompts | Name each agent with distinct perspective |144| Keep bugs only in chat | File GitHub Issues (session context dies) |145146## Tips for Better Results147148- **"Take your time, do it right"** in prompts produces better output than rushing149- **Test before merge** — you playtest, file issues, dev fixes, then merge150- **Run team consiliums** before major sprints — each agent reviews the plan from their perspective151- **Save lessons to memory** after every milestone152153---154155## Claude Code Notes156157**Adapted from:** `skills/ai-team-orchestration/SKILL.md` (awesome-copilot community)158by Denis Evdokimov159160**Key adaptations:**161- "VS Code window" → "Claude Code session" / "terminal window" throughout — the162 multi-session architecture is identical; the tool is different163- `runSubagent` references removed — this skill's parallel-team model uses separate164 Claude Code sessions, not nested Task calls. The human carries messages between them.165- GitHub Issue commands updated to `gh` CLI syntax166- `references/` subdirectory content (project-brief-template.md, brainstorm-format.md,167 sprint-plan-template.md, anti-patterns.md) preserved as-is — they're plain markdown168 with no Copilot-specific dependencies
asleekgeek