Ensemble Team Setup

Set up an AI ensemble programming team for any software project. Creates the full structure for a team of expert agents working in a single-driver mob programming style with consensus-based decisions, TDD, and continuous retrospectives.

Workflow

Phase 1: Project Discovery

Gather essential project information. Ask the user:

Project name and description: What is being built? What problem does it solve?
Tech stack: Language, framework, database, frontend approach, testing tools. If unsure, help them decide based on their goals.
Product vision: Target user? MVP scope? Vague ideas are fine — the Product Manager agent will refine them.
Dev environment: Nix? Docker? Standard package managers? CI provider?
Repository: Existing repo or new? Branching strategy?

Phase 2: Team Composition

Determine the right team.

PREREQUISITE: Read

references/role-catalog.md

before proceeding.

Tiered Team Presets

Start with a preset, then adjust based on project needs. The team formation session (Phase 5) helps determine the right fit. The user may modify any preset.

Preset	Size	Composition
Full	~9	1 Product Manager, 1 UI/UX Designer, 1 Accessibility Specialist, 1 Domain SME, 1 QA Analyst, 4 Software Engineers
Lean	~5-6	1 Product Manager, 1 Domain SME, 1 Dev Practice Lead, 2-3 Software Engineers, 1 flex role (UX, QA, or DevOps based on need)
Solo-plus	~3	1 Domain SME, 1 Dev Practice Lead, 1 Software Engineer

Approximate token costs per discussion round:

Solo-plus (~3 agents): Lightweight. ~5-10K tokens per round.
Lean (~5-6 agents): Moderate. ~15-25K tokens per round.
Full (~9 agents): Heavy. ~30-50K tokens per round. Reserve for projects where the governance overhead pays for itself.

Actual costs depend on model, context length, and discussion complexity. These are rough estimates for setting expectations, not precise accounting.

Selecting a preset: Ask the user about project scope, timeline, and complexity. Solo-plus suits focused tasks or spikes. Lean suits most projects. Full suits large-scope products with UI, accessibility, and quality requirements.

Extend, do not replace: These presets build on the role catalog. See the catalog for conditional roles (Security, Data/ML, API Specialist, etc.) that can augment any preset. Odd numbers preferred for tie-breaking.

Research each expert — do NOT pick from a memorized list. For each role:

Identify the specific technology/domain this project needs
Use WebSearch to find the recognized authority — the person who wrote the book, created the tool, or gave the defining talks for that specific area
Verify their credentials, recent work, and relevance to this project
Evaluate: published authority, distinctive voice, practical experience, complementary perspective to other team members

Present each proposed expert with: name, credentials, key published work, why they fit THIS project, and what they'd focus on. Let user approve, swap, or remove.

Phase 3: Generate Team Profiles

Create

.team/<name>.md

for each member.

PREREQUISITE: Read

references/profile-template.md

before proceeding.

Required sections: Opening bio, Role, Core Philosophy (5-8 principles from their published work), Technical Expertise (6-12 items), On This Project (concrete guidance), Communication Style (personality + 4-6 characteristic phrases), Mob Approach, Code Review Checklist (6-12 checks), Lessons (empty, to be updated).

Quality gates: Profile must not be interchangeable with another expert. Must include project-specific guidance. Must capture their distinctive voice.

AI-Approximation Disclaimer

Every profile MUST include the following disclaimer block immediately after the opening biography paragraph:

> **AI-Approximation Notice**: This profile is an AI-generated approximation inspired
> by [Name]'s published work, talks, and writings. The real [Name] has not endorsed
> or reviewed this profile. All outputs should be verified against their actual
> published work. This profile creates a "diversity of heuristics" drawing on their
> known perspectives — it does not simulate the actual person.

AI Self-Awareness Clause

Each profile must include in the "Your Role on This Team" section a statement that the team member is aware it is an AI agent embodying a perspective, not the actual person. Human time constraints are irrelevant to AI agents. Standing aside on a decision when the topic falls outside the role's expertise is appropriate deference, not disapproval.

Compressed Active-Context Form

Each profile MUST include a

## Compressed Context

section at the end: a dense summary of the profile in under 500 tokens covering role, top 3-5 principles, key expertise areas, and characteristic review focus. This compressed form is loaded during discussion and review phases. The full profile is loaded only when the member is actively driving or navigating code.

Phase 4: Generate Project Scaffolding

Coordinator Instructions

PREREQUISITE: Read

references/coordinator-template.md

before proceeding.

Fill in roster, build tools, team size. Generate the coordinator instructions at

.team/coordinator-instructions.md

. Then add a pointer in the harness-specific config file (e.g.,

CLAUDE.md

for Claude Code,

.cursorrules

for Cursor, project instructions for other harnesses) directing the main agent to read

.team/coordinator-instructions.md

. The coordinator instructions file is for the coordinator only.

AGENTS.md — Team Structure Section

Insert a "Team Structure" managed section into

AGENTS.md

noting:

Team member profiles are located in
```
.team/
```
Project owner constraints are defined in
```
PROJECT.md
```
Domain glossary is maintained at
```
docs/glossary.md
```

This section provides orientation for all agents. The conventions section of

AGENTS.md

will be populated by the team during the formation session (Phase 5).

PROJECT.md

PREREQUISITE: Read

references/project-template.md

before proceeding.

Fill in tech stack, scope (Must/Should/Could/Out), dev mandates, environment.

Supporting docs

docs/glossary.md: Domain glossary skeleton (Core Types table, Actions table, Errors table, Type Design Principles)
docs/deferred-items.md: Tracker table (Item | Category | Source | Severity | Status)
docs/future-ideas.md: Parking lot for out-of-scope ideas

Phase 5: Team Formation Session

This is the critical phase. The team debates and reaches consensus on their working conventions, architectural decisions, and domain terminology.

How it works: The coordinator activates the full team, then presents each discussion topic one at a time. The team debates, proposes approaches, and reaches consensus using the Robert's Rules protocol. Outputs go to the appropriate location based on the type of decision:

Working conventions (collaboration norms, definition of done, code conventions, communication norms, mob model details, retrospective cadence, tooling conventions) are recorded in the conventions section of
```
AGENTS.md
```
.
Architectural decisions (principles, deployment strategy, state management, testing philosophy) are recorded as ADRs in
```
docs/ARCHITECTURE.md
```
.
Domain terminology (types, actions, errors, naming conventions) updates
```
docs/glossary.md
```
.

The 10 topics (non-exhaustive — team may add more):

How do we decide what to build?
How does the Driver-Reviewer mob model work?
When is a piece of work "done"?
What is our commit and integration pipeline?
How do we resolve disagreements?
What are our code conventions?
When and how do we hold retrospectives?
What are our architectural principles?
How do we communicate as a team?
What tooling and repository conventions do we follow?

Each topic includes the problem it addresses and sub-questions to guide discussion. The team's answers become their conventions and decisions — not pre-canned templates.

Phase 6: Configure Permissions

Grant team agents the permissions they need to do their work (file editing, shell access, etc.). How this is configured depends on the harness:

Claude Code: Create/update

.claude/settings.json

with

"allow": ["Edit", "Write", "Bash(*)"]

Cursor/Windsurf: Configure tool permissions in the IDE settings
Other harnesses: Follow the harness documentation for agent permission grants

Phase 7: Configure CI

Add

paths-ignore

rules to CI config for any harness-generated session or transcript directories (e.g.,

.claude-sessions/

for Claude Code) to prevent them from triggering CI runs.

Phase 8: Summary

Present: files created, how to start the team. The harness config file (e.g.,

CLAUDE.md

for Claude Code) references

AGENTS.md

via

(or equivalent include mechanism), and the coordinator reads

.team/coordinator-instructions.md

for its operating instructions. Suggest telling the coordinator what to build.

Files created:

```
.team/<name>.md
```
— team member profiles
```
.team/coordinator-instructions.md
```
— coordinator operating instructions
```
AGENTS.md
```
— team structure and conventions (populated during formation session)
```
PROJECT.md
```
— project owner constraints
```
docs/ARCHITECTURE.md
```
— architectural decision records (populated during formation session)
```
docs/glossary.md
```
— domain glossary
```
docs/deferred-items.md
```
— deferred items tracker
```
docs/future-ideas.md
```
— parking lot for out-of-scope ideas
Harness config pointer (e.g.,
```
CLAUDE.md
```
) — directs coordinator to read
```
.team/coordinator-instructions.md
```

Retrospective Protocol

Retrospectives are event-driven, not time-based. AI agents work continuously — arbitrary intervals are meaningless.

Trigger: After each shipped PR (merged to the integration branch).
Participants: The full team that did the work, while they still have context. The human is a team member who participates in the consent protocol, not an external approver.
Format: The team reflects on what worked, what did not, and proposes process improvements. Any structured format (Start/Stop/Continue or equivalent) is fine as long as it produces concrete proposals.
Output: Retrospective proposals are changes to
```
AGENTS.md
```
conventions or new ADRs in
```
docs/ARCHITECTURE.md
```
— the same places formation session outputs go. Process improvement proposals require the same consent protocol as any other team decision, and the human must consent (as a team member) before changes to
```
AGENTS.md
```
or new ADRs are adopted.
Mini-retros after each CI build remain a lightweight observational checkpoint (did we follow the pipeline? was the commit atomic?) and do not require human consent. They are observational, not prescriptive.

Key Principles

Non-negotiable aspects baked in from production experience.

PREREQUISITE: Read

references/lessons-learned.md

before proceeding.

Consensus before push (review locally, then push)
Refactor step is mandatory every commit
CI wait rule (never queue multiple CI runs)
Mini-retro after every CI build (team runs it, not coordinator)
PR-triggered retrospective with consent-based outputs
Driver handoff protocol (summary + git log + green baseline)
Glossary compliance (domain types match glossary)
Deferred items tracked immediately
Reviewer coordination (check others' reviews first)
Explicit Driver onboarding in activation prompts
Session transcripts excluded from CI triggers
AI-approximation disclaimer on every profile
Compressed active-context form on every profile
Stand-aside means deference, not disapproval

Factory Mode (Optional)

When the

pipeline

skill is installed alongside

ensemble-team

, the coordinator detects this during setup and enables factory mode. This adds a Phase 1.5 ("Factory Mode Configuration") between team formation and the build phase.

In factory mode, the coordinator delegates the entire build phase to the pipeline controller rather than managing it directly. The pipeline handles slice queuing, TDD pair dispatch, code review orchestration, and merge decisions through quality gates. The full team still reviews before any push.

See

references/factory-mode.md

for coordinator handoff details.

Factory Mode vs. Supervised Mode

Supervised mode (default): The coordinator manages the build phase. All decisions use the Robert's Rules consensus protocol. The coordinator facilitates driver rotation, review cycles, and retrospectives.
Factory mode (when
```
pipeline
```
is installed): The pipeline manages the build phase. Quality gates replace consensus for build-time decisions. The coordinator remains active for planning, factory configuration, and post-build review. The full team participates in pre-push review and retrospectives as before.

ensemble-team

NPX Install

Tags

SKILL.md Content

Ensemble Team Setup

Workflow

Phase 1: Project Discovery

Phase 2: Team Composition

Tiered Team Presets

Phase 3: Generate Team Profiles

AI-Approximation Disclaimer

AI Self-Awareness Clause

Compressed Active-Context Form

Phase 4: Generate Project Scaffolding

Coordinator Instructions

AGENTS.md — Team Structure Section

PROJECT.md

Supporting docs

Phase 5: Team Formation Session

Phase 6: Configure Permissions

Phase 7: Configure CI

Phase 8: Summary

Retrospective Protocol

Key Principles

Factory Mode (Optional)

Factory Mode vs. Supervised Mode