GSD Teams - GSD Skill Creator

Documentation > User Guides > GSD Teams

GSD Teams coordinate multiple AI agents working in parallel on complex tasks. Instead of a single agent handling everything sequentially, teams decompose work into tasks, assign specialists, and execute in waves — cutting wall-clock time dramatically while maintaining quality through structured coordination.

Team Composition Patterns

skill-creator supports 5 team topologies, each optimized for different work patterns:

Pattern	Structure	Best For
Leader-Worker	Single coordinator dispatches to workers	GSD phase execution, code generation
Pipeline	Sequential stages (research → plan → execute → verify)	GSD milestone workflows
Swarm	Homogeneous workers on same task type	Bulk file processing, test generation
Router	Intelligent dispatcher to specialists	Multi-domain projects
Map-Reduce	Parallel processing with aggregation	Codebase analysis, data collection

Creating a Team

Teams are defined in .claude/teams/ as JSON configuration files:

{
  "name": "feature-team",
  "description": "Build authentication module",
  "topology": "leader-worker",
  "members": [
    {
      "name": "lead",
      "role": "orchestrator",
      "model": "opus",
      "tools": ["Read", "Write", "Edit", "Bash", "Glob", "Grep", "Task"]
    },
    {
      "name": "implementer",
      "role": "executor",
      "model": "sonnet",
      "tools": ["Read", "Write", "Edit", "Bash", "Glob", "Grep"]
    },
    {
      "name": "tester",
      "role": "verifier",
      "model": "sonnet",
      "tools": ["Read", "Bash", "Glob", "Grep"]
    }
  ]
}

Team Validation

Before launching, teams undergo comprehensive validation:

# Validate team configuration
npx skill-creator team validate feature-team

# Checks performed:
# - All member agents exist and are loadable
# - No circular task dependencies
# - No redundant tool overlap between members
# - No contradictory skill instructions
# - Role semantics align with assigned skills

The validator catches issues like a “verifier” agent having write tools (role incoherence) or two agents with identical skill sets (redundancy).

GSD Integration

GSD commands use teams automatically during phase execution:

# Execute phase 3 with wave-based parallelization
/gsd:execute-phase 3

# Behind the scenes:
# 1. Discovers all PLAN.md files in phase directory
# 2. Analyzes dependencies between plans
# 3. Groups into parallel waves
# 4. Spawns worker agents per wave
# 5. Aggregates results across all workers
# 6. Commits changes atomically

Two specialized GSD team templates are built in:

GSD Research Team — 4 parallel researchers each exploring a different dimension (technology, architecture, quality, concerns), results synthesized by a coordinator
GSD Debugging Team — Adversarial debugging with multiple hypotheses investigated simultaneously

Real-World Example: The Brainstorm Team

The v1.32 brainstorm module uses an 8-agent team with a specialized communication bus:

# Brainstorm team (from src/brainstorm/chipset.yaml)
agents:
  - name: facilitator    # Session orchestration, pathway recommendation
  - name: ideator         # Raw idea generation (Diverge phase)
  - name: questioner      # Clarifying questions, problem exploration
  - name: analyst         # SCAMPER, Six Thinking Hats, structured analysis
  - name: mapper          # Mind mapping, affinity mapping, lotus blossom
  - name: persona         # Rolestorming, figure storming
  - name: critic          # Evaluation (GATED: Converge phase only)
  - name: scribe          # Capture and documentation (always-on)

communication:
  loops:
    - session   # Facilitator to/from all technique agents
    - capture   # All agents to Scribe
    - user      # CAPCOM to human
    - energy    # Energy signals to Facilitator

The Critic agent is architecturally gated — it literally cannot instantiate during non-Converge phases, enforcing Osborn’s brainstorming rules at the code level rather than by instruction.

The Den Staff: 10-Position Chipset

The v1.28 Den chipset defines a permanent 10-position staff with dedicated token budgets:

# From .chipset/chipset.yaml
name: den-v1.28
totalBudget: 0.59  # 59% of context window

positions:
  - id: coordinator    # 0.08 budget, persistent, session_start
  - id: planner        # 0.06 budget, task lifecycle, on_phase_enter
  - id: executor       # 0.10 budget, task lifecycle, on_execution_start
  - id: verifier       # 0.06 budget, task lifecycle, on_verification
  - id: relay          # Message routing between positions
  - id: monitor        # Health and performance tracking
  - id: dispatcher     # Work distribution
  - id: configurator   # Environment setup
  - id: analyst        # Data analysis and reporting
  - id: observer       # Pattern capture for skill-creator

Co-Activation Tracking

skill-creator automatically detects when skills frequently activate together and suggests composing them into agents:

# After 5+ co-activations over 7+ days:
# skill-creator suggests:
npx skill-creator agent suggest

# Output:
# Detected cluster: [typescript-patterns, test-generator, code-review]
# Co-activation count: 12 over 14 days
# Suggested agent: typescript-quality
# Composition: Leader-worker with test-generator as primary

This bridges the gap between individual skills and coordinated teams — the system learns which tools work together and proposes automation.