Node Team - Coordinated Agent Workflow
EXECUTION INSTRUCTIONS
Follow the orchestration procedure in references/orchestration.md.
DO NOT read references/builder-context.md, references/reviewer-context.md, or references/examples.md. Those are read by subagents only.
Overview
The Node Team skill implements features you define. You provide the WHAT (feature spec), it handles the HOW (implementation).
┌─────────────────────────────────────────────────────────────────┐
│ PLAN.md (you write this) │
│ - Feature name and description │
│ - Acceptance criteria │
│ - Any notes or constraints │
└─────────────────────────┬───────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ TASK MANAGER (subagent) │
│ - Explores codebase for patterns and conventions │
│ - Breaks down into 2-5 minute implementation tasks │
│ - Writes task specs to .tasks/ files │
│ - Returns only task count to orchestrator │
└─────────────────────────┬───────────────────────────────────────┘
│
▼
┌─────────────────────────────────────┐
│ For each task (in dependency order) │
└─────────────────────┬───────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ NODE BUILDER (subagent) │
│ - Reads: .tasks/task-{id}.yaml + builder-context.md │
│ - Follows TDD: RED -> GREEN -> REFACTOR │
│ - Writes results to .tasks/result-{id}-build.yaml │
│ - Returns only status + 1-line summary to orchestrator │
└─────────────────────────┬───────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ NODE REVIEWER (subagent) │
│ - Reads: .tasks/task-{id}.yaml + result-{id}-build.yaml │
│ - Combined review: spec compliance THEN code quality │
│ - Writes results to .tasks/result-{id}-review.yaml │
│ - Returns only verdict + issue count to orchestrator │
└─────────────────────────┬───────────────────────────────────────┘
│
┌───────────┴───────────┐
▼ ▼
┌──────────┐ ┌─────────────┐
│ APPROVED │ │ CHANGES │
│ Next task│ │ NEEDED │
└──────────┘ └──────┬──────┘
│
▼
┌───────────────┐
│ NODE BUILDER │
│ (reads review │
│ from file) │
└───────────────┘
Context-Saving Design
All subagents communicate via .tasks/ files. The orchestrator never reads
source code or detailed results - it only tracks status and dispatches agents.
This keeps the orchestrator's context lean across many tasks.
Invocation
# Use default PLAN.md in current directory
/node-team
# Specify a different plan file
/node-team plan="docs/features/user-auth.md"
# Implement only a specific task (after initial planning)
/node-team task=3
Arguments
| Argument | Default | Description |
|---|---|---|
plan | PLAN.md | Path to the plan file containing feature specification |
task | (all) | Specific task number to implement (optional) |
Plan File Format (PLAN.md) - BDD/Gherkin
The plan file uses BDD (Behavior-Driven Development) format with Gherkin syntax. This provides executable specifications that map directly to tests.
Feature: [Short descriptive name]
As a [role/persona]
I want [capability]
So that [benefit]
Background:
Given [common precondition for all scenarios]
Scenario: [Specific behavior being tested]
Given [initial context/state]
And [additional context]
When [action taken]
And [additional action]
Then [expected outcome]
And [additional outcome]
Scenario: [Another behavior]
Given [context]
When [action]
Then [outcome]
# Optional: Notes section for implementation hints
# Note: Use existing user model from src/components/users/
# Note: JWT secret should come from config
Gherkin Keywords
| Keyword | Purpose |
|---|---|
Feature: | High-level description of the capability |
As a / I want / So that | User story format (optional but recommended) |
Background: | Steps run before each scenario |
Scenario: | Specific testable behavior |
Given | Precondition/initial state |
When | Action being performed |
Then | Expected outcome |
And / But | Additional steps (continues previous keyword type) |
Agents and Their Context
Each agent is a subagent dispatched via the Task tool. The orchestrator does NOT read these files - subagents read their own context.
| Agent | Role | Context File |
|---|---|---|
| Task Manager | Parses plan, explores codebase, creates task breakdown | (explores codebase directly) |
| Node Builder | Implements tasks following TDD, component architecture | references/builder-context.md |
| Node Reviewer | Combined review: spec compliance + code quality in one pass | references/reviewer-context.md |
See references/orchestration.md for exact dispatch templates and the coordination loop.
Anti-Patterns
- Orchestrator reading source code, result files, or reference files (subagents do this)
- Orchestrator echoing or summarizing full subagent output (wastes context)
- Inlining plan content into dispatch prompts (reference by file path instead)
- Dispatching multiple builders in parallel (causes conflicts)
- Proceeding with CHANGES_NEEDED status
- Ignoring security vulnerabilities
- Marking task complete with failing tests/lint
Integration with Other Skills
- planner: Can be used to create the initial plan that Task Manager refines
- prd/rfc: Use these to write the feature specification before implementing