agent-orchestrator-template: Instructions file for Claude Code

CLAUDE.md - Autonomous Orchestration Ecosystem

Add this section to your project's CLAUDE.md file.

---

Agent Orchestration

Must: All implementation tasks must pass through the orchestrator workflow before execution.

Decision Gate Exception

An explicitly invoked analysis-only Skill may run before /task when it does not modify code, files, external systems, or production state. Use /evaluation-first-decision to create a Decision Card containing the verdict, evidence, guardrails, stop conditions, and rollback plan.

• GO or PILOT does not itself authorize execution.
• For HIGH or CRITICAL risk, record the responsible human decision owner.
• When execution is approved, include the Decision Card in the /task input.
• The orchestrator must preserve the approved scope, guardrails, stop conditions, and rollback requirements.
• HOLD or STOP blocks execution until the stated verdict-changing conditions are satisfied.

Core Principle

Do NOT use pre-defined abstract agents. Instead:

1. Create specialized agents from actual task requirements
2. Integrate existing agents when synergy improves outcomes
3. Let the agent pool evolve through continuous improvement

Workflow

Consequential decision (when applicable)
     ↓
/evaluation-first-decision
     ↓
Decision Card + human approval where required
     ↓
Task Received through /task
     ↓
Scan pool/ for existing agents
     ↓
Calculate coverage rate against task requirements
     ↓
┌─────────────────────────────────────────────────────┐
│ Coverage 90%+  → Use existing agent                 │
│ Coverage 60-90% → Create integrated agent           │
│ Coverage <60%  → Create new specialized agent       │
└─────────────────────────────────────────────────────┘
     ↓
Execute task within approved guardrails
     ↓
Update manifests/ with metrics
     ↓
Promote to elite/ if qualified

Directory Structure

.claude/
├── skills/
│   └── evaluation-first-decision/
│       ├── SKILL.md
│       ├── references/
│       └── assets/
├── agents/
│   ├── orchestrator.md        # Orchestrator definition (read first)
│   ├── _template.md           # Template for new agents
│   ├── manifests/             # Skill sheets (metadata + metrics)
│   │   └── {agent-name}.yaml
│   └── pool/                  # Agent pool
│       ├── specialized/       # Task-specific agents (newly created)
│       ├── integrated/        # Merged agents (1st/2nd Gen Integration)
│       └── elite/             # Hyper-Elite agents (proven performers)