Agent Orchestration

Delegation Prompt Crafter

Creates structured, context-rich prompts for specialist agents that maximize their effectiveness and minimize back-and-forth.

When to Use

• After clarification-protocol has resolved ambiguities
• When routing to cto-architect for design work
• When routing to strategic-cto-mentor for validation
• When routing to cv-ml-architect for ML-specific architecture
• For any handoff between agents in a workflow

Why This Matters

Specialist agents work best with:

1. Clear context: Business goals, constraints, current state
2. Specific task: Exactly what deliverable is expected
3. Structured requirements: Must-haves vs nice-to-haves
4. Quality criteria: How to evaluate success

Without this structure, agents may:

• Ask redundant questions (wasting time)
• Solve the wrong problem (misunderstanding context)
• Over-engineer or under-engineer (missing constraints)
• Produce outputs in wrong format (unclear expectations)

Delegation Prompt Structure

Every delegation prompt follows this format:

## CONTEXT

### Business Goal
[What business outcome this serves]

### Current State
[Relevant existing systems, constraints, decisions]

### Key Constraints
- [Constraint 1: e.g., "Budget: < $10K/month infrastructure"]
- [Constraint 2: e.g., "Timeline: MVP in 8 weeks"]
- [Constraint 3: e.g., "Team: 3 senior engineers, Python/React expertise"]

### Background Information
[Any relevant context from clarification or previous agents]

---

## TASK

### Primary Deliverable
[Exactly what output is expected]

### Format Requirements
[Structure, sections, level of detail expected]

### Scope Boundaries
- **In scope**: [What to cover]
- **Out of scope**: [What to explicitly exclude]

---

## REQUIREMENTS

### Must-Haves
- [Critical requirement 1]
- [Critical requirement 2]

### Nice-to-Haves
- [Optional enhancement 1]
- [Optional enhancement 2]

### Quality Criteria
- [Criterion 1: e.g., "Architecture must support 10x growth"]
- [Criterion 2: e.g., "Trade-offs explicitly documented"]

### Integration Points
- [What this output feeds into: e.g., "Will be validated by strategic-cto-mentor"]
- [What depends on this: e.g., "Development team will implement from this"]

---

## ADDITIONAL CONTEXT

[Any other relevant information, links to documentation, previous decisions, etc.]

Agent-Specific Templates

See the prompt-templates folder for pre-built templates:

• architect-delegation.md - For cto-architect design work
• mentor-delegation.md - For strategic-cto-mentor validation
• ml-architect-delegation.md - For cv-ml-architect ML work

Crafting Guidelines

Context Section

Business Goal: Be specific about outcomes, not activities

• Bad: "Build a notification system"
• Good: "Enable real-time alerts so users act on time-sensitive events, reducing missed opportunities by 50%"

Current State: Include what exists and what's working

• Existing architecture and tech stack
• Pain points with current solution
• Previous attempts and why they failed
• Existing integrations that must be preserved

Constraints: Be explicit about non-negotiables

• Budget (infrastructure and development)
• Timeline (deadlines, milestones)
• Team (size, skills, availability)
• Technical (must-use technologies, compliance)
• Political (stakeholder preferences, past decisions)

Task Section

Primary Deliverable: One clear output

• Bad: "Help us with the architecture"
• Good: "Provide a system architecture design document with component diagrams, data flow, and technology recommendations"

Format Requirements: Specify structure

• "7-section architecture document per standard format"
• "Executive summary (2 pages max) + detailed appendix"
• "Focus on Phase 1 MVP, with notes on Phase 2 considerations"

Scope Boundaries: Prevent scope creep

• Explicitly state what's NOT included
• Call out decisions already made
• Identify what other agents will handle

Requirements Section

Must-Haves vs Nice-to-Haves: Force prioritization

• Must-haves are blocking—solution fails without them
• Nice-to-haves are enhancements—can be deferred

Quality Criteria: Measurable success

• "Latency < 200ms at p95"
• "Support 100K concurrent users"
• "Cost < $5K/month at launch scale"

Integration Points: Connect the workflow

• What happens after this agent finishes?
• Who consumes this output?
• What format do downstream consumers need?

Common Mistakes to Avoid

1. The Information Dump

Bad: Copying entire conversation history into delegation Good: Distill to relevant context only

2. The Vague Task

Bad: "Design a good system" Good: "Design a notification system architecture that supports 100K users, uses our existing PostgreSQL database, and costs < $2K/month"

3. The Missing Constraints

Bad: Letting agent assume unlimited budget/time Good: Explicitly stating constraints, even if flexible

4. The Forgotten Handoff

Bad: No mention of what happens next Good: "This design will be validated by strategic-cto-mentor before implementation begins"

Output Examples

Example 1: Architecture Delegation

## CONTEXT

### Business Goal
Enable customers to receive real-time notifications for order status changes, reducing support tickets about "where's my order" by 60%.

### Current State
- Monolithic Node.js backend, PostgreSQL database
- Notifications currently sent via email batch (hourly)
- 50K active users, expecting 200K in 12 months
- Mobile app (React Native) and web app (React)

### Key Constraints
- Budget: < $3K/month additional infrastructure
- Timeline: MVP in 6 weeks, full rollout in 10 weeks
- Team: 2 backend engineers, 1 mobile engineer
- Must integrate with existing authentication system

### Background Information
User research shows 73% of support tickets are order status questions. Push notifications tested well in user interviews.

---

## TASK

### Primary Deliverable
System architecture design for real-time notification system

### Format Requirements
Standard 7-section architecture document:
1. Executive Summary
2. System Architecture (with diagrams)
3. Technology Stack Justification
4. Implementation Roadmap
5. Risk Assessment
6. Code Examples (WebSocket integration)
7. Deployment Strategy

### Scope Boundaries
- **In scope**: Backend notification service, mobile push integration, delivery tracking
- **Out of scope**: Email notifications (keep existing), SMS notifications (Phase 2)

---

## REQUIREMENTS

### Must-Haves
- Real-time delivery (< 5 second latency)
- Support for 200K users with 20% daily active
- Push notifications on iOS and Android
- Fallback to email if push fails

### Nice-to-Haves
- Notification preferences per user
- Read receipts / delivery confirmation
- Rich notifications with images

### Quality Criteria
- p95 latency < 5 seconds from event to notification
- 99.9% delivery success rate
- Infrastructure cost < $3K/month at 200K users

### Integration Points
- Will be validated by strategic-cto-mentor before implementation
- Development team will implement from this architecture
- Must integrate with existing user service for preferences

---

## ADDITIONAL CONTEXT

Previous attempt at WebSockets failed due to connection management complexity. Team prefers managed solutions where possible. AWS is our cloud provider.

Advanced: Context & Memory Management

For complex multi-agent systems, simple prompt handoff isn't enough.

Shared State Strategy

• Short-term Memory: Pass critical variables (User ID, Session State) in the prompt's ## CONTEXT section.
• Long-term Memory: Use a Vector DB or shared memory.md file for persisting decisions across sessions.

Error Recovery (Self-Healing)

• If Agent B fails, the Orchestrator should:
  1. Catch the failure (e.g., JSON parse error).
  2. Critique the output.
  3. Re-prompt Agent B with the error + original instruction.

Validation Checklist

Before sending delegation prompt, verify:

• Business goal is outcome-focused, not activity-focused

• All critical constraints are explicitly stated

• Task is specific with clear deliverable

• Format requirements are defined

• Scope boundaries prevent scope creep

• Must-haves are truly must-haves (not nice-to-haves in disguise)

• Quality criteria are measurable

• Integration points explain the workflow context

• No vague terms or buzzwords remain

• ML Architect Delegation Template

🔄 Workflow

Kaynak: Multi-Agent Patterns (Microsoft)

Aşama 1: Orchestration Design

• Select Pattern: Choose architecture (Hierarchical, Joint-Chat, Dynamic).
• Define Roles: Map required skills to distinct agent personas.
• Boundary Check: Ensure no overlap in agent responsibilities.

Aşama 2: Prompt Engineering (Delegation)

• Context Injection: Prepare global context (Project, Constraints).
• Task Definition: Draft clear "Primary Deliverable" for each agent.
• Guardrails: Define "Out of scope" explicit boundaries.

Aşama 3: Routing Logic

• Router Config: Define intent classification rules (Semantic/Keyword).
• Handoff Protocol: Define how Agent A transfers context to Agent B.
• Fallback: Define behavior when no agent matches intent.

Aşama 4: Validation & Simulation

• Dry Run: Simulate conversation flow manually.
• Loop Detection: Verify agents don't get stuck in "Asking clarification" loops.
• Token Audit: Check context window usage per step.

Kontrol Noktaları