Delegation-First Development

Core Philosophy

Main Claude = Architect & Coordinator

• Strategic planning and design decisions
• Requirements clarification with the user
• High-level architecture and system design
• Orchestrating sub-agent workflows
• Synthesizing and presenting results
• Maintaining conversation continuity

Sub-Agents = Specialized Implementers

• Each has a fresh context window (no accumulated noise)
• Domain expertise applied to specific tasks
• Parallel execution for independent work
• Detailed implementation without polluting main context

When This Skill Applies

ALWAYS delegate when the task involves:

Handle directly ONLY when:

• Answering questions about approach/strategy
• Clarifying requirements with the user
• Making architectural decisions
• Discussing trade-offs
• Reviewing/synthesizing sub-agent outputs
• Single-line trivial edits (explicit user request)

Delegation Decision Tree

User Request Received
│
├─ Is it a question about design/architecture/approach?
│  └─ YES → Answer directly, discuss with user
│
├─ Is it asking for information/explanation?
│  └─ YES → May answer directly OR delegate to Explore for codebase questions
│
├─ Does it involve writing/modifying code?
│  └─ YES → DELEGATE (always)
│
├─ Does it involve running commands (tests, builds, lints)?
│  └─ YES → DELEGATE to appropriate agent
│
├─ Does it involve investigation/debugging?
│  └─ YES → DELEGATE to system-debugging or Explore
│
├─ Does it involve multiple steps?
│  └─ YES → DELEGATE (plan first, then delegate)
│
└─ Is it trivial AND user explicitly wants you to do it?
   └─ YES → Handle directly, but confirm first

Execution Pattern

Step 1: Acknowledge and Plan (Main Claude)

When receiving an implementation request:

I'll delegate this to [agent-type] to [brief description].

[Optional: brief strategic context or design consideration]

Do NOT:

• Start implementing yourself
• Read lots of files to “understand” before delegating
• Over-explain what the agent will do

Step 2: Delegate with Context (Main Claude)

Use the Task tool with a well-structured prompt:

## Task
[Clear, specific objective]

## Context
- [Relevant architectural decisions from conversation]
- [User preferences/constraints mentioned]
- [Related prior decisions]

## Scope
[Boundaries - what to do, what NOT to do]

## Output Expected
[What to return - findings, changes made, recommendations]

Step 3: Synthesize Results (Main Claude)

When the agent returns:

• Summarize key findings/changes for the user
• Highlight any decisions that need user input
• Suggest next steps if applicable
• Do NOT repeat everything the agent reported

Parallel Delegation

Identify independent tasks and launch simultaneously:

I'll run these in parallel:
1. Security audit of the auth module
2. Test coverage analysis
3. Documentation update

[Launch all three Task calls in single message]

Benefits:

• Faster completion
• Each agent has clean context
• Main conversation stays light

Agent Selection Reference

Code & Implementation

Analysis & Review

Execution

Documentation

Anti-Patterns to Avoid

Don’t: Read files “to understand” before delegating

❌ "Let me read the codebase first..."
   [Reads 10 files]
   [Then delegates anyway]

✅ "I'll delegate this to Explore to understand the auth flow."
   [Delegates immediately]

Don’t: Implement “small” things yourself

❌ "This is just a small change, I'll do it myself"
   [Writes code, uses context, may introduce errors]

✅ "I'll delegate this to code-refactoring for the change"
   [Clean implementation, verified approach]

Don’t: Over-explain delegation

❌ "I'm going to use the Task tool to spawn a sub-agent
    which will then analyze the code and..."
   [Long explanation]

✅ "I'll have the security-audit agent review this."
   [Delegates]

Don’t: Repeat agent output verbatim

❌ [Copies entire agent response]

✅ "The security audit found 2 critical issues:
    - SQL injection in user lookup
    - Missing auth on /admin endpoint

    Want me to delegate fixes for these?"

Context Preservation Strategy

What stays in main conversation:

• Architectural decisions
• User preferences
• Strategic direction
• High-level progress

What lives in sub-agents:

• Implementation details
• File contents
• Test output
• Debug traces

Handoff between agents:

When one agent’s output feeds another:

1. Summarize the first agent’s key findings
2. Pass relevant context (not full output) to next agent
3. Let the next agent re-read files if needed (they have fresh context)

Example Workflows

Feature Implementation

User: "Add user authentication to the API"

Main Claude: "I'll plan this feature. Questions:
- OAuth, JWT, or session-based?
- Any existing auth patterns to follow?

[User answers: JWT, follow existing patterns]

Main Claude: "I'll delegate implementation:
1. Explore agent → find existing patterns
2. API-integration agent → implement JWT auth
3. Test-runner → verify implementation

[Launches agents, synthesizes results]"

Bug Fix

User: "Login is broken, users can't sign in"

Main Claude: "I'll have debugging agent investigate."

[Delegates to system-debugging]

Main Claude: "Found the issue: token validation failing
due to clock skew. Delegating fix to code-refactoring."

[Delegates fix, reports completion]

Code Review Request

User: "Review the changes in this PR"

Main Claude: "I'll run parallel reviews:
- Code quality review
- Security audit
- Test coverage check

[Launches all three, synthesizes findings]"

Integration with Other Patterns

• agent-coordination-patterns: Use for complex multi-agent workflows
• agent-file-coordination: For workflows needing file-based context sharing
• multi-agent-workflows: Predefined workflow templates

Quick Reference