delegation-advisor

name	delegation-advisor
description	Advises on when to delegate to subagents. Use when user is working on complex multi-file tasks, exploring codebases, researching, implementing large features, or needs help deciding if delegation would be beneficial.
title	Delegation Advisor
stratum	2
branches	["agentic"]

Purpose

I help you decide when to delegate to subagents. When I detect a task that would benefit from a fresh context, I ASK you first rather than auto-invoking.

Key principle: Delegation should be explicit and user-controlled.

---

Recognition Patterns

I suggest delegation when I detect:

Pattern	Suggested Agent	Confidence
"How does X work?" + codebase	Explore	High
Multi-file changes needed	Plan	High
Research/synthesis tasks	knowledge-curator	Medium
"Create a skill for..."	skill-writer	High
"Create an agent that..."	agent-writer	High
Scaffold/workflow questions	workflow-expert	High
New project/vault setup	seacow-scaffolder	High
Doc maintenance needed	claude-md-updater	Medium

---

How I Ask

When I recognize an opportunity, I use AskUserQuestion:

"This looks like [task type]. Would you like me to:
 • Spawn [agent] agent (fresh context, deep work)
 • Handle directly (uses current context)
 • Tell me more first"

I respect your choice. Quick "handle directly" is always an option.

---

Tool-Specific Delegation

Claude Code

When you approve delegation:

1. I use the Task tool with subagent_type: Explore, Plan, or general-purpose
2. I provide context summary to the agent
3. Agent works in fresh context, returns summary
4. I synthesize results for you

Built-in types available:

• Explore - Fast codebase exploration
• Plan - Design implementation approach
• general-purpose - Complex multi-step tasks

OpenCode

When you approve delegation:

1. I can spawn child sessions with specific agents
2. For large tasks, I may suggest parallel exploration
3. Each child returns distilled summaries
4. I synthesize across all children

OpenCode enables:

• TRUE recursive agents (agents spawning agents)
• Isolated context per session
• Different LLMs per level
• Session navigation (parent/child)

---

Parallel Delegation (OpenCode Only)

For large codebases, I may offer:

"This codebase has multiple major areas. Would you like:
 • Parallel agents for /src/, /api/, /docs/ (faster)
 • Single depth-first exploration (sequential)
 • Handle directly"

---

Quick Decline

If you prefer direct handling, just say:

• "Handle it directly"
• "No delegation needed"
• "Just do it"

I'll adapt to your preference for similar future tasks.

---

When NOT to Suggest Delegation

• Simple questions with obvious answers
• Single-file operations
• Tasks you've indicated preference to handle directly
• When context is already loaded and sufficient

---

Agent Dispatch Reference

You're asking about...	Suggested Agent	Why
How codebase works	Explore	Multi-file discovery
Implementation approach	Plan	Design before code
This workflow scaffold	workflow-expert	Deep guide knowledge
New project structure	seacow-scaffolder	SEACOW-based design
Creating skills	skill-writer	Follows conventions
Creating agents	agent-writer	Follows conventions
Doc accuracy	claude-md-updater	Maintenance checks
Knowledge synthesis	knowledge-curator	Research + organize

---

Installation

Global (All Projects)

cp -r .claude/skills/delegation-advisor/ ~/.claude/skills/

Per-Project

Keep in your project's .claude/skills/ directory.

Hybrid (Recommended)

• Global: delegation-advisor/SKILL.md (universal delegation)
• Project: Domain-specific skills and agents

---

See Also

• workflow-meta skill - Scaffold-specific concierge
• workflow-expert agent - Deep scaffold knowledge
• .claude/ARCHITECTURE.md - Tool capabilities and constraints
• docs/tool-comparison.md - Claude Code vs OpenCode details