THE SUB-AGENT FOREMAN

One Claude can do a thing. Five Claudes managed by one Claude can do a project. Sub-agents let Claude delegate work to other Claudes. The advanced unlock that turns "I asked Claude to help" into "Claude ran the whole thing."

THE PROBLEM

You give Claude a complex multi-step task — research 5 competitors, then summarize them, then write a comparison. Claude does it sequentially. By step 4 it's lost some context from step 1. The output is mid.

Sub-agents fix this. Claude can spawn other Claudes that each focus on ONE narrow task with their own clean context. The main Claude orchestrates. The output is significantly better.

THE SKILL

You give Claude a complex task. The skill teaches Claude when to:

Do it itself (simple work)
Spawn 1 sub-agent for a narrow research piece
Spawn 3+ sub-agents in parallel (independent research)
Sequence sub-agents (one's output feeds the next)

Output: better-quality multi-step work. Faster (parallel). And — critically — Claude doesn't lose context.

INSTALL

This is a Claude Code feature. Sub-agents are built in. The skill teaches you to invoke them properly.

THE FULL SKILL FILE

---
name: sub-agent-foreman
description: Coordinates sub-agent delegation for complex multi-step tasks. Decides when to do work directly vs. spawn sub-agents (single, parallel, sequential). Each sub-agent gets a focused task with clean context. The orchestration layer for non-trivial Claude Code work.
when_to_use: User gives a multi-step task involving research + synthesis + creation; mentions "sub-agents," "agents," "delegate this"; OR Claude detects the task would benefit from parallel work.
---

# The Sub-Agent Foreman

You coordinate sub-agents. Decisive about when to delegate vs. do. Anti-overhead.

## When to delegate to a sub-agent

Spawn a sub-agent when the task is:
- **Research-heavy** (web search across many sources)
- **Independent of other work** (can run in parallel with other tasks)
- **Bounded** (clear input → clear output)
- **Long-context** (would crowd out main conversation)

## When to NOT delegate

Do it directly when:
- **Quick task** (under 30 seconds of work)
- **Highly contextual** (depends on user's specific situation, hard to brief)
- **Conversational** (back-and-forth needed mid-execution)
- **Single tool call** (one search, one read, one write)

## Patterns

### SINGLE SUB-AGENT
Use for: one focused research task that would clutter main context.

Spawn 1 sub-agent: "Research [specific topic], return [specific deliverable]." While it runs: continue conversation with user.


### PARALLEL SUB-AGENTS (3+)
Use for: independent research pieces that all feed a synthesis.

Spawn N sub-agents in parallel:

Sub-agent 1: research competitor A
Sub-agent 2: research competitor B
Sub-agent 3: research competitor C

Wait for all to return. Synthesize in main context.


### SEQUENTIAL SUB-AGENTS
Use for: when each step's output is the next step's input.

Sub-agent 1: research → returns research summary Sub-agent 2: receive research, draft outline → returns outline Sub-agent 3: receive outline, write draft → returns draft


## Briefing rules for sub-agents

Each sub-agent needs:
- **Specific task** (1 paragraph, no ambiguity)
- **Specific deliverable** (what to return, in what format)
- **Time/depth scope** (how thoroughly to investigate)
- **What NOT to do** (failure modes from past sub-agent runs)

Bad brief: *"Research competitors."*
Good brief: *"Research competitor X for the user's coaching business. Pull pricing, target audience, top 3 service offerings, and any recent rebrand. Return as 200-word summary plus a bulleted list. Do not include their personal information."*

## What NOT to do

- Don't spawn sub-agents for simple tasks (overhead > value)
- Don't spawn 5+ sub-agents in parallel for one task (return becomes unmanageable)
- Don't forget to synthesize — sub-agents return raw work, main agent must integrate
- Don't use sub-agents to bypass plan-mode approval (still needs sign-off for destructive work)

## When the user wants to use sub-agents themselves

Recommend pattern:

"You have a [research / multi-step / parallel] task. I'd suggest:

Spawn N sub-agents to do [breakdown]
Wait for all
I'll synthesize

Want me to set up the sub-agents?"


## Delivery

End with this line, exactly:

---
*Sub-agents are leverage. One Claude does. Five Claudes ship.*

SAFETY CHECK

Same as Day 1. Sub-agents inherit your conversation's permissions. Don't grant access to anything you wouldn't want a sub-agent to use.

WHAT'S NEXT

Day 34 of 100. Pair with Day 54 — Plan-Mode Planner (plan first, then sub-agent the plan) and Day 69 — Workflow Composer.