multi-agent-split

A lightweight Claude Code / Codex Skill for deciding when to split work across native subagents.

The default is simple: do not split. Use this Skill only when the user explicitly asks for multi-agent work, subagents, or parallel review, or when a host agent has a clear need for 2-3 independent read-only checks.

Tags

ai-skill multi-agent subagents codex claude-code privacy-safe

What It Does

• Decides whether a task should be split at all.
• Keeps subagents narrow, scoped, and read-only by default.
• Provides a delegation prompt for independent subagent checks.
• Helps the host agent judge and merge subagent findings.
• Allows self-improvement only as proposed rules, not automatic Skill edits.

When To Use

Use it for:

• Explicit multi-agent or subagent requests
• Parallel review across 2-3 independent questions
• Tasks where separate checks reduce blind spots or context pollution

Do not use it for:

• Simple Q&A, translation, or explanation
• Small edits or single-file changes
• Writing that needs one unified final voice
• Ordinary reviews that the host agent can handle in one pass

Quick Start

Clone the repository:

git clone https://github.com/kelin3296-jpg/multi-agent-split.git
cd multi-agent-split

Install for Claude Code:

mkdir -p "$HOME/.claude/skills"
ln -s "$PWD/multi-agent-split" "$HOME/.claude/skills/multi-agent-split"

Install for Codex:

mkdir -p "$HOME/.codex/skills"
ln -s "$PWD/multi-agent-split" "$HOME/.codex/skills/multi-agent-split"

If a target already exists, remove or rename the old target first.

Usage

Ask Claude Code or Codex:

Use multi-agent-split to decide whether this task should be split:
review this PRD with product value, execution risk, and MVP scope in parallel.

For agent-assisted installation, see docs/INSTALL_WITH_AGENT.md.

Expected Output

For a simple task, the Skill should keep the work in one pass:

Split: no
Reason: Simple translation task; one pass is enough.
Subtasks if split: none

For a suitable multi-agent task, it should first define narrow subtasks:

Split: yes
Reason: The task has three separable review questions.
Subtasks if split:
1. Product value review
2. Execution risk review
3. MVP scope review

Project Structure

multi-agent-split/
|-- multi-agent-split/
|   `-- SKILL.md
|-- docs/
|   `-- INSTALL_WITH_AGENT.md
|-- README.md
|-- LICENSE
`-- .gitignore

Privacy

This repository does not include private task records, account data, API keys, local absolute paths, or full subagent outputs.

The Skill also tells host agents not to store private user details, account information, local paths, or one-off task records in future Skill updates.

License

MIT