Run any Skill in Manus with one click

subagent-first

Stars3

Forks1

UpdatedJune 13, 2026 at 23:16

Orchestrate work by delegating to sub-agents. Main agent handles only planning, judgment, and user communication. Always active for multi-step work.

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Run Skill in Manus

Source

iimuz

iimuz/dotfiles

View GitHub Repository View Creator Repositories

Download

Run Skill in Manus

Related occupationsSOC

Based on SOC occupation classification

Software DevelopersComputer and Mathematical Occupations·SOC 15-1252

SKILL.md

readonly

name	subagent-first
description	Orchestrate work by delegating to sub-agents. Main agent handles only planning, judgment, and user communication. Always active for multi-step work.
user-invocable	true
disable-model-invocation	false

Subagent-First

Overview

Preserve the main agent's context window by delegating all investigation, analysis, execution, and validation to sub-agents. The main agent acts strictly as an orchestrator: planning, high-level judgment, user communication, and state tracking.

Main Agent Boundaries

DO NOT read or edit codebase files directly. All file modifications, tests, and analysis must be done via sub-agents.
Exception: Small, self-contained inline tasks (under 50 lines of expected output) that are directly needed for the next immediate decision may be handled by the main agent. However, lint, test, and build commands must ALWAYS be delegated to a sub-agent regardless of size.

Dispatch Rules

Minimize Return Context: Sub-agents must write verbose outputs (e.g., test logs, lint results, raw search results) to {session_dir}/ and return ONLY:
1. Success or failure status
2. File path to the full log in {session_dir}/
3. A concise summary of items requiring action (e.g., specific test failures)
Explicit Context: Always provide each sub-agent with:
- Exact file paths (never let them guess)
- Specific scope: what to do AND what NOT to do
- Expected output format or success signal

session_dir resolves to ~/.copilot/session-state/{session_id}/files/.

Model Selection

When invoking a sub-agent, choose the cheapest model that is sufficient for the task:

claude-haiku-4.5: Default for trivial work: status checks, simple reads/searches, basic commands, or short summaries.
gpt-5.3-codex: Default for coding work: codebase inspection, targeted edits, debugging, test-fix loops, and medium-to-complex implementation.
claude-sonnet-4.6: Use for deep refactoring, architecture analysis, complex test design, or decisions requiring broad engineering judgment.
gpt-5.4: Use for second opinions, alternative perspectives, or review of another model's plan/output.

More from this repository

same repository

gh-ops

iimuz/dotfiles

Use for GitHub operations: creating issues, posting issue comments, creating draft PRs, and adding PR review comments.

2026-07-163

scamper

iimuz/dotfiles

Facilitate structured divergent idea generation using the SCAMPER technique (Bob Eberle). An orchestrator elicits a concrete target, applies seven transformation lenses (Substitute, Combine, Adapt, Modify, Put to another use, Eliminate, Reverse) in parallel via focused subagents, then converges into evaluated, prioritized ideas. Use when the user asks for a SCAMPER analysis, systematic ideation, or improvement ideas for an existing product, service, process, or idea.

2026-07-043

six-hats

iimuz/dotfiles

Facilitate structured multi-perspective analysis using Edward de Bono's Six Thinking Hats. A Blue Hat orchestrator applies one hat at a time in sequence over a single shared record, using five focused hat subagents. Use when the user asks for a Six Hats or 6 hat analysis, a multi-perspective breakdown of a decision, idea, or problem, or systematic structured thinking on a topic.

2026-07-043

markmap

iimuz/dotfiles

Author markmap-format Markdown. Use when writing a mindmap (マインドマップ / markmap) in Markdown.

2026-07-043

sample-greeting

iimuz/dotfiles

Sample skill for verifying plugin wiring. Trigger when the user asks to run the sample greeting or test the sample plugin.

2026-07-033

gh-pr-resolve

iimuz/dotfiles

Use when resolving or addressing PR review comments.

2026-07-023

Subagent-First

Overview

Main Agent Boundaries

DO NOT read or edit codebase files directly. All file modifications, tests, and analysis must be done via sub-agents.

Exception: Small, self-contained inline tasks (under 50 lines of expected output) that are directly needed for the next immediate decision may be handled by the main agent. However, lint, test, and build commands must ALWAYS be delegated to a sub-agent regardless of size.

Dispatch Rules

Minimize Return Context: Sub-agents must write verbose outputs (e.g., test logs, lint results, raw search results) to {session_dir}/ and return ONLY:

Success or failure status
File path to the full log in {session_dir}/
A concise summary of items requiring action (e.g., specific test failures)

Explicit Context: Always provide each sub-agent with:

Exact file paths (never let them guess)
Specific scope: what to do AND what NOT to do
Expected output format or success signal

session_dir resolves to ~/.copilot/session-state/{session_id}/files/.

Model Selection

When invoking a sub-agent, choose the cheapest model that is sufficient for the task:

claude-haiku-4.5: Default for trivial work: status checks, simple reads/searches, basic commands, or short summaries.

gpt-5.3-codex: Default for coding work: codebase inspection, targeted edits, debugging, test-fix loops, and medium-to-complex implementation.

claude-sonnet-4.6: Use for deep refactoring, architecture analysis, complex test design, or decisions requiring broad engineering judgment.

gpt-5.4: Use for second opinions, alternative perspectives, or review of another model's plan/output.