metagent

name	metagent
description	Design and build new agents from user goals. Use this skill whenever the user wants a new agent, subagent, assistant, tool-calling workflow, coordinator-specialist system, or asks how to architect prompts, tools, memory, guardrails, or evals for an agent — even if they do not explicitly ask for 'agent architecture'.

Metagent: Build New Agents

This file is the top-level frame for the metagent skill.

Keep SKILL.md lightweight:

• global routing logic lives here
• core knowledge for this skill lives in handbook.md, examples.md, and rules.md
• deeper reference material for this standalone package is folded into the three sibling files above
• detailed chapter knowledge should not be duplicated back into this file

Purpose

Use this skill whenever the main agent needs to design, implement, review, or explain agents and agentic systems.

Typical triggers:

• the user asks to build a new agent or assistant
• the user wants a tool-using workflow, subagent, or multi-agent system
• the user wants help choosing prompts, tools, memory, guardrails, or evals for an agent
• the user has an agent design but it is incomplete, unsafe, or hard to evaluate
• the user asks conceptual questions about how agents should be built or why a given pattern is used

Global context

At the highest level:

• start from the user’s job to be done, not from a preferred architecture
• an agent is more than an LLM; runtime, prompts, tools, memory, guardrails, planning, and evaluation all matter
• use the smallest agent shape that can solve the task
• use examples to ground implementation details
• use rules as the review gate before recommending or finalizing a design

Modes

1. Design mode

Use when the task is to architect a new agent or decide how an agent should work.

Primary action:

• read handbook.md

Output focus:

• mission
• trigger conditions
• architecture choice
• tools, memory, guardrails, and eval shape

2. Implement mode

Use when the task is to turn an approved or implied design into a concrete implementation shape, prompt harness, runtime pattern, or tool contract.

Primary action:

• read examples.md

Output focus:

• concrete patterns to reuse
• implementation scaffolding
• example-backed choices
• anti-pattern avoidance

3. Review mode

Use when the task is to critique, validate, or improve an existing agent design or implementation.

Primary action:

• pass the design or implementation through rules.md

Output focus:

• violations
• missing constraints
• safety and eval gaps
• recommended fixes

4. Knowledge-answer mode

Use when the user is asking conceptual questions and does not necessarily want a new agent built right now.

Primary action:

• answer from skill knowledge, starting with handbook.md
• pull examples.md when concrete illustration helps
• pull rules.md when the answer should end in operational guidance

Output focus:

• direct explanation
• grounded tradeoffs
• concise recommendations

Pipeline

1. Classify the request into one primary mode: design, implement, review, or knowledge-answer.
2. Read the mode’s primary file first:
   • design -> handbook.md
   • implement -> examples.md
   • review -> rules.md
   • knowledge-answer -> handbook.md
3. Pull supporting material from the local companion files only when the task needs deeper detail:
   • handbook.md for architecture, tools, memory, guardrails, evals, and operating models
   • examples.md for reusable patterns, prompt harnesses, and anti-patterns
   • rules.md for review gates, safety checks, and operating constraints
4. Produce the deliverable that matches the mode:
   • design -> agent specification
   • implement -> concrete implementation guidance or artifacts
   • review -> findings plus prioritized fixes
   • knowledge-answer -> direct explanation
5. Before finalizing substantial design or implementation guidance, run a review pass against rules.md.

Local file contract

The local metagent directory should contain:

• SKILL.md — routing frame and usage policy
• handbook.md — distilled concepts and mental models
• examples.md — good patterns, anti-patterns, and source-backed snippets
• rules.md — actionable review criteria and operating rules

Maintenance rule

When evolving this skill:

1. keep SKILL.md focused on mode selection and pipeline
2. put conceptual material into handbook.md
3. put concrete implementation patterns into examples.md
4. put review criteria into rules.md
5. keep the package self-contained and prefer extending the three companion files over adding extra directories
6. make sure every addition helps the main agent build, refine, or evaluate other agents