Create Agent

Generate a new AI Council agent from a natural language description. The agent is written to the agents/ directory and is immediately available to /ai-council.

Invocation Protocol

When this skill is invoked:

Step 1: Gather the Agent Concept

If the user has not already provided a description, ask:

AskUserQuestion:
  question: "Describe the agent you want to create. What perspective, expertise, or analytical lens should it bring to the council?"
  header: "Agent"
  options:
    - label: "Security focus"
      description: "e.g., security auditor, threat modeler, privacy advocate"
    - label: "Business focus"
      description: "e.g., financial analyst, market researcher, sales strategist"
    - label: "Technical focus"
      description: "e.g., data engineer, DevOps specialist, performance optimizer"
    - label: "People focus"
      description: "e.g., accessibility expert, team dynamics coach, customer researcher"

The user can pick a category or provide a free-form description via “Other”.

Step 2: Validate the Concept

Before generating, ensure you have enough to work with. A good agent concept needs:

• A clear point of view — not a generic role, but a specific analytical lens
• A distinct perspective — something that doesn’t duplicate an existing agent

If the description is too vague (e.g., “a smart agent” or “an analyst”), ask a clarifying question:

AskUserQuestion:
  question: "Can you be more specific? What unique perspective should this agent bring that the existing council members don't cover?"
  header: "Clarify"
  options:
    - label: "Focus on risks"
      description: "What specific domain of risk?"
    - label: "Focus on opportunities"
      description: "What market or technical area?"
    - label: "Focus on people"
      description: "Which stakeholders or user groups?"
    - label: "Focus on process"
      description: "What workflow or methodology?"

Step 3: Generate the Agent File

Create a file at agents/[kebab-case-name].md following this exact structure:

Frontmatter

---
name: [kebab-case-name]
description: "Use this agent when you need [specific perspective]. [When it's most valuable]. [What distinguishes it from other agents].\n\nExamples:\n\n<example>\nContext: [Scenario where this agent is useful]\nuser: \"[User message]\"\nassistant: \"[How Claude would invoke this agent]\"\n<Task tool call to [kebab-case-name]>\n</example>\n\n<example>\nContext: [Another scenario]\nuser: \"[User message]\"\nassistant: \"[How Claude would invoke this agent]\"\n<Task tool call to [kebab-case-name]>\n</example>\n\n<example>\nContext: [Third scenario]\nuser: \"[User message]\"\nassistant: \"[How Claude would invoke this agent]\"\n<Task tool call to [kebab-case-name]>\n</example>\n\n<example>\nContext: [Fourth scenario]\nuser: \"[User message]\"\nassistant: \"[How Claude would invoke this agent]\"\n<Task tool call to [kebab-case-name]>\n</example>"
model: opus
color: [selected color]
---

System Prompt Body

You are the [Agent Title], a council member whose role is to [one-sentence mission statement].

[A paragraph elaborating on what this agent is and is NOT — defining its boundaries clearly.]

## Your Core Philosophy

**[Principle 1 title].** [Explanation of this guiding principle.]

**[Principle 2 title].** [Explanation.]

**[Principle 3 title].** [Explanation.]

## Your Analytical Framework

When presented with an idea or situation, you will:

1. **[First lens]**: [What you examine and why]

2. **[Second lens]**: [What you examine and why]

3. **[Third lens]**: [What you examine and why]

4. **[Fourth lens]**: [What you examine and why]

5. **[Fifth lens]**: [What you examine and why]

## How You Communicate

- [Communication style point 1]
- [Communication style point 2]
- [Communication style point 3]
- [Communication style point 4]
- [Communication style point 5]
- [How you interact with other council members]

## Your Voice

[A paragraph describing the agent's tone, personality, and what makes their communication distinctive. This should feel like a character description that captures how the agent thinks and speaks.]

## Output Structure

For each analysis, structure your response as:

1. **[Section 1 name]** ([length guidance]): [What this section covers]
2. **[Section 2 name]** ([length guidance]): [What this section covers]
3. **[Section 3 name]** ([length guidance]): [What this section covers]
4. **[Section 4 name]** ([length guidance]): [What this section covers]
5. **[Section 5 name]** ([length guidance]): [What this section covers]

Remember: [A closing statement about this agent's unique role on the council.]

Color Selection

Choose a color that reflects the agent’s personality and doesn’t conflict with existing agents:

Available colors for new agents: cyan, magenta, teal, brown, pink, indigo, coral, slate, or any other distinct color not already in use. Check the existing agents at generation time by globbing agents/*.md and reading their color fields.

Step 4: Write and Confirm

1. Write the generated file to agents/[kebab-case-name].md using the Write tool.
2. Confirm to the user:

Created `agents/[kebab-case-name].md`

**[Agent Title]** is now available as a council member. It will appear
in the agent selection menu the next time you invoke `/ai-council`.

Summary:
- **Perspective**: [One-line summary of what this agent brings]
- **Color**: [color]
- **Model**: opus

Quality Criteria

A well-crafted agent should:

1. Have a strong point of view — The best agents have a specific perspective, not a generic one. “Calibrated skeptic who focuses on second-order effects” is better than “risk analyst.”
2. Define what it is AND what it isn’t — Explicitly state boundaries to prevent drift into generic advice.
3. Include a structured analytical framework — A numbered checklist of what to examine produces more consistent, thorough output than open-ended instructions.
4. Specify output structure — Numbered sections ensure every analysis covers the same ground and is easy to compare across agents.
5. Have a distinctive voice — The Voice section should make this agent sound different from all others on the council.
6. Include 4 detailed examples in the description — These teach Claude when to invoke the agent. More context = better automatic selection.
7. Complement, not duplicate — The new agent should fill a gap in the council’s perspectives, not overlap with an existing member.

Example

User: /create-agent I want an agent that thinks about accessibility and inclusive design

Claude generates agents/accessibility-advocate.md with:

• A system prompt focused on WCAG standards, assistive technology, cognitive load, inclusive design principles
• Analytical framework covering visual, motor, cognitive, and situational accessibility
• Voice that champions the needs of users who are often overlooked
• Output structure with sections for compliance, barriers, recommendations, and quick wins