create-agent

name	create-agent
description	Create a new .agent.md file for the investment partner agentic system. Use when building a new agent, adding a subagent, or scaffolding agent definitions. Guides through template, authoring rules, tool selection, and handoff design.
argument-hint	Name and purpose of the agent to create

Create Agent

When to Use

• Building a new agent for the investment partner system
• Adding a subagent to handle a specific responsibility
• Scaffolding an agent definition from a design document in artifacts/

Procedure

1. Ask the user for the agent's name and purpose (one sentence)
2. Determine project prefix (if multi-project workspace):
   • Check projects/ directory for multiple agentic systems
   • Use domain abbreviation as prefix: lf- (litigation finance), ip- (investment partner), hr- (human resources)
   • Examples: lf-orchestrator, ip-research-agent, hr-onboarding-agent
   • Single project workspace: no prefix needed
3. Determine scope:
   • Default: omit user-invokable — agent is visible in agent picker
   • Subagent-only: uncomment # user-invokable: false to hide from picker
4. Read the agent template as the starting scaffold
5. Select tools from the tool registry — pick only what this agent needs. The registry includes input shapes for every MCP tool and a per-agent assignment matrix.
6. Define handoffs using the system handoff map:
   • Inbound: which agents call this one, with what payload shape
   • Outbound: which agents this one delegates to, with trigger conditions and payloads
7. Write the agent body following authoring rules:
   • Role line (one sentence)
   • Constraints (DO NOT / ONLY rules)
   • Procedure (numbered steps with explicit #tool: and @agent references)
   • Handoffs (inbound + outbound with structured payloads)
   • Output format (table, JSON, or Mermaid)
8. Save to .github/agents/{[prefix-]agent-name}.agent.md
9. Verify: under 150 lines body, no prose paragraphs, all tools named explicitly, all handoff payloads structured

Validation Checklist

• name in frontmatter is kebab-case matching the filename (e.g., lf-compliance-agent)
• Project prefix used consistently in multi-project workspace
• description in frontmatter is keyword-rich and specific
• tools array is minimal — no unused tools
• Every tool in Procedure uses #tool: syntax
• Every subagent call uses @agent-name syntax
• Handoffs section has both Inbound and Outbound
• All handoff payloads are structured { key: type }
• Output section specifies exact format
• Body is under 150 lines
• No prose paragraphs — bullets and numbered steps only
• Imperative voice throughout