agent-creator

agent-creator

A Claude Code plugin for building, testing, and iterating on custom agents — the markdown files that define specialized Claude personas you can start with claude --agent name or invoke mid-session with @agent-name.

Why agents matter

Agents are the foundation that skills delegate to. When you type create a Go expert agent or /code-review, Claude often hands off to a specialized agent under the hood. That agent's quality determines the quality of the entire interaction.

Most agents are poorly designed — not because the author didn't try, but because there's almost no guidance on how to build them. People guess at which tools to give an agent, write vague system prompts, and never verify that the agent actually behaves as intended. The result: a narrowly scoped agent that drifts out of scope, a task agent that hallucinates steps, or a skill that calls an agent that breaks silently.

This plugin solves that. It guides you through designing an agent correctly, tests it against realistic scenarios, and iterates on the system prompt automatically until it behaves the way you intended.

Quick Start

claude plugin marketplace add lgbarn/agent-creator
claude plugin install agent-creator@agent-creator

Then in Claude Code, just describe what you want:

create a Go expert agent
build me a code review agent
my security-advisor isn't staying in scope

Or invoke the skill directly:

/agent-creator

What this plugin does

1. Design — Asks the right questions to determine agent type, scope, tool access, and system prompt structure. Applies smart defaults so you only make decisions that actually matter.
2. Test — Validates the agent file, runs multi-turn test scenarios against it, and grades behavioral assertions — including whether the agent stays in scope when asked about things outside its domain.
3. Iterate — Runs an automated improvement loop that identifies failures, rewrites the system prompt using extended thinking, re-tests, and saves the best-performing version.

What you're building

Here's a complete agent file — a Go expert you can talk to about architecture, idioms, and code review:

---
name: go-expert
description: Go programming expert for architecture, idioms, and code review
model: inherit
color: blue
tools: Read, Grep, Glob
---

<role>
You are a senior Go engineer with 10+ years of production Go experience. You
specialize in idiomatic Go, concurrency patterns, error handling, and performance
optimization. You're direct — you lead with answers, not caveats.
</role>

<knowledge>
You know Go deeply: the standard library, common patterns (functional options,
table-driven tests, context propagation), and the reasoning behind Go's design
decisions. Your expertise stops at Go — you don't advise on Python, JavaScript,
or other languages.
</knowledge>

<style>
Direct. Lead with the answer, then explain. Include code examples. When something
is genuinely debatable, say so and explain the tradeoffs. No unnecessary hedging.
</style>

<rules>
- Stay within Go. If asked about another language, acknowledge it and redirect.
- Never recommend patterns the Go community considers bad practice (global state,
  empty interfaces used as shortcuts, ignoring errors).
- If you're uncertain, say so rather than guessing.
</rules>

Each part serves a specific purpose:

Part	What it controls
name	How you invoke it: claude --agent go-expert or @go-expert
description	When Claude auto-delegates to it in normal sessions
model	Which Claude model it uses (inherit = same as your current session)
color	Terminal color for the agent's responses
tools	What the agent can do (Read, Grep, Glob = read-only; Bash, Write, Edit = task agent)
<role>	Who the agent is and what it knows — be specific about expertise depth
<knowledge>	What it knows deeply and where its boundaries are
<style>	How it communicates — don't leave this to chance
<rules>	Hard behavioral constraints — what it must never do

Two archetypes

Conversational agents — domain experts you talk to. The system prompt defines expertise, boundaries, and conversation style. Tools are read-only (Read, Grep, Glob) or none. These live in ~/.claude/agents/ and are available across all your projects.

Examples: go-expert, security-advisor, recipe-helper, interview-coach

Task agents — automated workers that perform specific actions. The system prompt defines procedures, verification steps, and output format. Tools include Bash, Write, Edit. These often live in .claude/agents/ (project-scoped, version-controlled) because they're tied to a specific codebase.

Examples: code-reviewer, test-generator, deployment-helper