---
title: "Agent Development for Claude Code Plugins: review, setup & limitations · Undominated.ai"
canonical: https://undominated.ai/skills/anthropic-agent-development/
description: "Explains how to write a Claude Code plugin agent file (frontmatter, trigger description, system prompt) and ships a Bash script that checks one agent file."
---

# Agent Development for Claude Code Plugins: review, setup & limitations · Undominated.ai

> Explains how to write a Claude Code plugin agent file (frontmatter, trigger description, system prompt) and ships a Bash script that checks one agent file.

[← Explore all skills](/skills/)

AGENT DEVELOPMENT / Anthropic

# Agent Development for Claude Code Plugins

Explains how to write a Claude Code plugin agent file (frontmatter, trigger description, system prompt) and ships a Bash script that checks one agent file.

 [See setup guidance ↓](#setup)[Original source ↗](https://github.com/anthropics/claude-plugins-official/tree/41f32e407b450055044f167db7ded5e8ff67d203/plugins/plugin-dev/skills/agent-development)

SOURCE REVIEW

 Reviewed 2026-10-21
 Evidence 6 linked sources
 Publisher Anthropic
 Licence [Apache-2.0 ↗](https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/plugins/plugin-dev/LICENSE)
 Revision 41f32e407b45
 [Read what was—and wasn’t—checked ↓](#review)

“This skill should be used when the user asks to "create an agent", "add an agent", "write a subagent", "agent frontmatter", "when to use description", "agent examples", "agent tools", "agent colors", "autonomous agent", or needs guidance on agent structure, system prompts, triggering conditions, or agent development best practices for Claude Code plugins.”

 [Anthropic · upstream description ↗](https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/plugins/plugin-dev/skills/agent-development/SKILL.md) Our analysis follows below.

01 / THE REASONING

## Why this made the selection.

 - It sets the agent frontmatter rules: a name of 3-50 lowercase letters, numbers and hyphens, a required description, model (inherit, sonnet, opus or haiku) and color, and an optional tools list it says to keep to the minimum needed.
- It ships scripts/validate-agent.sh, which checks an agent file for frontmatter markers, the four required fields, the name format and the system-prompt length, and it includes an agent-generation prompt it says was extracted from Claude Code.

### A good fit for

 - Write an agents/<agent-name>.md file for a Claude Code plugin with name, description, model, color and an optional tools list
- Draft an agent description that names two to four trigger scenarios in prose and points to a "When to invoke" section in the body
- Check one agent file's frontmatter and system-prompt length with scripts/validate-agent.sh before committing it

### Weigh up before choosing

 - Last commit touching plugins/plugin-dev/skills/agent-development at revision 41f32e40: 2026-04-28 (ce721c1f).
- The SKILL.md lists scripts/test-agent-trigger.sh under Utility Scripts, but the skill's scripts/ directory at this revision holds only validate-agent.sh.
- validate-agent.sh prints a warning when a description has no <example> block, while the SKILL.md tells authors to name trigger scenarios in prose and keep worked scenarios in a "When to invoke" body section.
- The SKILL.md says an agent with no tools field gets access to all tools; the restriction is left to the author.
- The License section of the plugin-dev README reads "MIT License - See repository for details", while plugins/plugin-dev/LICENSE at the same revision is the Apache License, Version 2.0. This record follows the LICENSE file.
- The README's install line names a marketplace called claude-code-marketplace. The repository's marketplace file is named claude-plugins-official, and its plugin-dev entry gives anthropics/claude-plugins-public as the homepage.

02 / THE REVIEW RECORD

## What we actually inspected.

Source review has boundaries. A clear record is more useful than a “safe” badge.

### Material inspected

 - plugins/plugin-dev/skills/agent-development/SKILL.md
- plugins/plugin-dev/skills/agent-development/scripts/validate-agent.sh
- plugins/plugin-dev/README.md
- plugins/plugin-dev/LICENSE
- .claude-plugin/marketplace.json (plugin-dev entry)
- README.md

### Our findings

 - The SKILL.md frontmatter declares no allowed-tools. The skill is Markdown guidance plus one Bash validator.
- Validation limits in the SKILL.md: identifier 3-50 characters, description 10-5,000 characters, system prompt 20-10,000 characters.
- Manual creation ends with saving the file as agents/agent-name.md; all .md files in agents/ are auto-discovered.
- validate-agent.sh reads only the file named on its command line and prints its checks to the terminal. It has no network call and writes no file.

### Not established by this review

 - validate-agent.sh was not run against any agent file.
- references/agent-creation-system-prompt.md, references/system-prompt-design.md, references/triggering-examples.md and the two files in examples/ were not opened.
- Not executed: it was not installed or run.

The review applies to the material and revision named here. A newer upstream release can change its behavior.

03 / PUT IT TO WORK

## Add a skill to your workflow.

[Upstream setup instructions ↗](https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/plugins/plugin-dev/README.md)
 - The plugin-dev README prints `/plugin install plugin-dev@claude-code-marketplace`. That line names a marketplace called claude-code-marketplace, which is not this repository's marketplace (.claude-plugin/marketplace.json names it claude-plugins-official), and it installs the whole plugin-dev plugin (seven skills, the /plugin-dev:create-plugin command and the plugin's three agents), not this skill alone, so install.command is null.
- For development the same README prints `cc --plugin-dir /path/to/plugin-dev`.
- The repository README gives the general form `/plugin install {plugin-name}@claude-plugins-official`, with the plugin name as a placeholder.
- This review ran none of these lines.

### Before you start

 - Claude Code, where the /plugin lines run
- Bash, for scripts/validate-agent.sh (#!/bin/bash)

### Compatibility

Claude Code

### Guidance text plus a Bash validator that reads one agent file

 - No allowed-tools list in the SKILL.md, so tool use follows the host session's own permissions
- Creating an agent means writing agents/<agent-name>.md inside the plugin directory
- scripts/validate-agent.sh reads the one agent file passed to it and prints to the terminal; it makes no network call

### Cost model

Apache-2.0 per plugins/plugin-dev/LICENSE; the skill names no paid service or API key

04 / FOLLOW THE EVIDENCE

## The source trail.

Our notes are separate from the original resource. Check upstream before adopting a new version.

 - [agent-development SKILL.md ↗](https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/plugins/plugin-dev/skills/agent-development/SKILL.md) Checked 2026-10-21 https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/plugins/plugin-dev/skills/agent-development/SKILL.md Supports: upstreamDescription, summary, whySelected[0], whySelected[1], bestFor[0], bestFor[1], limitations[1], limitations[2], limitations[3], compatibility, review.findings[0], review.findings[1], review.findings[2], access.permissions[0], access.permissions[1]
- [agent-development scripts/validate-agent.sh ↗](https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/plugins/plugin-dev/skills/agent-development/scripts/validate-agent.sh) Checked 2026-10-21 https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/plugins/plugin-dev/skills/agent-development/scripts/validate-agent.sh Supports: whySelected[1], bestFor[2], limitations[2], review.findings[3], access.permissions[2], install.prerequisites[1]
- [plugin-dev README ↗](https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/plugins/plugin-dev/README.md) Checked 2026-10-21 https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/plugins/plugin-dev/README.md Supports: install.sourceUrl, install.steps[0], install.steps[1], install.prerequisites[0], limitations[4], limitations[5], compatibility
- [plugin-dev LICENSE ↗](https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/plugins/plugin-dev/LICENSE) Checked 2026-10-21 https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/plugins/plugin-dev/LICENSE Supports: license, limitations[4], access.cost
- [Marketplace file, name and plugin-dev entry ↗](https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/.claude-plugin/marketplace.json) Checked 2026-10-21 https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/.claude-plugin/marketplace.json Supports: limitations[5], install.steps[0]
- [Repository README ↗](https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/README.md) Checked 2026-10-21 https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/README.md Supports: install.steps[2]

KEEP COMPARING

## Other approaches to consider.

Related by category or shared topics. These are alternatives to inspect, not a measured quality order.

 [### Command Development for Claude Code ↗ Explains how to write Claude Code slash-command files: frontmatter fields, $ARGUMENTS and positional arguments, @file references, inline Bash and plugin command patterns.](/skills/anthropic-command-development/)[### Hook Development for Claude Code Plugins ↗ Explains how to write Claude Code plugin hooks, prompt-based and command-based, with their event list, input and output JSON and exit codes, and ships Bash utilities to validate, lint and test them.](/skills/anthropic-hook-development/)[### MCP Integration for Claude Code Plugins ↗ Explains how a Claude Code plugin bundles MCP servers, through .mcp.json or plugin.json, for stdio, SSE, HTTP and WebSocket servers, and how plugin commands pre-allow the resulting MCP tools.](/skills/anthropic-mcp-integration/)

 [AI Tools ↗](/tools/)[Skills ↗](/skills/)[Agents ↗](/agents/)[MCP Servers ↗](/mcp-servers/)[Workflows ↗](/workflows/)

## Continue your investigation

 - [Choose a toolkit](/tools/)
- [Inspect execution hosts](/agents/)
- [Check tool access](/mcp-servers/)
- [Check project dependencies](/check-repo/)
