Claude Code Plugin Agent Development
A practical guide to designing, writing, validating, and testing autonomous agents for Claude Code plugins.
The skill explicitly recommends restricting tools according to least privilege and explains that omitting tools grants access to all tools. It does not define user confirmation, sensitive-data handling, external-effect disclosure, rollback, or recovery procedures, so points are deducted. Repository provenance and Anthropic ownership support attribution only, not safety claims.
The skill supplies structural rules, a validator script, trigger-testing steps, and edge-case guidance, making the happy path understandable. However, this is a static review with no executed reproduction or committed tests covering key paths, and claims such as “exact” and “production-ready” are not independently verifiable, so the score is capped and reduced.
The audience, scenarios, trigger phrases, input structure, and expected agent-file format are reasonably clear. Non-fit and negative-trigger boundaries are incomplete, Chinese-language guidance is absent, and reachability of Claude Code or related documentation from mainland-China networks is not addressed, so points are deducted.
The material is layered across overview, fields, design, examples, validation, and references, and includes a version, examples, and script locations. It lacks installation/dependency notes, known limitations, a changelog, explicit maintenance ownership, and an update path. License information is available only from repository context, not the skill itself, so points are deducted.
The skill provides copyable frontmatter, system-prompt patterns, triggering examples, and an implementation workflow that addresses the core agent-creation task. But claims such as “battle-tested” and “production-ready” lack static execution evidence, and generated agents still require user review and real testing, so points are deducted.
The source is auditable and includes rules, scripts, and examples, providing limited traceability. It lacks sufficient third-party execution evidence, key-path test coverage, and corroboration; no scripts were executed in this review, so only a low score is justified.
- If tools is omitted, the agent receives all tools by default; explicitly restrict permissions and confirm before write, Bash, or other externally effectful actions.
- The material recommends trigger testing and validation but provides no executed results here; do not treat “production-ready” or “battle-tested” as verified facts.
- Referenced references/, examples/, and scripts/ paths depend on the repository layout; confirm that files exist, scripts are executable, and rules match the current Claude Code version.
- No Chinese-language trigger examples or mainland-China network reachability guidance is provided.
What it does & when to use it
This skill helps Claude Code plugin developers build agents for complex, multi-step work. It covers Markdown file structure, YAML frontmatter, triggering conditions, and system prompt design. It also explains model, color, and tool configuration, plus validation and testing practices. It is a fit for developers who need to add autonomous agents to a Claude Code plugin and want explicit guidance on structure and behavior.
Guides users in creating Markdown agent files under a plugin's agents/ directory; writing the name, description, model, color, and optional tools fields; designing system prompts with responsibilities, processes, output formats, and edge cases; validating identifiers, descriptions, and prompts against stated rules; using the documented validate-agent.sh and test-agent-trigger.sh utilities for validation and trigger testing; and documenting the agent in the plugin README.
- A plugin developer adding a code-review, test-generation, or security-analysis agent.
- A plugin author defining autonomous workflows for complex, multi-step tasks.
- A developer improving agent triggering through multiple concrete description examples.
- A team choosing models, colors, and least-privilege tool sets for several agents.
- A maintainer validating agent structure, trigger behavior, and prompt edge cases before release.
Pros & cons
- Covers the workflow from agent structure and prompt design through validation and testing.
- Includes reusable frontmatter, system prompt, and minimal-agent templates.
- Defines formats and recommendations for identifiers, descriptions, models, colors, and tools.
- Supports both AI-assisted and manual agent creation.
- It targets Claude Code plugin concepts such as the agents directory, namespacing, and automatic discovery, so other clients require adaptation.
- The source does not document a standalone installation process for the skill.
- The source provides no actual test results or platform compatibility matrix.
- The repository metadata lists the license as unknown.
How to install
The README documents installation of Claude Code, but it does not provide a separate installation command for this skill. Install Claude Code using a documented method, such as curl -fsSL https://claude.ai/install.sh | bash on macOS/Linux, then enter the project directory and run claude. The skill is located at plugins/plugin-dev/skills/agent-development/SKILL.md; a standalone procedure for copying or registering it is not documented in the source.
How to use
In Claude Code, make a request such as Create an agent for reviewing API security issues in this plugin, How should I write the agent frontmatter and triggering examples?, or Validate this agent's system prompt and tool permissions. The documented trigger topics include creating or adding an agent, agent frontmatter, when-to-use descriptions, examples, tools, colors, and autonomous agents.