# Build Focused Claude Code Subagents

Building a subagent can require careful decisions about triggers, tools, prompts, and output. This skill provides structured guidance, examples, and validation steps for dependable Claude Code agents.

## Install

```bash
npx skillstore add anthropics/agent-development
```

## Metadata

- Status: approved
- Slug: anthropics-agent-development
- Version: 0.1.0
- Author version: 0.1.0
- Skillstore revision: r2
- Version status: valid
- Tree hash: 497a0b25b113dffd131c29c9e9f69fe0d76def5e01f30bf01f41844ceadb464c
- Author: anthropics
- GitHub username: anthropics
- License: MIT
- Repository: https://github.com/anthropics/claude-plugins-official/tree/main/plugins/plugin-dev/skills/agent-development
- Ref: 80999bf530a7874d7bedf8ce202001ecb4c4f5e0
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: allowed
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: filesystem, external\_commands
- Quality score: 84
- Quality tier: gold
- Public page: https://skillstore.pages.dev/skills/anthropics-agent-development
- Manifest: https://skillstore.pages.dev/api/skills/anthropics-agent-development/manifest

## Capabilities

- Explains the required frontmatter fields for Claude Code subagent files.
- Helps write focused system prompts with responsibilities and output expectations.
- Recommends triggering descriptions and examples for appropriate agent selection.
- Guides least-privilege tool selection for analysis, generation, and testing tasks.
- Provides agent patterns for code review, testing, documentation, and security analysis.
- Includes a local shell validator for agent file structure and basic content checks.

## Use Cases

- Create a review agent: Define a focused code-review subagent with clear triggers, scoped tools, and a consistent findings format.
- Standardize team agents: Create repeatable frontmatter and prompt conventions for a plugin or repository agent collection.
- Validate an agent file: Check a draft agent file for required fields, frontmatter boundaries, and basic prompt expectations.

## Prompt Templates

### Beginner: Create a test agent

```
Help me design a test-generation subagent for this repository. Propose its name, triggering description, tool scope, and system prompt.
```

### Intermediate: Improve triggers

```
Review this subagent description and examples. Rewrite them so Claude Code selects the agent only for API documentation tasks.
```

### Advanced: Apply least privilege

```
Review this agent design and recommend the smallest tool set needed to inspect code and report security issues without modifying files.
```

### Advanced: Design an agent suite

```
Design a small subagent suite for a TypeScript service. Define boundaries, triggers, tool scopes, handoff rules, and validation checks for each agent.
```

## Limitations

- It does not automatically create, install, or publish an agent.
- The validator checks structure and basic conventions, not agent quality or safety in every context.
- Tool access recommendations require adaptation to the repository and organization policy.
- Examples may need updates when Claude Code agent formats change.

## Best Practices

- Give each agent one clear responsibility and explicit selection examples.
- Grant only the tools required for the agent task and repository workflow.
- Validate frontmatter and test the agent with representative user requests.

## Anti Patterns

- Do not use vague names or descriptions that overlap with unrelated agents.
- Do not grant full tool access when read-only or limited write access is sufficient.
- Do not rely on a validator as the only review of agent behavior or permissions.

## Security Audit

- Audited at: 2026-07-18T10:06:09.957\+00:00
- Summary: All 82 static findings are false positives after contextual review. Most matches are Markdown backticks and instructional examples; the local validator uses quoted standard shell utilities without evaluating input as code. No prompt-injection attempt, data exfiltration intent, or malicious business logic was found.

## Stats

- Views: 523
- Downloads: 13
- Favorites: 0
- Popularity score: 0