What is the difference between project-level and user-level agents?

Project-level agents are stored in the .claude/agents/ directory and can be version-controlled for team use, while user-level agents in ~/.claude/agents/ are personal and available across all your local projects.

Can I prevent a custom agent from making changes to my files?

Yes, you can configure 'disallowedTools' in the agent's YAML frontmatter to explicitly block tools like Write or Edit, or use 'permissionMode: plan' for read-only exploration.

How does Claude Code decide when to use a custom agent?

Claude uses the agent's description field to determine delegation. Providing a clear, keyword-rich description helps Claude understand exactly when the agent's specialized expertise is required.

What are subagent lifecycle hooks?

Hooks are scripts that run at specific stages, such as PreToolUse or PostToolUse. They are used for validation, setup tasks, or cleanup, ensuring the agent operates within defined safety boundaries.

Claude Agent Architect

Name: Claude Agent Architect
Author: adbutterfield

byadbutterfield

0•

生产力与工作流

Streamlines the creation and configuration of specialized Claude Code subagents with precise tool restrictions and custom lifecycle behaviors.

The Claude Agent Architect skill provides a comprehensive framework for designing and deploying custom Claude Code subagents. It guides developers through the entire lifecycle of agent creation—from selecting the appropriate scope (user vs. project level) to configuring granular tool permissions and implementing advanced lifecycle hooks. By using this skill, teams can create domain-specific experts like security auditors or database readers that follow strict safety protocols and consistent implementation patterns, ultimately enhancing the autonomy and reliability of AI-driven development workflows.

主要功能

01Granular tool access management via allowlists and denylists

02Optimization strategies for agent descriptions to improve auto-delegation

03Guidance on implementing lifecycle hooks for automated validation

040 GitHub stars

05Step-by-step design process for project-specific or global subagents

06Pre-configured patterns for read-only reviewers and domain specialists

使用场景

01Building a specialized database agent with pre-tool SQL validation scripts

02Setting up personal productivity agents for consistent code review patterns

03Creating a team-wide security auditor agent with restricted write access

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add adbutterfield/coding-agent-plugins creating-agents

For use in Claude.ai and ChatGPT

Download Skill