Can I use different AI models for specific custom agents?

Yes, the 'model' field in the agent definition allows you to specify Haiku for fast lookups, Sonnet for standard development, or Opus for tasks requiring complex reasoning.

Where are custom agent definitions stored?

Custom agents are typically defined in the .claude/agents/ directory or within a plugin's agents/ directory using Markdown files with YAML frontmatter.

What is context forking in Claude Code custom agents?

Context forking creates an independent copy of the current conversation history for the sub-agent. This allows the agent to perform experiments or research without its changes or history polluting the main parent conversation.

What is the primary benefit of using custom agents?

Custom agents allow you to enforce the principle of least privilege and specialized focus, ensuring that an agent only has the tools and instructions necessary for a specific task like auditing or documentation.

How do I restrict a custom agent's capabilities?

You can restrict an agent by using the 'disallowedTools' field to blacklist specific tools (like Bash or Write) or by using 'allowed-tools' to create a strict whitelist of permitted actions.

Custom Agent Architect

Name: Custom Agent Architect
Author: laurigates

bylaurigates

•

開発者ツール

Defines and configures specialized AI agents with granular tool control and isolated execution contexts for tailored development workflows.

The Custom Agent Architect skill provides the expert framework needed to extend Claude Code by creating domain-specific agents beyond the default capabilities. It allows developers to implement strict security boundaries through tool whitelisting and blacklisting, select specific models (Haiku, Sonnet, Opus) for performance optimization, and utilize 'context forking' to run isolated research tasks that don't pollute the main conversation history. This skill is essential for teams looking to standardize complex workflows such as security auditing, automated documentation, and parallel code exploration while maintaining the principle of least privilege.

主な機能

01Strict capability restrictions via the disallowedTools blacklist

02Standardized YAML schema for plugin-based agent distribution

033 GitHub stars

04Context forking for isolated agent execution and research

05Granular tool whitelisting using the allowed-tools field

06Per-agent model selection for cost and performance optimization

ユースケース

01Spawning isolated research agents to explore library documentation or codebase patterns without affecting the primary workspace context

02Creating a read-only security auditor that analyzes code without the ability to modify files or execute bash commands

03Defining specialized documentation writers with access to file writing tools but restricted from executing system commands

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add laurigates/claude-plugins custom-agent-definitions

For use in Claude.ai and ChatGPT

Download Skill