agent-skill-use_skill

Overview

This skill helps you create and manage AI agent skills following proven patterns and operational best practices. It provides a three-layer architecture and progressive disclosure for efficient context management. Use it to design subagents, wrap external tools, and keep token usage and load times under control.

How this skill works

The skill defines a lightweight index-first loading strategy: an index always loaded, metadata loaded on trigger, and full content loaded only on activation. It prescribes a three-layer architecture (connector, skill, subagent) that separates external connections, task logic, and isolated execution contexts. It also supplies checklists and spawn rules for when to use subagents versus in-skill workflows.

When to use it

• Creating a new agent skill that needs clear purpose, inputs, and triggers
• Optimizing context size and token usage across many skills
• Designing multi-agent systems with coordinator and worker roles
• Implementing progressive disclosure to reduce load time and costs
• Deciding whether tasks require parallel, isolated subagents

Best practices

• Adopt the three-layer architecture: connector (MCP), skills, subagents for isolation
• Use progressive disclosure: index → metadata → full content to save tokens
• Wrap external tools behind skill interfaces; load full tool definitions on demand
• Define clear spawn conditions for subagents: parallelism, isolation, long-running work
• Name skills in kebab-case, lowercase, and avoid forbidden trademark words
• Keep base context under target token budgets and measure load times

Example use cases

• Automating a CI workflow: skill handles orchestration, subagents run parallel test suites
• Creating a code-quality skill: wraps linters and PR reviewers and loads tools on activation
• Building an integration skill: minimal index loaded, full API bindings loaded only when used
• Designing a multi-agent review pipeline: coordinator issues tasks, review team agents process items
• Progressively disclosing documentation for large content generation tasks to reduce token costs