Agent Skills

Intermediate

Reusable instruction packages that give agents specialized capabilities on demand.

Last updated: Sep 13, 2026

What are Agent Skills?

Agent Skills are folders of instructions, prompts, examples, and resources that an LLM can load when relevant to perform specialized tasks consistently. Instead of putting everything in the system prompt, skills let you modularize expertise.

"Skills are like apps for your agent—install them once, use them whenever needed."

Just as you install apps on your phone for specific functions, skills give agents specialized capabilities without bloating every conversation.

How Agent Skills Work

1

Skill Discovery

The host initially exposes names and descriptions. The model can choose a relevant skill, or the user invokes one explicitly. Keyword rules are an optional host implementation, not part of the skill format.

2

Skill Loading

The relevant skill's instructions, examples, and context are loaded into the agent's working memory. This adds specialized knowledge without polluting the base system prompt.

3

Skill Execution

The agent follows the skill's instructions to complete the task, using any provided templates, checklists, or scripts. Results are returned to the user.

Interactive Skill Demo

Inspect a local selection rule, standard metadata and progressive loading.

Select a skill, then load its instructions

This local teaching matcher uses the displayed whole-word rules in English and German. It is not an Agent Skills standard or a model-confidence estimate. Real hosts commonly let the model select from skill descriptions.

No local rule matched. Choose a skill explicitly or revise the request.

Initially available: name and description

---
name: code-review
description: "Performs thorough code reviews following team standards"
---
📁

Skill Structure

Anatomy of a skill folder

my-skill/
SKILL.md # Metadata and main instructions
instructions.md # Detailed guidance for the task
examples/ # Sample inputs and outputs
templates/ # Reusable output formats
scripts/ # Helper scripts if needed

SKILL.md is the entry point. Standard required frontmatter fields are name and description; the Markdown body contains the instructions. Load supporting files only when needed.

Types of Agent Skills

1

Domain Skills

Domain-specific references and procedures. They can support specialized work but do not replace domain expertise or evidence checks.

2

Workflow Skills

Multi-step processes with defined stages—code review workflows, content publishing pipelines, incident response procedures.

3

Format Skills

Consistent output formatting—API documentation, changelog entries, meeting summaries. Ensure outputs match your standards.

4

Integration Skills

Instructions for working with specific tools or services—GitHub workflows, Jira ticket creation, Slack notifications.

Skills vs. Tools

Tools

  • •Execute actions (read files, call APIs, run code)
  • •Defined by function signatures and schemas
  • •The "hands" of the agent

Skills

  • •Provide knowledge and methodology (how to approach tasks)
  • •Defined by instructions and examples
  • •The "expertise" of the agent

Skills and tools work together: a code review skill tells the agent what to look for, while tools let it read the code and leave comments.

Benefits of Skills

Specialization Without Bloat

Keep the base system prompt lean. Load specialized knowledge only when needed, preserving context window for the actual task.

Consistency

Make the intended procedure reviewable and repeatable. Verify whether the agent follows it; written instructions alone do not guarantee consistent execution.

Shareability

Skills are just folders—share them across projects, teams, or publicly. Build once, use everywhere.

Iteration

Improve skills independently of the agent. Update the code review skill without touching the rest of your agent setup.

Example: Code Review Skill

SKILL.md
---
name: code-reviewer
description: Performs thorough code reviews following team standards
---

# Example: Code Review Skill

When reviewing code, analyze for correctness, performance, security, and maintainability.

## Test
- [ ] Check for common security vulnerabilities
- [ ] Verify error handling is comprehensive
- [ ] Look for performance anti-patterns
- [ ] Ensure code follows style guidelines

Best Practices

1

Keep Skills Focused

One skill, one purpose. If a skill is doing too many things, split it up. Focused skills are easier to maintain and compose.

2

Include Examples

Show, don't just tell. Include input/output examples that demonstrate exactly what good execution looks like.

3

Version Your Skills

Track changes to skills over time. When behavior changes unexpectedly, you can trace it back to a skill update.

Key Takeaways

  • 1Skills are reusable instruction packages that give agents specialized expertise on demand
  • 2Unlike tools (which execute actions), skills provide knowledge and methodology
  • 3Well-designed skills improve consistency and reduce system prompt bloat
  • 4The skill framework enables modular agent development—build once, share everywhere

Primary sources