What Are Skills

A skill is a reusable bundle of procedures and know-how an AI agent needs to carry out a specific kind of task. Instead of re-explaining everything from scratch in every prompt, a skill is prepared in advance as a set of instructions and loaded only when it’s needed.

Why Skills Matter

Rewriting a long prompt every time you make a similar request is inefficient. Knowledge you reuse repeatedly — “what to check in a PR review,” “how articles should be structured,” “the pre-deploy checklist” — is easier to manage as a standalone set of instructions than as something rewritten inline in every prompt.

Skills vs. Prompts vs. Tools

Scroll horizontally→

Aspect	Prompt	Skill	Tool
Role	Communicates the current request	Bundles procedures and judgment criteria for a specific task, reusably	Executes an external function
Frequency	Written for every request	Written once, invoked repeatedly	Executed every time it’s called
Contents	Goal, constraints, output format	Steps, checklists, common failure modes and fixes	The executable function itself

This table scrolls horizontally. Keyboard users can focus the table and use the left and right arrow keys.

A skill is “knowledge of what should be done.” A tool is “the means to actually do it.” A skill’s procedure often combines multiple tools along the way.

Elements of a Good Skill Definition

Scroll horizontally→

Element	Description
Name and scope	What kind of request should trigger this skill
Steps or checklist	Exactly what to do, in what order
Inputs or prerequisites	What needs to be in place before using this skill
Common failure modes	Past mistakes and how to avoid them

This table scrolls horizontally. Keyboard users can focus the table and use the left and right arrow keys.

Skill name: PR review checklist
Scope: When asked to review a pull request
Steps:
1. Confirm the purpose and scope of the change
2. Check whether tests were added or updated
3. Check whether the change deviates from existing design patterns
4. Check for security risks (input validation, permissions)
Common failure mode: Focusing too much on style and missing design-level issues

Load Only What’s Needed

Keeping every skill permanently in the AI’s context turns irrelevant procedures into noise that interferes with judgment. Loading only the skills relevant to the current task follows the same principle covered in context engineering: supply only the information that’s needed.

Pages in This Section

Scroll horizontally→

Page	Content
Loading Only When Needed	Progressive disclosure and how to write the summary
Skills Versus Prompts Versus MCP	How the four mechanisms differ in role
Writing a Good Skill	Guidance on granularity, steps, and completion conditions

This table scrolls horizontally. Keyboard users can focus the table and use the left and right arrow keys.

Summary

• A skill is a reusable bundle of procedures and know-how organized around a specific task
• If a prompt is “the current request,” a skill is “a reusable playbook”
• If a tool is “the means to execute,” a skill is “the knowledge of what and how to execute”
• Loading skills only when relevant avoids adding noise to context