Data compiled: 2026-09-21
On this page, “AI skill” means Agent Skill: a reusable capability package for AI agents.
One-line definition #
An AI skill is a reusable, version-controlled package of instructions and resources that hands an AI agent the knowledge, workflow, checklist, and optional program tools for a specific task, to load and run as needed.
It is not a new AI model, nor a one-off prompt; it is more like an AI agent’s “SOP + professional toolkit”.
What does a skill usually contain? #
The core file is SKILL.md:
- YAML frontmatter: at least
nameanddescription - Main instructions: when to use, input requirements, work steps, output format, pre-completion checks
A skill can also include:
scripts/: executable helper scriptsreferences/: detailed docs, API notes, specsassets/: templates, images, examples, fixed resources
How does an AI agent use a skill? #
- Discovery: on startup, it reads only each skill’s name and description, keeping context light.
- Activation: when a user’s task matches the description, the agent loads the full
SKILL.md. - Execution: the agent follows the instructions, reading references and using scripts or assets as needed.
This progressive disclosure lets an agent install many skills but only consume full context when relevant.
Why is it useful? #
| Benefit | Description |
|---|---|
| Consistency | Locks multi-step SOPs in place, reducing variance between repeated interpretations |
| Specialization | Supplements company, team, or personal domain knowledge |
| Reusable | The same skill works across many tasks, projects, or compatible agents |
| Maintainable | Version, review, test, and update it like code |
| Auditable | Clearly records what the agent should do, output, and check before finishing |
| Context savings | Only loads the full workflow when needed |
How it differs from other AI components #
| Component | Primary role |
|---|---|
| Model | The base AI that produces reasoning and text |
| Skill | A reusable SOP that tells the agent how to do a class of task |
| MCP / tool | Connects external data, services, or operational capability |
| Agent | Combines model, skills, tools, and context to complete tasks |
| Prompt | A one-off request; a skill is a workflow that can be saved, reused, and triggered on demand |
A skill can call or pair with tools, MCP, and scripts; it provides the “how to do it” workflow, while a tool provides the “what it can do” capability.
Your examples #
- Taste: front-end design taste, design system mapping, image-to-code, redesign, and other workflows.
- humanize: finds AI writing patterns, rewrites them into a natural tone, then runs a final anti-AI audit.
- superpowers: brainstorming, systematic debugging, TDD, writing plans, code review, verification, and other software development workflows.
Minimal skill example #
---
name: meeting-summary
description: Turn meeting notes into a concise decision log and action list.
---
# Meeting summary
1. Extract decisions, owners, deadlines, and unresolved questions.
2. Preserve uncertainty instead of inventing missing facts.
3. Return a decision table followed by action items.
4. Check that every action has an owner or is marked unassigned.
A real skill can add scripts, references, and assets on top.
Checklist for a good skill #
- Define a clear task scope and trigger description
- Write repeatable workflows as explicit, executable steps
- Specify input, output format, and pre-completion verification
- Put long docs and code into references/scripts instead of writing
SKILL.mdas a giant encyclopedia - Set least privilege for scripts, network access, file changes, and secrets
- Test with real cases that the skill triggers and completes correctly
- Bump the version and keep sources for significant changes
Limits and caveats #
A skill is instructions and resources, not a security guarantee, and it does not automatically make a model smarter. You still need to check the output, code, external operations, and permissions. Skills that include shell commands, network requests, or file writes should especially be reviewed before enabling.
References #
| Source | Highlights |
|---|---|
| Agent Skills overview | Open format, SKILL.md structure, progressive disclosure, cross-product reuse |
| OpenAI API Skills | Versioned bundle, SKILL.md manifest, upload and usage rules |
| OpenAI Codex Build Skills | Codex skill directory, how to build, name/description requirements |
| Claude Code Skills | Building skills, load locations, triggers, and supporting files |