Guides
What are AI agent skills?
AI agent skills (or AI skills) are reusable instruction packages that give an agent a defined workflow, domain knowledge, and supporting resources for a task.
An AI agent skill is a reusable unit of operating knowledge. It tells the agent what the task is, when the workflow applies, which steps matter, and which resources support the result. Instead of restating the same process in every conversation, the workflow lives in a skill and can be loaded when relevant.
Skills are usually organized as a folder built around a file named SKILL.md.
That file explains what the skill is for, when to use it, and how the workflow should run.
The folder can also include scripts, templates, reference notes, and other resources that
support the task.
my-skill/
├── SKILL.md
├── scripts/
├── references/
└── assets/A short history of the skills standard
The current Agent Skills standard was introduced by Anthropic in October 2025 as a way to package instructions, scripts, and resources for agents. In December 2025, that work was published as an open cross-platform standard, which helped skills spread more easily across different AI products and agent environments.
Why skills matter
Skills help generative models and LLM-based agents produce more reliable outcomes. Teams still need repeatable workflows, shared standards, and a way to preserve knowledge that should not live inside one prompt or one person’s memory. Skills solve that by turning good procedures into reusable building blocks.
A well-written skill can teach an agent how to review pull requests, prepare a report, analyze a spreadsheet, onboard a customer, write in a house style, or follow an internal operating process. The agent still reasons on its own, but the skill gives that reasoning a stronger structure.
- Reusable context for recurring tasks
- Repeatable workflows across conversations and teammates
- Less prompt repetition for common jobs
- More consistent outputs for domain-specific work
AI products that can use skills
Popular AI products that can use skills or skill-like reusable agent setup include:
Each product uses its own interface and naming, but all of them can work with reusable instructions, reusable context, and reusable workflow logic.
How skills work
Skills usually work through progressive loading. The agent does not need to read every workflow in full at startup. It can keep a short description for discovery, then load the full instructions only when the task actually needs them.
- Discovery: the agent sees the skill name and description.
- Activation: when the task matches, it reads the full
SKILL.md. - Execution: it opens referenced files or runs bundled scripts if the workflow needs them.
This keeps context usage efficient. An agent can have many skills available without paying the full context cost of every workflow upfront.
Some skills also include executable scripts. That is useful when part of the workflow is better handled by code than by language generation alone. Data extraction, file transformation, deterministic validation, and structured export are common examples.
What goes inside SKILL.md
A good SKILL.md explains when to use the skill and how to execute the
workflow. It usually includes a short description, the task steps, tool instructions,
quality checks, and references to other files in the skill folder.
Many skills also include lightweight metadata at the top of the file. That helps the agent understand the purpose of the skill before it loads the rest. The body of the file then carries the real workflow: what to do, in what order, with which tools, and what a good result looks like.
The best skills are specific. They do not try to teach everything. They focus on one job or one family of jobs and make that workflow clear.
This format is easy to read, easy to edit, and easy to store in version control. Teams can review a skill like any other working document, improve it over time, and share it across products that support skills.
Pre-built and custom skills
Some skills come from the AI product itself or from software vendors. These pre-built skills give the agent product-specific workflows without requiring the user to create them from scratch.
Custom skills come from teams, communities, or individual builders. They package internal standards, reusable research methods, coding workflows, writing systems, and domain expertise that matter in a specific environment.
Together, pre-built and custom skills make the ecosystem useful. Official skills bring product accuracy. Custom skills bring organizational knowledge and specialized methods.
Two broad categories of skills
In practice, skills tend to fall into two broad groups.
1. Domain knowledge skills
These skills teach an agent how to work well in a subject area. An SEO skill is a good example. It can define how to review metadata, structure content, evaluate internal links, or draft search-focused copy. The same idea applies to legal review, security analysis, product management, customer support, or content editing.
Domain knowledge skills are useful because they encode methodology. They help the agent follow a discipline instead of improvising from general knowledge alone.
Some domain skills focus on work style rather than business software. A communication
skill such as caveman teaches the agent to answer in a compressed, highly
efficient style while keeping technical meaning intact. Research-oriented skills can go in
the other direction. An autoresearch skill teaches the agent to run a more
structured research loop, and Karpathy-inspired research workflows push that idea further
toward iterative investigation, experimentation, and synthesis.
2. Official vendor and software skills
These skills teach an agent how to work with a specific product or platform. A Mixpanel skill can show the agent how to work with analytics questions, dashboards, tracking implementation, or reporting workflows. A Stripe skill can guide integration work. A Figma skill can help with design workflows. A Supabase skill can encode product-specific best practices.
Official vendor skills come from the companies that own the tools. People trust them because they cover the products directly and reflect the workflows those products support.
A Canva skill set is a clear example. It can teach an AI to turn a brief into a branded presentation, resize one design into the standard formats for social platforms, and translate an existing design into another language while preserving the original layout as much as possible. That is product-specific operating knowledge, not general design advice.
You can browse many of these vendor skill sets in the Skillscout official skills directory.
Where skills are used today
Skills are now used across coding tools, chat products, APIs, internal copilots, and agent frameworks. Common use cases include document creation, spreadsheet work, coding tasks, support workflows, analytics operations, research methods, internal reporting, and company-specific procedures.
People often discover them through GitHub skills searches, agent skills GitHub lists, official vendor directories, and collections such as skills.sh. That discovery layer is a big part of how AI agent skills spread in practice.
Skills vs prompts
A prompt asks for work in the current conversation. A skill packages a reusable way of doing that work. Prompts are one-off instructions. Skills are reusable operating logic.
You can still trigger a skill with a prompt, of course. The difference is that the prompt starts the task, while the skill shapes how the task gets done.
Security matters
Skills often include instructions, scripts, references, and file access. That makes them useful, and it also makes review important before installation.
A random skill from GitHub or a registry can ask for network access, read sensitive files, or run commands that deserve closer review. Teams should treat skills as working code and check the source before using them.
A skill such as skill-vetter helps with that review. It checks the source,
scans for red flags, and looks at the permissions a skill would need before installation.
How Skillscout helps
As skills become more common, discovery becomes the harder problem. A useful workflow may already exist, but it still needs to be found at the right moment. That is especially true for vendor skills, because the right skill often depends on the exact product page a person is visiting.
The Skillscout Chrome extension helps by reading the current site, matching that context to relevant skill catalogs, and surfacing the skills that fit the page. On a product site, it can point you toward official product skills. On a broader workflow page, it can surface domain knowledge skills such as SEO, writing, security, or research guidance.
It is handy for quick search. The relevant skill appears closer to the work itself.