Work with us
Have a product or AI idea?
Tell us what you want to build and we'll suggest a practical first step.
Every team that uses AI assistants hits the same wall. The assistant writes decent code or decent copy, but it does not know how your team does things. It names files the wrong way, skips your release checklist, or writes in a tone your brand would never use. So people paste the same long instructions into chat, again and again, and each person's version drifts a little.
Agent Skills are a fix for that problem, and in the last four months they went from one company's feature to a format that most major AI coding tools can read. This post explains what a skill is, who supports the format as of February 2026, and how a business can use skills to make AI assistants follow its own processes.
What a skill is
Anthropic introduced Agent Skills (opens in a new tab) on October 16, 2025. A skill is a folder that contains instructions, scripts and resources an agent can load when it needs to do a specific task. Anthropic's own examples include working with Excel files and following an organization's brand guidelines.
The heart of each folder is a file called SKILL.md. It starts with a short block of YAML frontmatter and continues
with plain Markdown instructions:
---
name: release-checklist
description: Steps to prepare and verify a production release. Use when the user asks to cut, tag or ship a release.
---
1. Confirm the changelog has an entry for every merged pull request.
2. Run the full test suite and the smoke tests against staging.
3. ...
The Agent Skills specification (opens in a new tab) requires only two fields: a name (lowercase
letters, numbers and hyphens, matching the folder name) and a description that says what the skill does and when to
use it. Next to SKILL.md, a folder can hold a scripts/ directory for code the agent can run, a references/
directory for longer documentation, and an assets/ directory for templates and data files.
The design choice that makes skills practical is called progressive disclosure. At startup, the agent reads only the name and description of each installed skill. It loads the full instructions when a request matches, and it opens the reference files or runs the scripts only when the task calls for them. You can install dozens of skills without filling the model's context window with text it does not need for the current task.
From one vendor's feature to an open standard
On December 18, 2025, Anthropic published Agent Skills as an open standard (opens in a new tab) at agentskills.io, arguing that skills, like the Model Context Protocol, should be portable across tools. The same announcement added two things for companies on Claude's Team and Enterprise plans. Admins can now provision skills centrally, and those skills are switched on for every user by default. There is also a directory of partner-built skills from companies including Notion, Canva, Figma and Atlassian.
Other tools moved quickly:
- GitHub Copilot. GitHub announced Agent Skills support (opens in a new tab)
the same day, for the Copilot coding agent, Copilot CLI and agent mode in VS Code Insiders, with stable VS Code
support scheduled for early January. Copilot also picks up skills that a repository already keeps in
.claude/skills. - OpenAI Codex. OpenAI added skills to Codex on December 19, 2025. The Codex documentation (opens in a new tab)
says its skills build on the open standard. In the CLI and IDE extension you can call one directly with
$skill-name, or let Codex choose a skill from your prompt. - Vercel and skills.sh. On January 20, 2026, Vercel released a
skillsCLI and the skills.sh directory (opens in a new tab). You install a skill package withnpx skills addfollowed by the package name, and skills.sh lists packages by category and popularity, with install counts. At launch Vercel said the CLI had been used with 17 agents, including Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Goose, OpenCode and Windsurf.
For a business, the practical meaning is portability. A skill your team writes once can be read by the AI tools your developers already use, even if different people prefer different tools. Your process knowledge stays in plain Markdown files that you own and can keep in version control.
What a business can put in a skill
A good candidate for a skill is any task your team does repeatedly, where there is a right way to do it and people keep forgetting part of it. Some examples by team:
| Team | Skill idea | What goes in the folder |
|---|---|---|
| Engineering | Add an API endpoint the way our codebase does | Folder layout, naming rules, test checklist, an example handler |
| Engineering | Prepare a release | Checklist, versioning rules, a script that drafts the changelog |
| Marketing | Write in our brand voice | Voice rules, banned phrases, three approved sample posts |
| Marketing | Publish a blog post | Front matter template, SEO checklist, image size rules |
| Operations | Draft a client proposal | Proposal template, pricing rules, approval steps |
| Support | Triage a bug report | Severity definitions, required fields, escalation contacts |
Engineering skills tend to pay off first because coding agents already run inside the repository, where skills can live next to the code. Marketing and operations skills work the same way in chat tools that support them, such as the Claude apps.
How to write skills that agents actually follow
The format is simple, so most of the work is in deciding what to write. These habits help.
Write the description for the agent. The description is the only part the agent sees before it decides whether to load the skill. Say what the skill does and list the words a person would use when they need it: "Use when the user asks to cut, tag or ship a release." A vague description such as "helps with releases" means the skill never triggers.
Keep one task per skill. A skill called engineering-standards that covers testing, naming, deployment and code
review will load for everything and be precise about nothing. Split it into a few smaller skills with clear names.
Write steps, then examples. Agents follow numbered steps well. After the steps, show one or two real inputs and the output you expect. A short example of a correct commit message does more than a paragraph describing one.
Move long material into reference files. Keep SKILL.md short and link to files in references/ for details
such as a full style guide or an API schema. The agent reads them only when it needs them.
Use scripts for anything that must be exact. If a step is mechanical, such as checking that an image is the right
size or that a changelog entry exists, write a small script in scripts/ and tell the agent to run it. A script
gives the same answer every time, which instructions alone cannot promise.
Test with real requests. Ask the assistant to do the task the way a colleague would ask, without naming the skill. Check whether it loaded the skill and followed the steps. If it did not, the description usually needs work.
Rolling skills out across a company
Treat skills like code, because they change how your tools behave.
- Start with three skills. Pick the tasks where mistakes cost the most or where people ask the same questions most often. Write those, test them for two weeks, then expand.
- Keep them in version control. Repository skills live in the repo and go through pull requests like any other change. For non-engineering teams, a shared repository of skills gives you history and review.
- Name an owner for each skill. When the release process or the brand voice changes, someone has to update the skill, or the AI will keep following the old rules.
- Review third-party skills before installing them. A skill can include scripts that your agent will run on your machine or your servers. Read the files before you add a skill from a public directory, the same way you would review a new dependency.
- Use central provisioning where your tools offer it. On Claude Team and Enterprise plans, admins can push a skill to everyone at once, which beats asking each person to install it.
Start with the process you repeat most
The Agent Skills format is small and young, and some details will keep changing as more tools adopt it. The core is
stable enough to build on today: a folder, a SKILL.md with a clear description, and optional scripts and references.
Pick one process your team explains over and over, write it down as a skill, and test it with the AI tools your team
already uses.
Want help turning your team's processes into skills and AI workflows? Book a free 30-min call (opens in a new tab) and we will map out where skills fit in your stack.
Related service: AI Development