AI Skills: What They Are, and Why You Probably Want to Use Them
For hands-on training in designing structured AI agent workflows with built-in checks, human oversight, and clear documentation, join Professor Crabtree for AI Agent-Driven Research Workflows.
If you’ve read any AI news in the last several months, you’ve likely read about something called skills. Or perhaps you’ve heard of a colleague using them. In this post, I’ll talk about what they are, how they can help, how you can find them, how you can make them, and how you can use them in your daily workflows.
In October 2025, Anthropic introduced ‘Agent Skills’, a way to extend Claude’s capabilities by adding customized abilities and knowledge. A couple of months later, Anthropic launched the approach as an open standard, and a growing number of AI agents, including OpenAI’s Codex and Google’s Gemini CLI, now support skills. In AI-focused corners of the internet, skills have become ubiquitous. People create them, often share them on GitHub, remix them to suit their own purposes, and boast that their use has radically changed their work and even their lives.
So, what are skills? While one might think that anything so radically effective might be technically complicated, skills are fairly easy to understand and create, which likely explains their tremendous growth as part of users’ daily toolkits. In its most basic form, a skill is a folder containing a SKILL.md file that begins with basic YAML at the top, followed by Markdown-formatted instructions below. The figure below contains a simple example.
If YAML or Markdown sound too intimidating, don’t worry. You can build skills just by asking your favorite AI agent to create one for you based on the description you provide. Agents can even help by suggesting potentially useful skills given their working history with you. One of the nice features of skills, though, is that they can also be created or edited by hand in a basic text editor.
In more complicated forms, a skill can include programming scripts, templates, reference documents, or other supporting materials alongside its SKILL.md file. Once you create or download a skill, you can place it in ~/.claude/skills/ for Claude or ~/.agents/skills/ for Codex or Gemini CLI, and the agent will discover it automatically. You can also invoke or manage skills manually using the product’s skill interface or slash commands.
While this technical scaffolding is fairly simple, skills can be used to powerful effect. One way to use them is to provide additional knowledge beyond the scope of the training data about your tastes. You can create skills that capture big, important things, such as how to run a multi-model workflow with a frontier model as an orchestrator, your preferences for analyzing and presenting the results of conjoint experiments, your ideas about how a replication package should be produced, or how your research team likes to prepare and organize internal reports.
You can also create skills that capture your predilections for smaller things, such as how you like to write emails, format table captions, or structure your acknowledgments section. You can, and probably should, create at least one “adversarial” skill at some point that audits your code and empirical choices. Scott Cunningham’s influential ‘Reviewer 2’ skill serves as an excellent base here.
Another important use case that sometimes overlaps with the above is using skills to help agents consistently perform a task. One problem I’ve encountered in the past is that Claude Code and Codex would sometimes generate code that creates figures in which text elements run outside the plotted area. This is a minor issue in some ways but an important one to deal with before circulating any figures in drafts or in papers to be submitted. So, I created a skill to help ensure that every text element in a plot fits within the figure (no titles or legends trailing off into the ether). The figure above displays that skill.
While skills are easy to create, a good place to start for the novice is probably to download some skills from one of the excellent GitHub repositories or lists. (You don’t need a GitHub account to download or use them.) Some great stops for social scientists include Scott Cunningham’s repo, Steven Denney’s repo, and Christopher Kenny’s curated list of valuable skills.
In my upcoming seminar, AI Agent-Driven Research Workflows, I’ll show how you can use skills and other techniques to move through the entire research process. You’ll learn how to break projects into steps, design multi-stage workflows with built-in checks, and document AI use clearly so that your process remains transparent and reviewable. Examples draw from quantitative, qualitative, and mixed-methods research, with an emphasis on principles that transfer across disciplines. No prior experience with AI agents is assumed.
Hope to see you there!
AI disclosure statement: Claude Code was used to identify resources on this topic. Claude Design was used to create the figure. Grammarly was used for spelling and grammar assistance.