Intro
Giving your agentic coding assistant a precise description of what to do—and how to obtain the relevant project context—can be critical to both developer experience and outcome quality. Clear instructions reduce friction, improve alignment, and significantly increase the likelihood that the agent delivers useful results.
To make coding agents more flexible, helpful, and better tailored to day-to-day engineering tasks, Anthropic—after introducing the Model Context Protocol (MCP)—has added Claude Code Skills. A skill is a markdown-based standard that follows a defined convention. Each skill is described in a SKILL.md file whose frontmatter is provided to the language model. This frontmatter explains what the skill does and, crucially, when it should be used. Based on this metadata, the agent can decide autonomously when to load and apply a specific skill.
The real power of the concept lies in timing: the agent receives the right instructions at the right moment. Conceptually, this is very similar to what were once called Semantic Functions in Microsoft’s Semantic Kernel. Skills allow you to shape distinct roles for your agent. For example, you might define:
- an architecture skill focused on analyzing codebases and suggesting refactorings aligned with your architectural principles, or
- a review skill that applies strict scrutiny to changes to ensure compliance with internal coding standards.
In practice, skills solve a common pain point: repeatedly reusing nearly identical prompts to get an agent to analyze features or generate implementations. With skills, those instructions are captured once and reused consistently—at least if you were using Claude Code.
That limitation is now gone. As of recently, the VS Code Insiders build supports Claude agent skills, which means we can reuse the growing ecosystem of community-defined skills directly inside VS Code. This is a significant step forward for agent-driven development workflows.
The SKILL.md Standard
Before diving into setup with VS Code, it helps to understand how agent skills work. The official specification is available at agentskills.io if you want a deeper look at the mechanics.
Below is an example of a skill taken from the claude-superpowers skill library. This particular skill helps the agent collaborate with you on brainstorming new features or technical concepts.
As mentioned earlier, the frontmatter briefly describes the skill and defines when it should be used. Each SKILL.md file lives inside a folder named after the skill itself. In this case, the file would be located in a folder named brainstorming.
Setting Up Skills with VS Code Insiders
At the time of writing, agent skills are only supported in VS Code Insiders, so this guide assumes you are using the green (Insiders) build rather than the stable blue one.
First, you need to explicitly enable the feature:
- Open Preferences
- Navigate to Features → Chat
- Enable Use Agent Skills
Once enabled, you can configure skills either globally or per project:
-
Global skills:
Place them in ~/.copilot/skills. These will be available across all projects. -
Project-specific skills:
Place them in .github/skills at the root of your repository. These skills will only apply to that project.
If you do not want to write your own skills initially, you can start with existing libraries such as claude-superpowers. Simply copy the skills folder from the repository into your preferred location.
After completing the setup, restart VS Code. When you ask the agent to perform a task that matches one of the available skills, it will automatically activate that skill. The corresponding SKILL.md file will be attached to the conversation context, making the behavior transparent and inspectable.
Limitations and Considerations
As of now, agent skills in VS Code are still experimental. Settings, behaviors, and APIs may change over time, so expect some churn.
Another important consideration is billing. Many skills—especially interactive ones like brainstorming—encourage multi-step dialogue. This does not bypass GitHub Copilot’s request-based billing model. Each back-and-forth interaction still consumes requests.
For exploratory or highly interactive workflows, it may therefore be sensible to use cheaper or free models (such as GPT-5-mini, Claude Haiku, or Gemini Flash) and reserve more capable—and more expensive—models for implementation-heavy tasks.
Despite these limitations, agent skills already represent a powerful abstraction. They move prompt engineering out of ad-hoc conversations and into reusable, composable building blocks—bringing us one step closer to truly programmable developer agents.