Agent Skills in VS Code — capabilities that load only when needed

VS Code picked up a feature I’d been waiting to see there: Agent Skills (added in 1.108, currently experimental). It’s worth a look — partly for what it does, but mostly for how cheaply it does it.

Enable it with the chat.useAgentSkills setting (still in preview), and the docs cover the rest.

What are Agent Skills?

Originally developed by Anthropic for Claude, Agent Skills are now an open standard (agentskills.io) adopted across multiple tools — including GitHub Copilot in VS Code, the Copilot CLI, and the Copilot coding agent. Like MCP, it’s a way to extend an LLM’s capabilities. The difference in emphasis: MCP gives an agent new tools; a skill gives it new know-how — specialized instructions plus the resources to act on them.

How it works

A skill is, refreshingly, just a folder with a SKILL.md file containing:

detailed instructions,
optional scripts,
examples,
and any additional resources.

The agent loads these dynamically, on demand, based on the conversation. Where do they live? This is in slight flux right now: the open spec defaults to .agents/skills/, while VS Code currently auto-detects .github/skills/ in your workspace (and .claude/skills/ for backwards compatibility), with personal skills in your user directory. (So yes — the newer .agents/skills/ location is the standard one to know.)

Why it costs fewer tokens

This is the part that makes it more than a renamed instructions file. Unlike custom instructions or MCP tool definitions — which are always injected into context — skills use a three-level progressive loading scheme:

Discovery — only the name and description (a few lines of YAML frontmatter) are permanently loaded. That’s all the agent needs to know a skill exists.
Instructions — the body of SKILL.md is loaded only when the skill is judged relevant to what you’re doing.
Resources — extra files (scripts, examples, references) load only when explicitly referenced.

The result: you can have dozens of skills installed without wasting tokens. Only what’s relevant to the current task ever enters the context window. If you’ve read my note on spending fewer tokens, this is the automatic, structured version of the “skill routing” trick — instead of you hand-writing “read database.md for schema changes,” the discovery layer does that routing for you.

Where to find skills

Anthropic (official): anthropics/skills — reference skills including document creation (docx, pdf, pptx, xlsx) and a skill-creator.
Community marketplaces: SkillsMP and Smithery, each listing tens of thousands of community skills (treat the counts as marketing, the good ones as gold).

Build your own

The real value is custom skills tailored to your team. Follow the specification (and VS Code’s create-a-skill guide), and let Anthropic’s skill-creator scaffold the boilerplate. Keep the main SKILL.md lean and push detail into a references/ folder.

Good candidates are anything you explain repeatedly:

how to implement feature flags in your codebase,
how to write and run tests following your conventions,
how to document code/APIs to your standards,
deployment procedures and checklists,
code-review guidelines,
spinning up a new project from your boilerplate.

Basically: anything you find yourself explaining over and over to the team can become a skill. That’s exactly what I’ve been doing on the 4D side — small skills for project info, scaffolding a project, running a method — and the progressive loading is what makes keeping a whole shelf of them painless: they cost nothing until the task calls one up.

Worth reading

Two ways to extend an agent, same shared-standard energy: MCP for tools, Skills for know-how — and Skills come with a built-in answer to the question every one of these features eventually runs into, “won’t this blow up my context?” No: it loads three lines until it needs more. 📜