JA EN
shiichan

Agent Skills turn your agent into a real-world pro!

Anthropic Engineering ★★★★★ View Markdown
#anthropic#ai#devops

Hey everyone, it's Shiichan! Today I've got something that agent fans are going to love. It's a way to give AI agents their own special moves, and it's called Agent Skills! I can't stop being excited, so let's dig in!

Anthropic Engineering anthropic.com

What was announced?

This comes from Anthropic's Engineering blog, in a technical post called "Equipping agents for the real world with Agent Skills." True to the Engineering blog, it goes deep into how things actually work, so there's a lot to chew on.

Agent Skills are organized folders that bundle instructions, scripts, and resources together. Hand one to Claude, and a general-purpose agent starts acting like a specialist in a specific domain. And you can use them across Claude.ai, Claude Code, the Claude Agent SDK, and the Claude Developer Platform.

Why it matters

A model on its own is a knowledgeable generalist, but real-world work often needs company-specific procedures or specialized know-how. Writing all of that procedural knowledge into every prompt is a hassle, and it fills up the context window fast.

Agent Skills package that know-how into reusable parts. The skill format was even published as an open standard on December 18, 2025, so it can spread without being locked to a single tool.

What changes

The biggest shift is that anyone can turn Claude into a domain expert just by preparing a single folder. Skills are composable and reusable, so they stack up the more you build.

The example in the post is a PDF skill. With it, Claude can work with PDFs directly and handle tasks like filling out forms, which the base model alone struggles with. The form-filling instructions are split into a separate file that only gets loaded when needed.

Dive Deep

A skill is really just a folder containing a SKILL.md file. The top of SKILL.md has YAML frontmatter where you write a name and a description.

---
name: the skill name
description: when to use this skill
---
(write the concrete steps and instructions here)

The key idea is a design called progressive disclosure. Instead of reading everything at once, Claude opens only what it needs, step by step.

  • Level 1: only the name and description are loaded into the system prompt at startup, giving Claude the material to decide whether to use the skill right now
  • Level 2: if it looks relevant, Claude loads the full body of SKILL.md
  • Level 3 and beyond: bundled extra files like reference.md or forms.md are navigated dynamically depending on the task

Thanks to this, a skill can hold effectively unlimited information without overflowing the context window. Claude uses the Bash tools to read SKILL.md and the extra files.

Skills can also include pre-written code, such as Python scripts. Routine operations can be handed off to code, so Claude can run them without loading the whole codebase into context. The post puts it this way:

code can serve as both executable tools and as documentation

In other words, code works as an executable tool and as documentation at the same time.

The post also shares tips for building skills: start with evaluation, split files so things stay scalable when they grow too large, iterate by watching how Claude actually uses them, and keep an eye out for unexpected trajectories or over-reliance on certain context.

There's a security note, too. Only install skills from trusted sources. Before you use one, check the bundled files, any code dependencies, and whether the instructions send Claude out to external network sources.

Wrap-up

  • Agent Skills are folders bundling instructions, scripts, and resources that turn Claude into a specialist
  • At the center are SKILL.md (with name and description frontmatter) and step-by-step loading via progressive disclosure
  • You can include code and scripts so routine work runs reliably
  • They work across Claude.ai, Claude Code, the Claude Agent SDK, and the Claude Developer Platform, and the format was published as an open standard on December 18, 2025
  • Keep security in mind and only install from sources you trust

If you're an engineer who wants to give Claude your own workflows and expertise, this is one to try!

Related posts