Introducing Agent Skills (article)
Introducing Agent Skills (article)
A quick read to understand the idea and how skills are designed.
An article by Anthropic's engineering team on why Agent Skills exist and how they are designed.
Overview
This is an article on Anthropic's engineering blog, written by Barry Zhang, Keith Lazuka and Mahesh Murag and published on October 16, 2025, introducing the ideas and design behind Agent Skills. It was updated on December 18, 2025 with the news that Anthropic has made Agent Skills an open standard that works across platforms.
The main argument: even a powerful model lacks procedural knowledge and an organization's own context. Instead of building a separate agent for each job, you can "specialize" a general-purpose agent by adding skills, much like handing a handbook to a new employee.
The article explains the progressive disclosure mechanism, how skills combine with code execution, principles for building and evaluating skills, and security considerations. It is a foundational document that helps product builders and engineers understand why skills work.
Highlights
Skills as an onboarding handbook
Each skill packages instructions, scripts and resources so the agent can find and load them when the work calls for it.
Three levels of progressive disclosure
Name and description are preloaded, SKILL.md content is read when relevant, and supporting files are opened only when truly needed.
PDF skill example
A PDF skill helps Claude fill in forms; the form-filling instructions live in a separate file that is read only when needed.
Code for work that must be exact
Bundled scripts handle deterministic operations such as sorting or extracting form fields, giving repeatable results.
Future direction
Anthropic wants to support the full lifecycle of skills, combine them with MCP, and move toward agents that create, edit and evaluate skills themselves.
What it's for
Shaping an AI strategy
Business owners read it to understand how to turn internal knowledge into skills instead of building many separate chatbots.
Designing skills properly
Engineers apply the principles of splitting reference files and using scripts for operations that need accuracy.
Evaluating before building
The AI team runs the agent on representative tasks, finds weak spots, then writes skills to fill the gaps.
Building a security process
The IT department uses the security section as a basis for a policy of installing skills only from trusted sources.
How to use it
- 1
Read the core idea
Understand why skills are needed: models are powerful but lack procedures and organizational context.
- 2
Learn the skill structure
Look into SKILL.md with its name and description frontmatter and how supporting files are referenced.
- 3
Understand progressive disclosure
See how the three levels of loading let a skill hold many documents without taking up the whole context.
- 4
Apply the principles for building skills
Start from evaluation, design for extension, observe how Claude uses the skill and improve it together with Claude.
- 5
Review security
Install skills only from trusted sources; for unfamiliar skills, check every file, especially scripts and calls to external networks.
Video guides
Don't Build Agents, Build Skills Instead – Barry Zhang & Mahesh Murag, Anthropic
AI Engineer
Agent Skills: Specialized capabilities you can customize
Anthropic
Real-world example
Suppose a small-to-medium logistics company wants to use Claude to draft shipping quotes. After reading the article, the tech lead doesn't write one long prompt but splits it into a skill: SKILL.md describing the quoting procedure, a separate file with the price table by route, and a script that calculates fees by weight so results are always consistent. The team has Claude try it on past quote requests, notes where it is wrong, and asks Claude itself to add those lessons to the skill.
Tips & notes
Invest in name and description
These decide whether the skill is triggered at the right moment.
Separate rarely co-used context
Content that is seldom needed at the same time should go in separate files to save tokens.
Say whether a script is to be run or read
Add a note in SKILL.md so Claude knows whether to execute the script or only consult it.
Let Claude record lessons itself
After each attempt, ask Claude to write successful approaches and common errors into the skill.
FAQ
Who is the article for?
Suitable for engineers, product builders and managers who want to understand the design of Agent Skills before applying them.
Where are skills supported?
The article says skills are available on Claude.ai, Claude Code, the Claude Agent SDK and the Claude Developer Platform.
Are skills related to MCP?
Anthropic plans to find ways for skills to complement MCP servers, by teaching agents multi-step workflows that use external tools.
Do I need to know programming to read it?
The ideas section is easy for non-specialists; the code execution and best practice parts are more useful for technical readers.
Useful links
Want to apply this in your business?
2dm looks at your workflows, picks the right tools and sets them up with your team. The first call is free.
Same topic
Anthropic Skills
4.5
Anthropic's official repository of example skills for Claude: each skill is a folder of instructions and resources that helps Claude do one specific kind of task well.
Agent Skills documentation
4.0
Explains what a skill is, the SKILL.md folder structure, and how Claude loads skills automatically when needed.
Skills in Claude Code
4.0
How to create, install and use your own skills for Claude Code, on your machine or per project.