Agent Skills are the new boom. If you’ve been following the agentic AI space since late 2025, you’ve likely noticed a shift. Previously, we spent most of 2024 and 2025 obsessing over tools—giving LLMs calculators, database access, and API keys. But giving a new employee a laptop and a login doesn’t make them productive. You still have to train them. For example, you have to show them how your organization works. That is exactly what Claude Agent Skills are.

Released by Anthropic in October 2025, Agent Skills have quietly become one of the most powerful features in the Claude ecosystem, and are now even expanding as an open standard across other toolings. In fact, this is perhaps even more significant than the Model Context Protocol (MCP) for everyday developers.

What Exactly Are Agent Skills?

1. Discovery: At startup, Claude only reads the metadata (Name and Description) of your skills. It knows what skills it has, but not how to do them yet.
2. Activation: When you ask a question that matches a skill’s description (e.g., “Generate a weekly report”), Claude dynamically loads the full SKILL.md and related files into its context.
3. Execution: Claude follows the procedure you defined to complete the task.

This approach had three fatal flaws:

1. Token Hemorrhaging: Sending a 2,000-word style guide with every request cost money and burned through rate limits, even if the user just said “Hello.”
2. Attention Dilution: When an LLM is given 50 instructions at once, the “Lost in the Middle” phenomenon occurs. The model prioritizes the beginning and end of the prompt, often forgetting the nuanced rules buried in the center.
3. Latency Spikes: Processing a massive system prompt takes time (Time to First Token), making chat interfaces feel sluggish.

The Big Question: Agent Skills vs. MCP

• MCP is for Connecting: It gives Claude eyes and ears. It connects Claude to a PostgreSQL database, a Slack channel, or a local file system. It is the “plumbing.”
• Agent Skills are for Experties: They give Claude procedural knowledge. They teach Claude the process of how to write a SQL query that matches your company’s naming conventions, or how to format a Slack message so it doesn’t annoy your boss.

How to Build Your First Agent Skill

1. Create the Directory Structure

.claude/
  skills/
    python-best-practices/
      SKILL.md
      linting-rules.txt

2. Write the SKILL.md

---
name: Python Best Practices
description: Use this skill whenever the user asks for Python code. It ensures code is type-hinted and PEP8 compliant.
---

# Instructions
When writing Python code, you must adhere to the rules defined in `linting-rules.txt`.

1. Always use Python 3.12+ type hinting.
2. Include docstrings for every function.
3. Before outputting the final code, cross-reference it with the attached linting guide.

Advanced Agent Skill Construction: Beyond Basic Markdown

Here is the anatomy of a “Senior” level skill.

The Variable Injection Pattern

Hard-coding values into your SKILL.md limits reusability. Instead, use placeholder variables that Claude can infer from the conversation context or ask the user to clarify.

Example SKILL.md for a cold-email generator:

# Instructions
You are an expert copywriter. You will draft a cold email based on the {{TARGET_AUDIENCE}} and {{VALUE_PROPOSITION}}.

<steps>
1. Analyze the user's input to identify the {{TARGET_AUDIENCE}}. If missing, ask the user to define it.
2. Review the attached `successful_examples.json` to match the tone.
3. Draft the subject line using the "Curiosity Gap" framework.
</steps>

The Context Library Strategy

A single Markdown file often isn’t enough. You can create a “Library” subfolder within your skill to hold reference materials.

Directory Structure:

skills
  legal-contract-review/
    SKILL.md
    reference/
      clause-library.txt
      risk-matrix.csv
      jurisdiction-guidelines.pdf

In your SKILL.md, you reference these specifically:

“When analyzing liability clauses, cross-reference the text against reference/risk-matrix.csv. If the liability cap exceeds $1M, flag it as ‘High Risk’ immediately.“

XML Guardrails

<critical_rules>
- NEVER output API keys or PII (Personally Identifiable Information).<br />
- IF the user asks for financial advice, output the standard disclaimer located in `disclaimer.txt`.<br />
- ALWAYS output the final code block in a copy-able artifact window.<br />
</critical_rules>

The “Agent Skills Chain” Workflow

For example, you might have:

1. Researcher Skill: Scrapes web data and summarizes findings.
2. Writer Skill: Takes summaries and turns them into blog posts.

To optimize for this, ensure your SKILL.md descriptions clearly state their inputs and outputs.

• Bad Description: “Helps with writing”
• Good Description: “Accepts raw technical summaries and converts them into SEO-optimized blog posts in Markdown format”

Agent skills for DevOps & Cloud Engineers

• Team Consistency: Commit a infrastructure-deployment skill to your team’s repo. Now, every engineer using Claude Code gets the exact same Terraform tagging standards automatically.
• Compliance: Create a security-audit skill that forces Claude to check for open S3 buckets whenever it reviews CloudFormation templates.

Agent skills as an open standard, and adoption

This shift is critical for enterprise adoption. It means companies can build a repository of Corporate Operating Procedures as Agent Skills without fearing vendor lock-in.

A Prime Example: VS Code Integration

We are already seeing the immediate benefits of this standardization in developer tooling. In its latest release, Visual Studio Code has launched a preview feature that natively supports the Agent Skill standard.

Because the standard is open, VS Code didn’t need to build a custom Microsoft version of skills. Instead, the IDE can simply parse the same .claude/skills (or the vendor-agnostic .agentskills) directory you already use. This allows developers to invoke their custom skills directly within the editor—applying linting rules, generating boilerplate, or running tests—without ever leaving their coding environment.

This is just the first wave of tools adopting the standard, bridging the gap between chat interfaces and local development environments. Similar to how the creation of MCP prompted a quick adoption across the developer ecosystem and tooling.

Conclusion

References

• Claude Skills are awesome, maybe a bigger deal than MCP
• Introducing Agent Skills | Claude
• Agent Skills – Claude Docs
• Use Agent Skills in VS Code
• Overview – Agent Skills
• Don’t Build Agents, Build Skills Instead – Barry Zhang & Mahesh Murag, Anthropic – YouTube