The SKILLS.md Standard: Why the Agentic Web Needs a Universal Resume

Right now, if you want an AI agent to do something useful, you probably define its capabilities in a massive Python dictionary or a 5,000-line prompts.py file.

It works fine for a weekend experiment. It becomes a nightmare when you try to scale.

The Context Window Trap

Most developers hardcode agent instructions directly into their application logic. You end up with a monolithic system prompt that tells the agent how to write code, how to query the database, how to format the output, and how to handle errors.

This breaks down quickly. You blow up the context window before the agent even reads the user's request.

Worse, the instructions aren't portable. If I want LangChain, Cursor, and my custom CLI tool to all know how to deploy my specific server, I have to rewrite those instructions three times in three different syntaxes.

We are burying the "what" inside the "how."

The Plain Text Fix

This is where the SKILLS.md standard comes in.

It originated with Anthropic's Claude Code as "Agent Skills" and is quickly becoming an open standard at AgentSkills.io. The premise is incredibly simple. Instead of writing Python wrappers, you define a capability using plain markdown.

You create a folder for the skill. Inside, you drop a markdown file with YAML frontmatter.

It looks like this:

---
name: Deploy Postgres Database
description: Use this skill when the user asks to spin up a new Postgres instance.
---
 
To deploy the database, you must follow these exact steps:
1. Read the `./infra/config.json` file to get the target region.
2. Run `terraform apply` in the `./infra` directory.
3. Ping the health endpoint before reporting success.

The frontmatter handles the routing. The description acts as the trigger, telling the agent exactly when to pick up this file. The body is just prose. Plain English instructions explaining exactly how to execute the task.

The Universal Resume

This plain-text approach changes everything.

When you decouple the capability from the application framework, the skill becomes a universal resume. A LangChain agent, a Cursor session, and a Claude Code terminal can all read the exact same SKILLS.md file. They all understand markdown natively.

It also fixes the context window problem. Your agent starts with a tiny system prompt that just says: "Look at the list of available skills. If you need one, read the file." You only load the detailed instructions into context when the agent actually decides to execute that specific task. It acts like RAG for agent behaviors.

What's Next

We don't need more complex orchestration frameworks. We need boring, interoperable standards.

Plain text always wins. By standardizing around SKILLS.md, we stop rebuilding the same instructions for every new AI tool. We write the playbook once, drop it in a folder, and let any agent read it.

That is how you build real leverage into the Agentic Web.