SKILL.md

As coding agents evolve from simple chat interfaces towards autonomous task execution, context engineering has become a critical challenge. Agent Skills provide an open standard for modularising contexts by bundling instructions, executable scripts and associated resources such as Test-Driven Development. Whilst AGENTS.md is loaded at the start of every session, skills are only loaded on demand based on their descriptions, which reduces token consumption and mitigates issues such as context window exhaustion or agent instruction overload.

At the start of a session, coding agents can scan all available skill files and read a brief description from the Markdown file for each one. This is very token-efficient: each skill consumes only a few dozen extra tokens, with the full details only being loaded when a task is requested that the skill can help solve.

See also

Skills and the programming environment

The Skills mechanism relies entirely on the model having access to a file system, tools for navigating it, and the ability to execute commands within that environment.

This is the main difference between Skills and previous attempts to extend the capabilities of LLMs, such as MCP. Skills therefore offer several advantages:

Advantages of Skills over MCP

Skills are more efficient

GitHub’s official MCP, on the other hand, consumes tens of thousands of context tokens on its own, and as soon as a few more are added, the LLM is left with hardly any room to actually do any useful work.

LLMs, on the other hand, know how to call CLI-TOOL --help, so we don’t need to use many tokens to describe its usage – the model can figure that out for itself later if necessary.

Skills can also be used with other models

You can simply take a Skills folder and point Codex CLI or Gemini CLI to it using:

Read uv-tdd/SKILL.md and then create a project structure

This will work even though these tools and models have no built-in knowledge of Skills.

Skills are more secure

The instructions can be executed in secure programming environments.

Skills are simpler

MCP is a complete protocol specification featuring hosts, clients, servers, resources, prompts, tools, samples, roots and three different transport protocols: stdio, streamable HTTP and, originally, HTTP with SSE. Skills, on the other hand, are based on Markdown with a little YAML metadata and some optional scripts that can be executed in the respective environment. They are therefore much closer to the concept of LLMs, as you can simply enter text that the model interprets.

See also

Skill plugins

As their popularity has grown, the surrounding ecosystem has also expanded. Plugin marketplaces are emerging as a way to version and share skills, and numerous projects are exploring how to assess the effectiveness of skills. Nevertheless, you should not use third-party skills without first verifying them, as they pose serious security risks within the software supply chain.

See also