I use AI agents every day. What really changed how I work isn’t a new model. It’s a skill.
The annoying problem with agent memory
Some agents now have memory between conversations. Useful, but limited: it depends on the agent, it’s not portable across platforms, and you don’t really control what it retains.
If you have a particular writing style, a way of structuring your PRs, or a way of thinking through architecture decisions, you want that to be explicit and reproducible everywhere.
Skills fix that.
What is a skill?
A skill is a folder with a SKILL.md file inside. That file contains a bit of YAML as a header and instructions in Markdown. That’s it.
my-skill/
└── SKILL.md
The YAML has a few fields, but two are required: name and description. The agent reads that description at startup to know when to use the skill. When your request matches, it loads the full instructions into its context and follows them.
---
name: grill-me
description: Interview the user relentlessly about a plan or design until reaching shared understanding.
---
Interview me relentlessly about every aspect of this plan...
This is the progressive disclosure principle: the agent first loads only the metadata (about 100 tokens per skill), then the full content only when relevant. You can have dozens of skills without blowing up your context window.
grill-me, my favourite skill
grill-me is a simple skill. When I bring it a plan, an idea, or a technical decision, it asks me questions one at a time until we’ve really dug into the subject. It pushes me to think about things I wouldn’t have thought of.
That’s what a good colleague does. Except this one is available at 11pm when I’m trying to decide how to structure a blog series.
This article was planned with grill-me, by the way.
One skill, every agent
The SKILL.md format is an open standard. Anthropic created it in December 2025, and adoption has been quick.
Today, the same SKILL.md works in:
- Claude (claude.ai Projects and Claude Code)
- GitHub Copilot (agent mode in VS Code, Copilot CLI, Copilot cloud agent)
- ChatGPT and OpenAI Codex CLI
- Cursor, Gemini CLI, and a dozen others
You build a skill once, it works everywhere. It’s a bit like what MCP did for tools: a common standard that keeps every platform from reinventing the wheel.
Where to start?
If you want to explore existing skills before creating your own, the two places to know are anthropics/skills for Anthropic’s official and community skills, and github/awesome-copilot for Copilot’s community collection.
The official documentation is available on the Agent Skills side and on the GitHub side.
In the next post, we build a skill from scratch.
Happy exploring, and if you try grill-me, be ready to answer some questions.
This post was written with AI assistance and edited by me.