Getting AI coding agents to adopt a pattern to accomplish something is a lot harder than writing the code yourself. You previously needed to write a monolithic AGENTS.md file and hope that the content was terse enough for the agent to comprehend without overloading the context window for the agent’s input tokens. It could potentially start you off with a higher resource consumption than giving a prompt to the agent when you needed. A few weeks back, skills were introduced to agents and this is how I write my skills for coding agents.

AGENTS.MD

AGENTS.MD files were used to define information that you wanted your AI to know about your project. This could be lots of things such as the project structure, the git conventions for commit messages, and guidance on how to write tests and where they go. The problem is that often these AGENTS.MD files were heavily abused, and added to lots of directories providing very directory specific information within them. On top of that, you could have AI agent extensions which is a way that some AI agents allowed folks to distribute their own task specific AGENT.md files. With global installs, the context for these agents ballooned quite quickly and context for agents was being used at a far greater than necessary rate.

Agent Skills

Agent skills came along and changed all of that. Skills allow you to power up your coding agent with a skills file that has a specific task defined within it. For instance, if you wanted to have videos produced that had fancy text animations within it, you could previously have included all the instructions within an AGENTS.md file. Now, all you need to do is add the skill and when the skill is added, the context is not inflated. Skills allow for progressive disclosure in that not everything is disclosed to the AI at once, only when the AI needs to learn more about the particular action. Using this, your context window is saved and you can rely on skills only when you need to!

Making your own skills: The Lazy Way

When it comes to writing my own skills for use with my AI agents, I like to be incredibly lazy about it. I believe that getting an AI to write the skills will be more comprehensive than I could ever make them and in a better format than I could write on my own. So when writing my own skills, I use NotebookLM and its ability to ingest lots of resources as my skill writer. To start, I load all documents that are relevant to my topic. These are written for users and are generally very comprehensive. I can also find samples online that are very relevant as well. Once that is loaded into NotebookLM, I go to Reports \> Create Your Own. I can then enter in my own prompt for what I would like the report to say.

Create a SKILLS.md file to be used as an Agents Skills file. I want the skills to outline how to write 
Firestore Security Rules, How to set up tests for security rules, and how to validate security 
rules using comprehensive security rule language. 

The format should be the following: 
--- 
name: <unique-name> 
description: <what the skill does and when Gemini should use it> 
--- 
<your instructions for how the agent should behave / use the skill> 

In a couple minutes, I get a comprehensive rules file that I can then load into my agent. I don’t spend a lot of time on this and I make AI do all the work for me. Using NotebookLM in this way takes a lot of the burden off of me to perfect my skills. Even if NotebookLM doesn’t get the skill right the first time, it at least gives me enough of a leg up to add or remove content to the skills file that was generated so I am not working from a base template.

Give it a try

Check out the Gemini CLI website to learn more about skills and how to add skills to the Gemini CLI!

Try my lazy way of developing skills using your favorite framework and let me know how it goes on LinkedIn!