Agent Skills

Something very interesting is happening right now in AI.

I’m seeing the development and adoption of something new: Skills. And I find it really fascinating.

What are Agent Skills?

At their core, skills feel like the next evolution of prompts.

Skills are reusable capabilities for AI agents. Install them with a single command to enhance your agents with access to procedural knowledge.

This is the definition of AI agent skills on Skills.sh, an initiative by Vercel.

Skills can have References

A skill can contain specific instructions on how to do something. It can also pull in references, which allow a skill to gain access to additional context about how to do something.

The beautiful part is that a single skill may be capable of doing 10 different things. But unlike a prompt, the core skill can remain lightweight and simple, and given instructions to use references for more specific knowledge.

A skill should do one thing really well

The approach is simple - One task, one skill. If you have multiple tasks, you use multiple skills.

A single skill does not need to know how to do everything.

The best skill can do one thing really well. And from experience, we have found that a chat with an AI agent performs really well when given a specific task with sufficient context. And after the task is complete, for best results, spin up a chat with another AI agent for another specific task.

That’s not very different from how humans operate.

We have specific skills, that we learn through practice and repetition and real-world experience. Those specific skills help us perform that specific task really well, and with low rates of error.

Skills can be combined

But in the real world, we don’t operate our skills in isolation. We combine one skill with another, and another.

A human being is a collection of skills, that can depend on, or rely upon each other. And together, they are capable of performing a bigger task.

I think in the same way, Agent Skills are meant to be combined in order to work together and build a bigger picture.

Combining Skills require Orchestration

When us human beings need to do something complex, we orchestrate the various skills we have in order to perform the task well.

I think that’s the right mental model to think about Agent Skills.

Individual skills should be good at specific tasks and solving specific problem statements.

An Orchestrator Skill should be responsible for overseeing the specific skills. The orchestrator knows what needs to be done, but does not necessarily know the best practices or how to get it done.

Orchestrators need Reviewers

When we do a task, we also need a way to verify that the task was done correctly.

We need a way to review the output and quality of the task, so that it matches up to expectations.

I think Orchestrators need Reviewer Skills, whose sole purpose is to review the output of a specific Agent Skill.

That way, the Orchestrator’s role becomes

• Do this task
• Is the task done?
• Is the task done correctly?
• Moving to the next task
• Do this task…

How does the Orchestrator know what list of tasks need to be done?

That should be a skill ;)

Mental Model - The Assembly Line

This is where things get interesting.

You can have a high-level skill that acts as an orchestrator of the entire task. Another set of skills are good at specific tasks. Another set of skills act as reviewers of their outputs.

Additionally, you can add references, which allow sharing real code snippets or patterns, allowing a skill to both handle ambiguity and be extremely precise.

The picture I want to impart to you is that of an assembly line.

There are tons of things happening, step by step.

The machinery at each step only knows how to take the previous input, transform it, and pass it on to the next machinery, who will review the output before passing it to the next step in the assembly line.

And the orchestrator used to be you, the human imagining this scene I just laid out in the previous paragraph.

This orchestrator can now be an AI Agent.

Skills enables true software-based process automation

The assembly line mental model has been around for a while. It’s the thought process behind any business process automation, or even software-based process automation.

But software automation of the past would be complex IF THEN ELSE statements. They would be rigid, and while they would perform admirably when everything goes according to the defined code and constraints, they would be very brittle in the face of ambiguity.

Humans are by default capable of handling ambiguity. We do it all the time, every single day. And in business process automation which involves humans doing the work, small shifts and ambiguities are being tackled all the time. And when the shift or ambiguity is too big, the issue is raised to the manager, or their manager, and all the way to the top.

I think that the arrival of AI, agents and skills makes it possible to have true software-based process automation.

LLMs are capable of handling these small ambiguities. And if they are unable to do so, they would raise it to their manager LLM, and all the way to the top.

Who’s sitting at the top?

I think it’s humans at the top, as of today. But I really can’t say what that would look like in the future.

Agent Skills Github Repository

I’ve started a new GitHub repository called agent-skills.

Right now, I actively use three specific skills in my development workflows:

• Code review: To enforce local codebase invariant rules.
• Git commit messages: To automate perfectly formatted history.
• Sentry analysis: To map production crash logs directly to actionable fixes.

It’s quite basic. But I think understanding the building blocks of it is powerful, so I’m calling this a win :)

My repo is super simple right now. But I’ll keep updating the repository with new skills - could be ones I create, or perhaps I can reference skills made by others that I’ve adopted.

Over time, I think these building blocks will fit into a much grander picture of how we build things in the future.

📖 Continue Reading