AI Agent Coding Skills

Paul Hudson has been posting several agentic AI skills this past week, such as this one for SwiftUI.

An AI Agent skill is an interesting yet simple thing. It’s a text file, and it tells the AI tool “do these things”. A snippet for the SwiftUI one…

Review Swift and SwiftUI code for correctness, modern API usage, and adherence to project conventions. Report only genuine problems - do not nitpick or invent issues.

Review process:

1. Check for deprecated API using references/api.md.
2. Check that views, modifiers, and animations have been written optimally using references/views.md.
3. Validate that data flow is configured correctly using references/data.md.
4. Ensure navigation is updated and performant using references/navigation.md.

This is actually a powerful thing, for two reasons. The first is shown by Sebastian who notes on Mastodon:

As a designer who doesn’t know what he is doing I’ve longed for something like this to make sure the code my AI is spitting out works well. I kinda even tried building skills and docs with Claude 2 weeks ago, but the thing is: I have no idea what I’m doing.

Paul knows exactly what he is doing, I’ve been learning so much from his resources already and these new skills and agent files are soooo valuable for all the people like me who had the barrier to entry lowered so much with agentic coding.

So – you’re getting expert direction for your AI to help write your code.

But there’s a second and more nuanced thing happening. You could include all of Paul’s information in your prompt to the AI. Or you could have an AGENTS.md markdown file that most agents¹ will pickup with each prompt. But if you include a few dozen lines and a few hundred words with each prompt, your context window² is going to shrink, faster because you’ve included the same information over and over again.

The cleverness behind the agentic AI skill is that you only use it one time, a la:

Use the Swift Concurrency Pro skill to look for potential concurrency problems in this project.