Incorporating AI Skills Into the Making Process

I think I found where programmers really add value to the LLM skill ecosystem. A lot of basic tasks can be handled by the LLM's training data. It's able to compose scripts on the fly to accomplish a lot of tasks, especially with small datasets.

In the real world, datasets aren't always small, consistent, or clean. Exceptions happen.

Skills let you marshal the AI into doing a specific job by running well-tested code. You have to program the skill like a little utility app. It took me hours to get a TSV-to-SQLite skill working, for instance.

You'll never one-shot some skills, but using the agent to create your skill outline then prompt-updating it (vibe coding) can result in a robust and repeatable task.

So what makes a skill different from just downloading a package on npm and running it?

You can talk to skills. I made one where it turns my questions about a database into SQL statements and runs them, but since it investigates the schema first it makes it very flexible. It seems to work on any DB I throw at it. That's way more flexible than any ORM I've ever used.

Code that ships is only part of the development process. We're constantly examining artifacts to fix bugs or writing one-off utility functions to transform data. Skills are amazing for this because they can be shared and used by just about any model.

I'm starting to get it.