The Complete Guide to Building Skills for Claude (PDF)

For a while I was very into maxxing out my productivity using LLMs. I still am, a little, but not so much. Still, I’m very intrigued by LLMs like Claude and what you can do with them.

A skill is essentially a standardized set of specific instructions for Claude. In this way it’s a bit like a program. Except, where most coded programs are deterministic–i.e., for a given set of inputs, they’ll return the exact same output every time, a Claude skill, like LLM prompts in general, are non-deterministic. This makes them pretty useless for tasks where accuracy and consistency are required.

Part of what got me down on working with LLMs is the realization that the quality of what you get from them can be pretty random. Sometimes you’ll have an amazing conversation, or generate a brilliant piece of code, or get a novel answer to an interesting question you’ve spent a lot of time with. Other times the LLM will just be confidently wrong and give you complete nonsense. In that way it’s a bit like drugs. You’ve probably heard of drug addicts chasing the feeling of their first high. Behavioral scientists call this action “random reinforcement” and it’s one of the most powerful levers in driving animal behavior. It’s essentially gambling. Sometimes when you press the button you win the jackpot, most of the time you get nothing. Your brain remembers the jackpot and wants to keep doing the action that caused it, even if most of the time you get nothing.

So people have a couple mind-blowing experiences with LLMs, interspersed among a bunch of trash nonsense slop, but they don’t remember the slop, they remember what was amazing, and chase after that.

Oddly enough, when we speak of a person being skilled, we usually mean the ability to do something well every time, not just randomly here and there. Anthropomorphism of AI is one thing but let’s at least get our metaphors right.