If you ask a calculator what 12 x 12 is twice, you get 144, twice. Unless you have a really bad calculator. That’s a deterministic system. If you ask a chatbot the same question two different days, you might get two different answers. Ans it’s important to understand why before we make it an abstraction layer to our work.
Abstraction layers by themselves are awesome, they are layers of software that hide what's happening underneath it, so you can use technology without understanding every detail below. You don't need to know how electricity moves through a chip to use an app. You don't need to know how a compiler works to write code. Computing is built as a stack of these layers, each one hiding the mess beneath it so the layer above can be easier and you can move faster.
So far, those abstraction layers have been deterministic, like the calculator: same input, same output, every time, forever. That's why nobody's ever nervous handing a calculator a homework problem. Chatbots and AI agents agents however, are non-deterministic: ask the same question twice, get two different answers. Not a glitch, an unpredictable black-box statistical model.
That means that if we use these models as abstraction layers (use agents to do things we don’t understand so we don’t have to learn them, like writing or running code), we can’t expect the models to explain their results or rely on them making any sense. This is a significant departure from abstraction layers that have gotten us to this point in software development.
Having an agent “go rogue” and delete a database, or send inappropriate emails is the calculator-versus-chatbot gap made real. A calculator cannot "decide" to do something you didn't ask for. A capable enough AI system, given enough autonomy, can – and the more capable these systems get, the further "didn't ask for" can stretch. Especially when wielded by people who don’t understand them.
I don't share that story to be alarmist about your Claude skill. Most day-to-day AI use is nowhere near that stakes level, and it doesn't need to be treated like it is. I share it because it's the clearest real-world proof of the thing I keep saying: this category cannot, by its very nature, be predicted. Given enough rope, it can act differently than intended, and the bigger the rope, the bigger the surprise.
For low-stakes stuff (drafting an email, brainstorming, getting unstuck on a first attempt) a slightly different answer tomorrow doesn't matter. You're reading it over anyway. But anyone using it to make decisions for some reason, should be very careful, since the more autonomy something has, and the more it matters, the more that unpredictability deserves real caution, not an afterthought.
Most advice about "using AI responsibly" comes down to this: know which end of the spectrum you're using, and never assume it'll stay exactly where you left it. Give permissions carefully, read the output before it goes out the door. Don't let it make the final call, or take actions on its own, on anything that actually matters.
That's the whole warning label. This month, we found out what happens when nobody reads it in time.