7 min · +100 XP
A Safe Space (Two Simple Rules)
Not to scare you — to un-scare you.
By the end of this lesson
You'll know the two habits that make all of this boring-level safe.
Before we wrap up: a few words about safety. Not to frighten you — the opposite. Once you know the two simple habits, there's nothing left to be scared of.
You'll occasionally hear scary terms like 'prompt injection' — that's when someone hides sneaky instructions inside text an AI reads, hoping it follows them. It's a real concern for big public systems with thousands of users. For your personal tools on your own computer, it's mostly a non-event — with ONE exception worth knowing: a tool that makes Claude read text from strangers (incoming emails, random web pages) while in don't-ask mode. Keep those tools look-but-don't-touch — read and summarize, not act and send — and you're covered. Know the word exists; don't lose sleep over it.
The two habits
- 1Give yourself a clean space. Don't run experiments on a computer that holds things you can't afford to touch — a crypto wallet whose seed phrase you never wrote down, the only copy of critical files. A regular home computer with normal stuff? Already fine.
- 2Dangerous outcomes require dangerous requests. Claude does what you ask, in the folder you're in. Ask it to optimize your photos — it optimizes them; it doesn't delete them. Don't ask for potentially dangerous things, don't point it at sensitive places, and you simply don't get dangerous consequences.
That's movie logic. In real life it edits files in your project folder because you asked it to build something there. The boring truth: the risk profile of building a habit tracker is the risk profile of editing a Word document.
Clean space + sane requests = nothing to fear. You've been following both habits this whole course without noticing.
What actually keeps your AI building safe?
Author
Evgeny Arsentyev
PhD · Chief Product Officer at a healthtech company