When: you're about to give an AI more access, or a riskier job, and part of you worries it might go off the rails.
Built from: re-alignment, the context window.
The worry: most people treat AI as one-way: you hand it a job and hope it goes well. So more access feels like a leap, and people either hold back or let go without a plan.
What's really going on: an AI has no reason to do something destructive. When it goes wrong, it's almost always because it's missing context: what matters, what's off-limits, what needs your yes. The risk isn't the access. It's the alignment that never got set up.
The fix: talk it through before you give it more. It's a two-way conversation, not a switch.
- Say what you want to give it, and why.
- Ask it what could go wrong with that access, and how bad each thing would be. Let it tell you what you didn't know to ask.
- Ask it what controls the tool you're using gives you. Most tools have them. Claude Code, for example, can ask before every action, accept edits on its own, or run on auto and only stop for you at risky steps.
- Agree the lines together: what it does on its own, what it checks with you first (paying, sending as you, deleting, anything that can't be undone), and what it never does.
- Ask it to tell you back what it understood. Correct it until it matches what you meant. That's re-alignment, done before the work instead of after a reset.
- Write the lines down somewhere it reads every time, so the next session starts aligned too.
- Then give it the access. Watch the first few runs. You can stop it at any time.
Why first: watching and stopping still matter, but they only catch a mistake after it happens. Agreeing the lines first stops most mistakes from happening at all. That's where the value is.
As a skill: ask the AI to turn this page into a skill, and run it whenever you're about to give it more.