AI Guides › Judgement & Guardrails

The Guardrail Habit Worth Building Before Any Other On This List

By Nigel Guy · 3 min read

Reading a library of guardrail guides is a satisfying way to feel prepared without actually changing what you do. Every mechanism in it makes sense on its own, and the honest failure mode isn't disagreeing with any of them — it's intending to apply all of them and, in practice, applying none of them consistently, because nothing tells you which one to reach for and when.

The mechanism: the stakes pause

1. Ask it before you act on anything AI-assisted, as a fixed habit, not a judgement call about whether this particular case seems risky. The whole value of the pause is that it runs even when nothing feels risky yet.
2. Answer honestly: what happens if this is wrong? Not much, in most day-to-day cases. Something real — money, a relationship, a decision already partly acted on, someone else's time or trust — in the cases that actually matter.
3. Answer the second half: would you know it was wrong before it caused the damage, or only after? A mistake you'd catch on the next read-through is a different category from one that ships and does its damage quietly.
4. If both answers are mild, proceed at normal speed. The pause isn't meant to slow down everything equally — most AI-assisted work genuinely doesn't need the full toolkit, and pretending otherwise is how people stop doing the pause at all.
5. If either answer is concerning, that's the trigger for the specific mechanism the situation calls for — trace the claim to a source, run the two-minute check on the number, treat the action as needing a real human checkpoint, read what permission is actually being granted. The stakes pause doesn't replace those mechanisms; it's the filter that tells you when one of them is actually needed.

What to skip

Skip trying to apply every named mechanism in this library to every AI-assisted task — that's not diligence, it's a plan that collapses under its own weight within a week. The stakes pause is deliberately the only habit meant to run on everything; everything else in this library is meant to be reached for, not run by default.

Guardrails

• This is a triage step, not a verification step in itself — answering the two questions honestly doesn't verify anything on its own. It decides whether verification is warranted, and if it is, you still have to actually do it.
• The honesty of your own answers is the whole mechanism, and there's no external check on that honesty — if you're inclined to answer "not much" reflexively because stopping is inconvenient, the pause isn't doing its job, and that's worth noticing in yourself specifically.
• This doesn't replace judgement that comes from experience in a specific domain — what counts as high-stakes in your particular line of work is something you'll calibrate over time, and this is the habit that keeps that calibration from lapsing, not a substitute for building it.

All 751 AI guides · JulieMango plans from £17/mo