Skip to content
Guardrails

The whole method, in four questions

The four questions I run any AI answer through before I trust it, shown end to end on one everyday example, with a prompt you can copy. No jargon.

The Bluff Filter is this kind of check, on one page. Take it with you →

Ask an AI “is this a good deal?” and it will almost always tell you yes. The deal has nothing to do with it. Agreeable is its factory setting. It wants to be helpful, a cheerful yes feels helpful, so that’s what arrives, catch and all.

The fix is four questions, asked in order. Each one does a job the single eager question skips. Here’s the whole method, start to finish, on one ordinary thing: a holiday deal someone’s sent you and you half want to believe.

One: make it argue against the deal, not for it

I make this the first move on every answer that matters. Left to itself, the AI sells you the sunshine. So give it the opposite job first. Tell it to be the tight-fisted friend who assumes every deal hides a catch and finds that catch before saying a single nice thing. Do this and it starts hunting the resort fee, the “airport” that’s a two-hour coach ride from the resort, the “from” price nobody has ever paid.

Two: split what it knows from what it’s guessing

Of the four, this is the step I lean on most. Now make it sort its answer into two piles: what the page in front of you says, and what it’s filling in from thin air. “Breakfast included” belongs in the first pile only if the deal page says so. If it’s a confident guess, it goes in the second, where you can see it for what it is.

Three: ask what would prove it wrong

This is the question almost nobody asks, and the one I added last. It’s the one that changed how much I trust the answer. Make the AI name the exact thing that would expose the deal as a dud: a tripwire. Something like “if the total at checkout comes out over £900, the headline price was a lie.” Now you’ve got one number to watch, and the deal either trips it or it doesn’t.

Four: make it commit to one answer

Here’s where I stop it hedging. Last, no fence-sitting. Book it or don’t, in one line, with how sure it is: low, medium or high confidence. Nobody has ever stood at the checkout with their card out, reading a balanced six-paragraph essay on the merits of both sides, and felt helped. One verdict, one confidence level, and you can decide.

THE EAGER ASK
"Is this a good deal?" A cheerful yes, the infinity pool described in full, catch and all.
THE FOUR-QUESTION ASK
Catches the resort fee, the "airport" that's a two-hour coach ride from the resort, the "from" price nobody has ever paid. Lands a verdict with one tripwire: over £900 at checkout, the headline price was a lie.

The whole thing, ready to paste

Here’s the version I keep to hand. Drop your deal in and run this:

// Paste this with the deal you're weighing up

Here’s a deal I’m considering: [PASTE THE DETAILS]. Run it through these four steps, in order, and label each one.

SCOPE: Work only from the deal details I paste below. Don’t fill in anything the deal doesn’t actually say, and if something’s missing, tell me rather than guess. Then come at it like a tight-fisted friend who assumes every deal hides a catch, and find the catch before you say anything good.

FILTER: Make two lists: what the deal states, and what you’re assuming or guessing.

RISK: Name the single thing that would prove this is a bad deal: a specific number or fact I can go and check.

VERDICT: Give one answer, go for it or don’t, with a confidence level: low, medium or high.

That’s the whole thing. It works on the holiday, and it works exactly the same on a builder’s quote, a contract, or an email you’re not sure how to answer. Four questions, always in that order, because each one clears the ground for the next.

That’s the method, free to keep. If you want the longer version, where each step came from, and prompts you can lift for harder calls, it’s all written up in one place: the Prompt Stack.

Ben Dixon
// Written by Ben Dixon

Ben tests how far you can trust the main AI assistants, and publishes exactly where they get things wrong. Every post here is a first-hand test with the receipts, including the times a tool simply wasn’t worth the trust. About Ben →

// Keep reading
Guardrails

Does ChatGPT just agree with you? Mostly no, but watch the numbers

Does ChatGPT just agree with you? Mostly no. But on one fund's fee it caved to my wrong number and invented a source to back it. The 30-second check.

Guardrails

How to check if a ChatGPT citation is real or fake

Looking for a ChatGPT citation checker? The manual check beats one, and here is why. The dangerous citation is not the dead link, it is the working link to a real page that does not back the claim.

Guardrails

How to tell if an AI answer is true in 30 seconds

Two questions, and a prompt you can copy, that catch a confident, wrong AI answer before you act on it. No jargon, no AI knowledge needed.

// New here?

The site tests how far you can trust the main AI assistants, on real decisions. Start with the Prompt Stack for the four-stage framework, free and ungated, or the Bluff Filter for the paste-ready version with a real before and after.

← All posts More in Guardrails →