How to Catch AI Mistakes Before You Send Them

Checking AI output takes about ninety seconds and it is a different job from writing a better prompt. A better prompt reduces how often the model is wrong. A check catches the times it is wrong anyway, which is the part that reaches your customer with your name on it.

The mistakes that hurt are almost never the obvious ones. Obvious mistakes get caught by reading. The ones that get sent are the confident, well-formatted, completely plausible sentences that happen to be false.

Why do AI mistakes slip past a normal proofread?

Because you are proofreading for the wrong thing.

A normal proofread looks for typos, awkward phrasing, and tone. AI output is clean on all three. It reads well, it flows, and it is formatted better than most first drafts a person writes. So your eye slides over it, and the part your eye is not checking is whether any of it is true.

There is a second reason, and it is uncomfortable. When something arrives already written, you edit it. You do not interrogate it. Editing is a lower-effort mode than verifying, and the polished draft invites the lower-effort mode.

What is the fastest way to check AI output?

Ask the model to grade its own work before you read it. Three moves, in order.

1. Ask which parts it is least sure about. "Which two or three claims here are you least confident in?" It will usually tell you, and that list is your proofread. You are checking three lines instead of twelve paragraphs.
2. Ask where each fact came from. "For every number or named fact above, say where it came from, or say you do not know." Anything that comes back as "commonly reported" or "general knowledge" is a claim with no source, and it does not go out.
3. Ask it to argue the other side. "What would someone who disagrees with this say?" If the counterargument is stronger than the draft, the draft was never finished.

What you are checking	The move	What a bad answer looks like
Confidence	"Which claims are you least sure about?"	"I am confident in all of it."
Sourcing	"Where did each number come from?"	"This is widely known."
Blind spots	"What would someone who disagrees say?"	A weak counterargument it clearly invented to lose

Which mistakes should you always check by hand?

Some things never get delegated to the check above, because the model has no way to know it is wrong.

Anything with a number a customer will act on. Prices, dates, quantities, hours, deadlines. Read those against the source yourself, every time.

Anything about a specific person or company. Names, titles, who said what, who works where. This is where confident invention is most common and most embarrassing.

Anything that is a promise. If a sentence commits you to a timeline, a refund, a scope, or a result, a person approves it. Not because the model is careless, but because the model is not the one who has to keep the promise.

Anything you cannot verify in under a minute. That is not a rule about AI. That is a rule about sending.

What does a pre-send check actually look like?

Ninety seconds, four questions, out loud if that helps.

1. Is there a number in this? Did I check it against the source, not against my memory of the source?
2. Is there a name in this? Do I know that person or company exists and that the detail is right?
3. Does this promise anything? Am I willing to be held to it exactly as written?
4. If one sentence in here is wrong, which one is it most likely to be? Go read that one again.

That fourth question does most of the work. You almost always know. The check is giving yourself permission to act on the thing you already suspected.

Does this mean AI output cannot be trusted?

It means it should be treated like a draft from a fast, well-read colleague who has never met your customer and does not know what they do not know. You would not forward that person's first draft to a client without reading it. That is the whole standard, and it is not a high one.

The people who get burned are not the ones who use it heavily. They are the ones who stopped reading because the last forty drafts were fine.

FAQ

Will asking it to check itself actually catch anything? It catches uncertainty reliably and factual errors inconsistently. Use it to narrow where you look, not as the check itself.

Is there a tool that does this automatically? Tools exist that flag unsourced claims. None of them know which promises you are willing to keep, so the judgment step stays yours.

Does a longer prompt fix this? It reduces how often you need the check. It does not remove the need for it.

What is the single highest-value habit here? Never send a number you have not read against its source. That one rule prevents most of the damage.

How long before this stops feeling slow? About a week. It compresses to a glance once you know which two questions matter for the kind of work you send.

---

The pre-send checklist from this post, plus 25 prompts I actually use and the template they are built on, is free in the Prompt Starter Pack. If you want the upstream fix as well, the ask-me-questions technique cuts down how many mistakes get generated in the first place.