Tell it what normal looks like.

The model has never seen your week. It can't spot what's off if it doesn't know what's usual.

If you want to know how to get AI to analyze your data and actually catch the problem, tell it what normal looks like first. A number only means something next to the number it's supposed to be.

Prompt Fluency is communicating so the model doesn't have to guess. "Normal" is one of the biggest guesses you hand it without noticing. You know your business. You know a slow Tuesday from a bad Tuesday. You know which spike is payday and which spike is trouble. That knowledge lives in your head, not in the spreadsheet. So when you paste the numbers and ask "anything look off?", the model has to invent its own idea of normal. And its idea is an average of a million places that aren't yours.

Try this ask:

"Here's last month's machine log for the laundromat. Anything look wrong?"

You get a tidy answer. Machine 7 ran fewer loads than the others, so maybe check it. Weekends were busier than weekdays. Revenue dipped mid-month. All true. None of it helpful. Machine 7 is the big front loader that only gets used for comforters, so of course it runs less. Weekends are always busier. And mid-month is when the college kids go home. The one thing that was actually wrong, a dryer that ran 40 cycles but only took money for 22, got buried in the list because the model didn't know that was the strange part.

The model didn't miss it because it's dumb. It missed it because it had no baseline. Everything looked a little unusual, so nothing looked very unusual.

Same break shows up everywhere.

"Look over this month's roofing bids and tell me what stands out."

"Here are the library's late returns. Is this bad?"

"These are the swim team's times from Saturday. How'd we do?"

Bids. Late returns. Lap times. Each one is a pile of numbers that can't judge itself. Is 30 late returns a lot? For a branch that checks out 2,000 books a week, no. For a branch that checks out 200, something happened. The model can't know which branch you are unless you say.

Give it the normal. Plain English. Right in the prompt.

"Here's last month's machine log for the laundromat. Normal for us: weekends run about double weekdays, machine 7 is the comforter washer and only does a few loads a day, and revenue always drops the second half of the month when the college is out. Every machine's paid cycles should match its total cycles within one or two. Flag anything that breaks those patterns and ignore anything that fits them."

"Look over this month's roofing bids. Our usual: a standard shingle tear off runs us about 3 to 4 days for a crew of five, and we win roughly one bid out of three. Tell me which bids look underpriced or overpriced against that, and tell me if the win rate moved."

"Here are the library's late returns. Normal is about 15 a week at this branch, mostly kids' books, mostly back within 10 days. Point out anything outside that."

"These are the swim team's times from Saturday. I've pasted each swimmer's best time from last season next to their name. Tell me who dropped time, who added more than two seconds, and leave the rest alone."

Look at what changed. The answers got shorter. The model stopped reporting things you already knew and started pointing at the thing you'd want to hear about. You didn't make it smarter. You gave it a ruler.

That's the part that clicks. Asking "what's wrong?" without a baseline is like asking a stranger if your truck sounds funny. They've never heard it run. You have. Tell them how it usually sounds and now they can actually help.

No shame here. I do this too. When I've stared at the same numbers every week, normal feels so obvious I forget it's not written anywhere. The model only has the words. If normal isn't in the words, it makes one up.

A quick check before you hit send:

What "normal" can look like in a prompt

• The usual range (about 15 a week, 3 to 4 days, 1 in 3)
• Patterns that are expected and should be ignored (weekends busier, mid-month dip)
• Things that should always match (paid cycles and total cycles, invoices and deposits)
• Last period's numbers pasted right next to this period's
• What you actually want flagged, and what you want left alone

When an analysis comes back as a long list of obvious stuff, ask yourself: did I tell it what normal looks like? Odds are it couldn't tell the strange thing from the usual things, so it listed all of them.

You're not doing the analysis for it. You're removing the need to guess. That's Applied Intelligence Mastery in practice: tell it what normal looks like, so it can see what isn't.