Section 1 — Why this recipe exists

You're three stores deep. Monday morning. Last week's Toast or Square or whatever-POS export landed in your inbox at 6am. You stare at the seven columns and twenty-one rows and your second coffee hasn't kicked in yet. Two years ago you'd open Excel, pivot the table, sort by week-over-week variance, and start dialing GMs. Today you don't have to. Claude can read the same export in 15 seconds and tell you the three things you'd have found in 25 minutes of Excel, in plain English, with store names. The recipe below is exactly how. Tested in our own kitchen, used by Expo customers running 8 to 70 locations.

Section 2 — What you'll need

Total time: 18 minutes

You'll need:

  • Last week's sales-by-store report. PDF, CSV, or screenshot all work.
  • 18 quiet minutes.

Tool:

  • Claude desktop app (free) or ChatGPT desktop app (free tier works for files under 10MB).

Section 3 — The recipe

Step 1. Open Claude. Drag your sales report PDF directly into the chat window. You'll see the file icon appear in the message bar — that confirms Claude has it.

Step 2. Paste this prompt exactly:

I'm a restaurant operator with 3 stores. This is last week's sales report. Tell me three things — worst store and why, best store and what's worth studying, and which numbers look suspicious enough to warrant a phone call. Use store names. Be specific.

If you have 5 stores, change "3" to "5". If you have 25 stores, change it to 25 and add: Focus on the bottom-quartile stores by week-over-week variance. Same logic; different cohort.

Step 3. Wait 15 seconds. Read what comes back. Pick the one finding you care most about and ask Claude to dig deeper on just that one. Don't ask Claude about all three — you'll get diluted answers. Pick the one that would matter most if it turned out to be true and follow up only on that one.

Section 4 — What to expect

A good response names your stores by name and points to specific numbers. "Westgate was down 12% week-over-week on dinner sales; Riverside was up 8% on lunch with the same staffing — worth a call to the Riverside GM to find out what changed." That's the shape.

A bad response talks in generalities. "Sales performance shows variance across locations; recommend investigating underperforming sites." If you get this, your report didn't have enough detail. Re-upload with day-of-week broken out, or with a daypart split, and re-prompt.

The most common failure is Claude inventing a store name that isn't yours. It happens when your report uses store codes (Store 4847) instead of names. Fix: tell Claude the real names in a follow-up message. It'll re-read with the correction.

Section 5 — FAQ

What if my Toast report comes out as a screenshot instead of a CSV?

Claude reads screenshots fine. Drag the screenshot in and it'll OCR the numbers itself. CSV is cleaner — fewer transcription errors — but screenshots work.

Can I use ChatGPT instead of Claude for this?

Yes. Both tools handle the file-upload-plus-prompt pattern identically. ChatGPT's free tier limits file size; if your report is over 10MB, use Claude or upgrade ChatGPT.

What if Claude gets a store name wrong?

Tell it the correct name and ask it to re-read with the right names. Don't restart the chat — Claude remembers the correction for the rest of the conversation.

Section 6 — Where this goes if you want to keep going

This is one move from the Expo recipe library — short, specific, do-this-on-Saturday guides for multi-unit operators learning to use AI without anyone selling them something. The longer story behind why this works lives in Start Using AI in Your Restaurant This Weekend. Next week's recipe extends the same chat window into a Monday GM email draft. If you want every Saturday recipe as it ships, the Learn page is the spine.