# The AI put 48 candles in my register. Saturday's market showed 71. > An AI agent loaded a candle shop's product sheet into Square twice and left old prices on the register. We tested a How I AI workflow that makes it show a dry run first. By Jordan Menlo · Published October 23, 2026 Section: Markets & Main Street › Small Business · Series: AI Prompts, Reviewed (https://mcafax.com/notebook/prompts) Canonical: https://mcafax.com/notebook/ai-prompt-review-product-catalog Source: MCAFax, The Broker's Notebook (https://mcafax.com/notebook). MCAFax is not a lender or a law firm. --- "Done! I've created your items in Square." That's what Marisol, owner of Wrenfield Candle Co. (a made-up candle shop), got back after she handed an AI agent her product spreadsheet. Her sheet had 48 candles. At the Saturday market, her register showed 71 tiles. Some had the old price. Never let the AI touch your register until it shows you the plan. AI Prompts, Reviewed takes one prompt or workflow from How I AI, Claire Vo's show where guests share their screens and walk through how they really use AI. We run it on a small-business job, score it, and tell you what we'd change. Credit: Based on "How to Automatically Create an E-commerce Product Catalog from a CSV using an AI Agent," a workflow from How I AI (hosted by Claire Vo, written up by the ChatPRD team), shown by Brad Axen, an engineering leader at Block, in the episode "How I AI: Block's 3-Step Workflow for Data Analysis, E-commerce, and Custom AI Tools with Goose." Original workflow: https://www.chatprd.ai/how-i-ai/workflows/how-to-automatically-create-an-e-commerce-product-catalog-from-a-csv-using-an-ai-agent. Full episode: https://www.chatprd.ai/how-i-ai/blocks-goose-ai-for-data-analysis-e-commerce. Guest site: none listed on the How I AI workflow page. What the original does In the episode, Brad gives Goose, Block's open-source AI agent, a farm stand CSV and an empty Square catalog in a sandbox account. The written workflow has 4 steps: * Connect the Square MCP with only the catalog permissions the job needs, and confirm which account you're in. * Ask for a dry run first: map every column, list creates, updates, skips and errors, and write nothing yet. * Apply a small batch, then check names, prices, tax and categories against the sheet before doing the rest. * Make a payment link only for an item you've already checked. The line we'd tape to the monitor: "Do not write yet and never delete based only on absence." The test Everything here is made up. Wrenfield sells soy candles at weekend markets and takes cards on Square. Marisol's sheet: * Rows: 48 candles, each with a SKU, scent, size, price and category * Sizes: an 8 oz jar at $18 and a 16 oz jar at $32 * Large jars: 6 scents Run 1, the shortcut: "Read this spreadsheet and create my items in Square." The agent created 23 items, then the connection dropped. Marisol typed "keep going." Before she did, she'd raised the 16 oz jars from $32 to $36 in the sheet. The agent started over from row one and created all 48 again. That's 23 + 48 = 71 items. The first 23 were now in the register twice. Of the 6 large jars, 4 were in that first batch. Each had a $32 tile and a $36 tile, side by side. At Saturday's market her helper tapped the first tile she saw. She sold 40 large jars at $32. At $4 short each, that's $160 gone in one day. Marisol spent the evening deleting 23 duplicates by hand. The rerun Run 2: same agent, a fresh sandbox, and our prompt below. It matches on SKU and writes nothing until she says go. The dry run came back as a table: 48 creates, 0 updates, and 3 flags. Three rows had a blank tax field, so she filled them in. She approved a batch of 5, spot-checked them, then ran the rest. The connection dropped again at item 31. This time the rerun matched SKUs: 31 unchanged, 17 creates. That's 31 + 17 = 48. No duplicates. When she raised the large jars, the dry run said 6 updates, 0 creates. Every 16 oz tile showed $36. Verdict: 7/10 Strengths: the dry run catches mistakes while they're still free. Matching on a stable ID like a SKU makes reruns safe. And the rule against deleting items that are missing from one export protects your live catalog. Weakness: it assumes you're at home with agents, MCP servers and sandbox accounts, and most shop owners aren't. The demo itself also used a one-line ask, the same shortcut that cost Marisol. Founder tweak: put a SKU on every row before you start, ask for the dry run as a table you can read in two minutes, and always test a batch of 5 first. Copy the MCAFax version This is our own rewrite, not the original. Copy the original from Brad Axen on How I AI: https://www.chatprd.ai/how-i-ai/workflows/how-to-automatically-create-an-e-commerce-product-catalog-from-a-csv-using-an-ai-agent ``` You are loading my product list into [Square, Shopify, or my POS] for [business name]. Account: [test or live]. Tell me which account you're connected to before anything else. Step 1. Read my sheet. Use the [SKU] column as the ID for every item. Map each column to a field: name, size or variant, price, category, tax. Show prices in dollars and cents exactly as the sheet has them. Step 2. Check what's already in the catalog. Match on SKU only, never on name. Step 3. Give me a dry run as a table: Creates, updates (old price and new price), unchanged, skipped, errors. Flag any row with a blank price, tax, or SKU. Write NOTHING yet. Step 4. Wait for my OK. Then create only the first 5 items and stop. Step 5. After I check those 5, do the rest. If anything fails, stop. On a rerun, start with Step 2 again so nothing is added twice. Never delete or hide an item just because it's missing from my sheet. ``` Paste into an AI agent that can connect to your store. No agent? Most store and POS systems have a spreadsheet import, and Step 3 still checks your sheet first. Before you load it 1. Give every row a SKU, and never reuse one. 2. Read the dry run's price column against your sheet. 3. Test a batch of 5 and ring one up as a sale you void. 4. Search the register for each product name after a rerun. Two tiles means a duplicate. 5. Keep an eye on what competitors charge for the same items: https://mcafax.com/notebook/prompts/competitor-price-monitoring More owner prompts: https://mcafax.com/notebook/prompts ## Questions readers ask **What tools does the original workflow use?** Goose, Block's open-source AI agent, connected to Square through the Square MCP, starting in a sandbox account. **Why did the agent create duplicates?** On the rerun it treated every row as new. Matching on SKU tells it which items already exist. **What is a dry run?** A preview of every planned change, with nothing written yet. You read it, fix the sheet, then approve. **Do I need an AI agent to do this?** No. Use your system's spreadsheet import, and run Steps 1 and 3 of our prompt to clean the sheet first. **Should I test in a sandbox account?** If your system offers one, yes. If not, load 5 items, check them, and ring one up before loading the rest. **Can the AI delete old products for me?** Only from a removal list you've reviewed. Never because an item is missing from a new sheet.