← All use cases

中文

Finance & bookkeeping · Claude Code

Build an invoice checker that flags duplicates and odd price jumps

Idea + starter prompt, not run by us yet

For: Small business owners who pay many supplier invoices and miss errors

A non-developer built a small invoice app by prompting. Claude reads each uploaded invoice into a database, and simple rules then watch for big amounts, upcoming due dates and repeats. Tested on 130 made-up invoices, it turned up four anomalies. Requiring three past invoices per supplier cut false alarms from 40% to about 15%.

Steps

  1. Collect a batch of sample invoices. Use made-up or old ones first, never live supplier data.
  2. Ask Claude to extract the fields into a table and show you a few rows to compare with the originals.
  3. Add rules for duplicates, big amounts and sudden price rises. Require a few past invoices per supplier before it flags a price change.
  4. Add a human review step. Extraction was around 95% on clean files but only 60 to 70% on handwritten or poor scans.
  5. Spot-check every flagged invoice yourself before you contact a supplier.

Starter prompt

Starter prompt. Change the [brackets] first.

Trial first: make [3] made-up invoices, with one duplicate and one big price jump, in a new folder called [invoice-trial]. Run the whole tool on those only and write the output to [trial-output]. Do not open my real invoice folder yet.

I want a small local tool that reads invoice files and puts supplier, date, invoice number, amount and due date into a table. It flags possible duplicates, unusually large invoices and sudden price jumps for the same item. Show me the plan first. Give each extracted field a confidence rating of low, medium or high. Do not flag a price change until a supplier has at least [3] past invoices.

Stop and wait for my yes before reading the real folder [folder]. The tool only reads and lists. It never emails or pays anyone, and a flag is for me to review, not to act on.

Original post: Medium · found via agentusecases.ai. Written up in our own words. We haven't run it, so test it on made-up data first.