The problem
Chatbots are happy to sum a few hundred bank rows in their head, and they get it wrong in quiet ways: the card payment counted twice (once from the bank, once on the card), a transfer to savings called spending, a refund added instead of subtracted, a friend paying you back called income. The totals look plausible, so nobody checks.
Why it works
Naming merchants and putting them in categories is a reading job, which a model does well. Adding up is not. So the prompt has the model look at each distinct merchant once and give it a category, and then code applies that table and does every sum. It also has to prove three checks: every row is in one group, the categories add up to total spending, and the rows balance against income, transfers and spending. In our test, those checks caught the refund and card-payment mistakes that earlier versions of the prompt made.
Step by step
- Download last month's (and the month before's) transactions as CSV from your bank and card websites. Delete account numbers if the export includes them.
- Open a chatbot that can run code (ChatGPT, Claude or Grok) and attach the files.
- Paste the Kickstart prompt and fill in the months.
- Check the row counts and date ranges against your statements first.
- Look at the merchant-to-category table and fix anything wrong, then ask it to re-run.
- Read the flags (double charges, new subscriptions, fees) and answer its five questions.
Pitfalls to avoid
- Never give a chatbot or app your bank login; an exported CSV is enough.
- If the chatbot can't run code, use the script it gives you, or treat its totals as rough.
- Card payments and transfers between your own accounts are not spending; check they were set aside.
- Refunds should lower a category, not raise it.
- Splitting rent or bills with someone: their payments are reimbursements, not income.
Free · Kickstart prompt
Get a rough first version today
Paste this into Claude, ChatGPT or Grok with your own data. It gets you a useful first pass, not the tested setup.
Act as my monthly money reviewer. Attached are my bank and card CSV exports for [MONTHS] (I removed account numbers). Use code for every number: if you can run code, write and run a short script over the files and don't add up rows in your head. If your code errors, fix it and run it again until it works. If you can't run code at all, say so and give me the script to run instead of guessing totals. 1. Load every file. Show the row count and date range per file so I can check nothing was dropped. Check the signs: in most exports money out is negative and money in is positive; say if a file uses the opposite and flip it. 2. Set aside rows that are not spending or income, and list them: the credit card payment (it shows in both files, as a payment from the bank account and as a "payment" line on the card: set aside both and count the card purchases instead) and transfers between my own accounts (savings). Money people paid me back is not set aside: show it as reimbursements, not income. 3. Any other money in on the card is a refund or return: put it in the same category as the purchase, so it makes that category's spending smaller. List each one. 4. Categorize by merchant, not row by row: have code print the list of distinct merchant names, then you decide one category for each merchant (Housing, Utilities & phone, Groceries, Dining & coffee, Transport, Shopping, Subscriptions, Health, Travel, Cash, Fees & interest, Income, Other), show me that merchant-to-category table, and have code apply it to every row. Use Other when you're not sure and list those merchants; don't guess. 5. For each month: income, reimbursements, spending after refunds (as a positive number), what's left over (income + reimbursements - spending; transfers to savings are still part of what's left over) and that as a % of income, and a table of spending by category with the change from the month before. 6. Flag: possible double charges (same merchant and amount within 3 days), merchants that are new this month and look like a subscription or membership, any fees or interest, and the 5 largest purchases (not rent, transfers or the card payment). 7. Check your work in code and show it: (a) every row is in exactly one group; (b) for each month, the categories add up to total spending; (c) for each month, the sum of all rows equals set-aside rows + income + reimbursements - spending after refunds. If a check fails, find the row that causes it and fix it before you answer. 8. Give me a downloadable CSV of every row with the category or set-aside reason you gave it, so I can correct it. 9. End with 5 short questions for me about the month (for example "Dining was up $X: was that planned?"). No investment advice and no lecturing.
✓ Prompt tested 15 Oct 2026: we ran it with gpt-5-mini on a synthetic sample and checked the output by hand.
This one is a free idea, not a paid Blueprint. If you want us to build and test the full version, tell us below.
Should we build this next?
One tap moves this up our build queue. Add an email only if you want a note when the tested Blueprint is ready.