AI and MCP

Claude prompts to categorise a list of bank transactions

Prompts that give Claude a fixed category list, ask for a table with a confidence column and flag uncertain lines instead of guessing, with an example.

Editorial illustration for Claude prompts to categorise a list of bank transactions

To get Claude to categorise bank transactions reliably, give it a fixed list of categories, tell it to use only those names, ask for a table with a confidence column, and instruct it to mark anything uncertain as Review instead of guessing. Then check the Review and Low rows yourself. Here is a prompt you can adapt and an example on ten lines.

What does the prompt need to say?

Three things make the difference: a closed list, a fixed output shape, and permission not to know. A prompt that says simply to categorise these will invent categories and stay confident about doubtful lines.

The base prompt

"You will categorise bank transactions from India. Use only these categories: Groceries, Milk and vegetables, Eating out, Fuel and FASTag, Medical, Rent, Utilities, Mobile and internet, Cash withdrawal, Transfer, Other. Return a table with columns Date, Narration, Amount, Category, Confidence (High, Medium or Low), and Note. If the narration does not make the merchant clear, use Category Review and Confidence Low, and say why in the Note. Never create new categories. Do not guess."

What does a ten-line before and after look like?

Narration (made up)CategoryConfidenceNote
UPI/FRESHMART GROCERSGroceriesHighMerchant name is clear
UPI/RAMU MILK DAIRYMilk and vegetablesHighDairy
POS PETROL PUMP KOLHAPURFuel and FASTagHighFuel station
FASTAG RECHARGEFuel and FASTagHighToll account
UPI/DR MEHTA CLINICMedicalMediumClinic, could be consultation or pharmacy
NEFT/RENT NOVRentHighText says rent
ATM CASH WDLCash withdrawalHighCash
UPI/9XXXXXXX21ReviewLowOnly a number, no merchant
IMPS/KIRAN STransferLowPerson's name; purpose unknown
BILLPAY*CITYPOWERUtilitiesMediumLooks like electricity, confirm

The two rows that matter are the Low ones. The assistant has no way of knowing what a bare number or a person's name was for. Only you do.

How do I make the output useful downstream?

  1. Ask for the table as plain rows you can paste into a spreadsheet, one line per transaction, in the same order as the input.
  2. Add: "The number of rows in your table must equal the number of lines I pasted. State both numbers at the end."
  3. Ask for a count by category and compare it with a pivot table or SUMIFS in your sheet.
  4. Paste your corrections back and say: "Remember these rules for the rest of this chat: FRESHMART always means Groceries."

Work in batches of thirty to fifty lines. Very long lists make it more likely that a row is skipped, and a row count check is quick to do when a batch is short.

When does AI categorising go wrong?

  • Ambiguous merchants, such as a shop that sells both groceries and household goods.
  • Person-to-person UPI payments, where the same name could be rent, a loan repayment or splitting a meal.
  • Merchants whose billing name differs from the shop name.
  • Amounts used as a clue: a ₹40,000 payment to a clinic is different from ₹400.

Treat the output as a draft that saves typing, not as a record. Your corrections are the real value, because they tell you which of your own rules to write down.

Can an app do this without a chat window?

SpendixAI categorises in stages: your own rules first, then your history and similar merchants, and only if AI is on, a model for what is left, using a normalised merchant name and an amount band, not your name, account numbers or balances. You can switch AI off for a feature, set a monthly AI budget, or have unmatched items held for review. And if you do prefer to work in Claude, it can be connected through MCP, where you choose the scope, can start read-only, and can disconnect any time.

SPENDIXAI JOURNAL

Useful ideas start with a clearer picture of your own spending.

Explore your reports