A good expense analysis prompt has five parts: the goal, the data, the categories, the output format and the checks. Leave out any one and the assistant fills the gap with a guess. You do not need a job title for the assistant or a clever trick, just those five plain pieces in this order.
The five parts in a sentence each
| Part | What to write | What it prevents |
|---|---|---|
| Goal | The one question you want answered | Rambling summaries |
| Data | Which rows, which dates, which currency | Answers about months you did not mean |
| Categories | The exact list to use, and what to do with misfits | Invented category names |
| Output format | Table or list, columns, how many lines of commentary | Walls of text |
| Checks | What it must verify and what it must say if it cannot | Confident wrong totals |
A weak prompt, and what usually comes back
The assistant has no rows and no list of categories, and has no idea what "save" means for you. Typical results are generic advice about cooking at home, cancelling subscriptions and avoiding impulse buys, none of it connected to your month. Some assistants also invent plausible numbers to look helpful.
The same request, rewritten with all five parts
How different is the result, in practice?
Compare what you get with a made-up example. Suppose September dining was ₹3,800 across 9 rows and October dining was ₹5,600 across 14 rows. The strong prompt produces a line showing a difference of ₹1,800 and five extra rows, which you can open and inspect. The weak prompt produces "reduce dining out". One is something you can verify, the other is something you can only nod at.
Which of the five parts do people skip most?
Checks. Without them you cannot tell a careful answer from a confident one. The two cheapest checks are asking for the row count beside every total, and asking for the grand total of all rows. Add both and then compare them with your own filtered view or report.
- Filter your own data for the same period.
- Compare the grand total with the assistant's.
- Compare the row count for one category.
- If either differs, ask which rows it left out.
How long should the prompt be?
Five parts do not need five paragraphs. A line each is plenty. Longer prompts are fine when you have special rules, such as "refunds reduce the category they came from" or "ignore credit card bill payments, since the individual purchases are already in the list". If you find yourself repeating such rules every month, move them to standing instructions or a project so they apply automatically.
Can you reuse the structure for other questions?
Yes. The five parts work for any numbers question.
- Subscription review: goal is to list every repeating payment with its monthly cost.
- Festival spending: goal is to total spending between two dates and group it by category.
- Cash versus UPI: goal is to compare the count and total of two payment types.
If the rows come from SpendixAI, a report or a CSV download gives you clean date, merchant, amount and category columns to paste under the data heading, and the same report is your cross-check afterwards.
Before you press send
- Is the question one sentence?
- Do the rows cover exactly the dates you mean?
- Did you name the categories?
- Did you say how the answer should look?
- Did you ask for row counts or a grand total?
Useful ideas start with a clearer picture of your own spending.
Explore your reports

