AI and MCP

Prompts that make an AI assistant show its working

Prompt additions such as list the rows used, show subtotals first and say what you could not classify, with before-and-after examples of clearer output.

Editorial illustration for Prompts that make an AI assistant show its working

To make an AI assistant show its working, add four requests to your prompt: list the rows you used, show subtotals before totals, say what you could not classify, and state how many rows you read. These turn a confident-sounding answer into one you can check line by line, which is how you find mistakes before they cost you anything.

Why does showing the working matter?

When an assistant only gives a final number, you can only trust it or not. When it shows the rows, the subtotals and the exceptions, you can test each piece against your statement. Mistakes then show up as a visible mismatch instead of a wrong figure that looks right.

Which phrases should you add?

  1. 'Before answering, tell me how many rows you read.'
  2. 'List the rows you included in each category, with date and amount.'
  3. 'Show subtotals for each category first, then the grand total last, with the addition written out.'
  4. 'Tell me which rows you could not classify and why, instead of putting them in Other.'
  5. 'If any figure is an estimate or an assumption, label it that way.'
  6. 'Do not round. Give amounts exactly as they appear.'

What is the difference in practice?

Here is the same question with and without those additions, using invented figures.

PromptWhat comes back
'How much did I spend on travel in March?''You spent about ₹8,500 on travel.'
Same, plus the working phrases'I read 62 rows. Travel rows: 03-03 cab ₹240; 11-03 fuel ₹2,500; 19-03 train ₹1,860; 25-03 fuel ₹2,400. Subtotal: 240 + 2,500 + 1,860 + 2,400 = ₹7,000. Could not classify: 14-03 UPI to a named person, ₹1,500.'

The second answer is better for two reasons. It gives ₹7,000 and not an approximation. It also tells you that ₹1,500 is undecided, so ₹8,500 in the first answer may have included something you would not call travel.

How do you use the working to find errors?

  1. Check the row count against your sheet.
  2. Compare each listed row with your statement; any row you cannot find is invented.
  3. Add the subtotal yourself. If your sum differs, the assistant slipped in the addition.
  4. Read the unclassified list and decide those lines yourself.

Do not use the working phrases on everything. A quick question about whether a merchant looks familiar needs none. Use them when the answer is a number you will act on: a total, a count, a monthly average, a percentage change.

You can also ask for a verification pass as a second step in the same chat: 'Now check your own answer. Recount the rows, re-add each subtotal and tell me whether anything changed.' It is not proof, since the same assistant can repeat the same mistake, but it catches slips that were random, and it is cheap. If the second pass gives a different number, you know which figure to test by hand.

What if the answer becomes too long?

Long answers hide errors as well. Ask for the working as a compact table, or ask for the rows only for the categories you care about. For very long statements, split by month and ask the same questions month by month, then combine the results yourself in a sheet.

Keep a small library of these phrases in a note and paste them at the end of every number-based prompt. After a few weeks you will find the answers arrive in the form you need, and you will spot a sloppy answer at a glance because it lacks the rows.

Does the app do some of this for you?

A tracker's report shows the transactions behind a number, which is the same idea built in. SpendixAI, for instance, lets you filter and group in the Report builder and download the result as CSV, so a figure can be traced back to its rows. Use that as a reference when an assistant gives you a summary, and trust the figure that has its rows listed.

SPENDIXAI JOURNAL

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

Explore your reports