---
title: "Prompt patterns for better financial questions"
slug: prompt-patterns-for-financial-questions
url: https://bankbridge.money/guides/prompt-patterns-for-financial-questions
category: tip
published: 2026-04-23
updated: 2026-04-23
---
# Prompt patterns for better financial questions

> The six patterns: anchor the date range, specify the grouping, request the math explicitly, chain follow-ups instead of starting fresh, use 'compared to me' framing (not 'compared to others'), and close with 'what should I do?' to get recommendations. Apply them in any order to any question and you'll get cleaner answers from any MCP-speaking agent.

Prompt Patterns for Financial Questions | BankBridge | BankBridge 

[Tip](/guides/tips)

# Prompt patterns for better financial questions

Updated Apr 23, 2026·5 min read

The six patterns: anchor the date range, specify the grouping, request the math explicitly, chain follow-ups instead of starting fresh, use 'compared to me' framing (not 'compared to others'), and close with 'what should I do?' to get recommendations. Apply them in any order to any question and you'll get cleaner answers from any MCP-speaking agent.

## Why patterns help

Modern agents are good at guessing what you want. They're great at inferring a date range when you say “last month.” They're great at picking the right tool. What they're less good at is guessing the _shape_ of the output you wanted — a one-line answer vs a detailed table, a percent vs a dollar amount, a list vs a sum.

These six patterns make the shape explicit. You get answers that look like what you had in your head when you typed.

### 1\. Anchor the date range

Instead of “recently” or “lately,” say a window. “Last 30 days,” “year-to-date,” “Q1,” or “since my last paycheck” all work.

> “Year-to-date grocery spending.” (not ‘recently’)

### 2\. Specify the grouping

If you want a breakdown, say _by what_. By merchant, by category, by month, by account. Without a grouping, the agent guesses.

> “Grocery spending this year grouped by month.”

### 3\. Ask for the math explicitly

Want a percent? Ask for a percent. Want a ratio? Ask for a ratio. Want “what would it look like if I cut X by 50%”? The agent can do the math if you tell it to.

> “What percent of my spending is groceries? And what would my savings rate be if I cut groceries by 20%?”

### 4\. Chain follow-ups

Don't start fresh for each question. The agent's context from the first tool call helps answer the second. Start with the broad question, then drill in.

> Q1: “Monthly cashflow.”  
> Q2: “What drove expenses this month vs last?”  
> Q3: “Show me the specific transactions in the biggest mover category.”

### 5\. Compare-to-me framing

“Is my spending high?” is hard to answer — high compared to whom? Reframe as “is my spending high for me, based on my history?” The agent knows your history specifically; it doesn't know cohort data for anyone else.

> “Is my April grocery spend high for me, vs the last 6 months of April data?”

### 6\. End with “what should I do?”

The agent won't give you investment advice, but it will happily suggest actions based on the data. Asking for recommendations at the end of a data-heavy question often produces the most useful output.

> “List my subscriptions sorted by cost. Which ones should I consider cancelling based on how often they appeared in my charges the last 90 days?”

## Combine them

A question using all six patterns at once:

> “For the last 90 days, group my dining-out spending by merchant. Compute what percent of my total spending this is. Compare to my previous 90 days for me. What should I do if I want to cut dining out by a third?”

That's a well-posed question. The agent will chain three or four tool calls, do the math, give you a specific action, and leave you with something to act on.

## FAQ

Does the agent actually care about prompt engineering?

For BankBridge questions, modestly yes. The agent picks tools on its own, so you don't need to name them. What you DO need is clarity about time windows, grouping, and what you want the output to look like. Those three things change the tool parameters the agent passes.

Can I skip the patterns and just wing it?

Yes, and it usually works. The patterns are for when you care about precision: tax prep, reconciliation, fraud checks. For casual 'what did I spend on coffee?' kind of questions, any phrasing works.

## FAQ

### Does the agent actually care about prompt engineering?

For BankBridge questions, modestly yes. The agent picks tools on its own, so you don't need to name them. What you DO need is clarity about time windows, grouping, and what you want the output to look like. Those three things change the tool parameters the agent passes.

### Can I skip the patterns and just wing it?

Yes, and it usually works. The patterns are for when you care about precision: tax prep, reconciliation, fraud checks. For casual 'what did I spend on coffee?' kind of questions, any phrasing works.