A small prompt pattern for checking bank statement conversions

I have been trying to make document-heavy finance work less fragile, and the most useful habit so far is treating the model as a reviewer, not as the first parser.

Bank statements are a good example. If you ask an LLM to read a messy PDF and immediately produce final bookkeeping rows, it may sound confident while quietly dropping page breaks, duplicated lines, or balance notes. I get better results when I split the workflow into two parts: first create a clean table, then ask the model to audit the table against a checklist.

For the extraction step, I usually want a structured file before I prompt the model. A dedicated Bank Statement Converter can turn statement PDFs or scans into CSV or Excel, which gives the AI reviewer something much more stable to inspect.

The review prompt I use is simple:

You are reviewing converted bank statement data before it is imported into bookkeeping software.

Check the rows for:
1. missing dates or repeated dates that look suspicious
2. deposits and withdrawals in the wrong column
3. negative numbers that should be positive amounts
4. descriptions that appear split across multiple rows
5. opening and closing balance inconsistencies
6. duplicate transactions created by page headers or footers

Return:
- a short risk summary
- the rows that need human review
- the likely reason each row was flagged
- a safe correction suggestion only when the evidence is clear

The key phrase is "safe correction suggestion only when the evidence is clear." Without that boundary, the model tends to fix too much. With it, the output becomes more like a review queue.

I also add one small rule when working with accounting data: the model should never invent a missing transaction. If a row is unclear, it should mark it for review and quote the surrounding rows that caused the concern.

This pattern has made the process calmer. The spreadsheet is still the source of truth, but the model becomes useful for catching boring mistakes before they travel downstream into reconciliation, tax prep, or monthly reporting.