Claude to find recurring costs in bank exports

Mahmoud Halat used Claude Code to analyze his own bank and credit-card exports without connecting a finance aggregator to his accounts. The agent read the files, wrote a Python pipeline and generated a self-contained interactive HTML dashboard.

The pipeline normalizes different export layouts, reconciles statement totals and deduplicates overlapping downloads. Classification then separates actual spending from transfers between the owner's accounts and credit-card bill payments. That distinction prevents a payment moving money between accounts from being treated as a second purchase.

On his own data, Halat reports approximately $178 a month in active digital subscriptions and about $574 a month when all recurring bills were counted. His AMEX parsing reportedly matched the statement totals to the cent. These are identified costs and reconciliation results, not money saved through cancellations.

The dashboard includes recurring-charge run rates, category trends, counterparty histories and money-flow views. Charts are embedded in the HTML so the finished file works offline. However, the statement contents are read by Claude during generation; the article explicitly acknowledges that they go to Anthropic. Offline output does not mean the entire analysis avoids a model service.

Halat exports the input files manually. The public demonstration uses fictional data, so it cannot substantiate his personal transaction amounts independently. The published prompt excerpt captures the dashboard goal without repeating its broader privacy claims.

Prompt
Original prompt

Source: finance.mahmoudhalat.com · Mahmoud Halat