Automating expense tracking with AI, Google Drive, and Google Sheets

Executive overview

Manually itemising receipts and logging expenses wastes hours each month. Dropping any receipt file into a watched folder triggers an automated pipeline that extracts, categorises, renames, and logs the expense — no manual entry required.

The core insight: a file-watcher plus a vision model plus a spreadsheet writer eliminates all manual expense work at the point of capture.

How the automation works

• Any image or PDF dropped into a Google Drive parent folder (or subfolder) triggers a watcher
• Gemini Flash 2.0 extracts date, merchant, amount, category, and purchase rationale from the receipt
• Rationale is inferred by the model from context clues (merchant type, time, location) — not read from the receipt
• Categories are supplied as a fixed list; the model picks the closest match
• The file is renamed to merchant_name + date format, replacing the default camera filename
• Extracted data is written as a row into the correct monthly tab in a Google Sheets expense tracker, with a link back to the Drive file

Building it with AI: three lessons

• Start a new Cursor Composer conversation after each micro-feature — context window degradation hurts output quality over time
• Reference live API docs when integrating services like Google Sheets; it resolves auth and insertion bugs faster than iterating blind
• For recurring errors, ask the AI to analyse root cause before suggesting fixes — prompt it to assess the full codebase and propose no fixes until the cause is agreed

Prompting strategy: PRD-first development

• Write an initial plain-English brief, then send it to a reasoning model (e.g. o3-mini) to generate a PRD (product requirements document)
• Structure the PRD as phases with numbered micro-steps, each with a checkbox — this keeps the AI navigating macro-to-micro without getting lost in detail
• Include explicit model versions, folder structure, sheet/tab naming conventions, and column schemas in the brief — AI coding tools have training cutoffs and will default to stale APIs otherwise
• Specify the runtime target (e.g. macOS launchd/plist for a persistent local process)

Test-driven development as a reinforcement loop

• For each micro-step, instruct the AI to write the unit test before writing the function
• A failing test feeds structured error output back into the next iteration — the model self-corrects toward a passing test
• This creates a reinforcing feedback loop that reduces accumulated errors across phases