For the first two years of running my consultancy, my monthly budget review process went like this: open the banking app, glance at the balance, feel vaguely fine, close the app. I owned QuickBooks the way some people own a gym membership — as a monument to good intentions and a small recurring charge. I knew roughly what I spent, but never by category, and never against a target, because I had never bothered to set one. My financial strategy was, in full, 'vibes, plus hope.'
The month revenue went up and my cash went down
In February I had a month where revenue climbed 18% and I still ended with less cash in the bank than January. That should be against the laws of physics. I had no theory for it, just a queasy feeling that something was leaking. So I finally did the thing I had avoided for twenty-four months: I exported the month's transactions to a CSV, hand-typed some rough budget targets for each category for the first time in my life, and pasted the whole lot into the Monthly Budget Review from Raw Transaction Data prompt, running it on GPT 5.2.
The category table did in 90 seconds what I avoided for two years
The breakdown table told me instantly what my gut could not. My software-subscriptions line had gone from $410 to $890 in a single month. GPT 5.2 flagged it as the single biggest absolute variance and, in the likely-driver note, pointed out that three new recurring charges had appeared mid-month. I opened my subscriptions list and found three tools I had trialled in January, told myself I would 'evaluate properly later,' and never cancelled. That was $340 a month of pure waste, discovered in roughly the time it takes to make a coffee.
What makes this work is that the prompt does not editorialize or pretend to be my financial advisor. The rules explicitly forbid it from inventing benchmarks or giving investment advice, so it never wanders off into generic personal-finance lectures. It just reflects my own numbers back at me, organized, with the three biggest swings sitting right on top and exactly three actions I can actually take this week.
The structure is the whole reason it lands. I had technically had all of this information sitting inside QuickBooks the entire time — the data was never the problem. The problem was that raw data in a software dashboard is a haystack, and what I needed was someone to hand me the three needles and a short note about what to do with each one. GPT 5.2 is strong enough at the reasoning that the 'likely driver' sentences are actually useful rather than generic, and disciplined enough about the rules that it never makes up a number to fill a gap. That combination is rarer than it sounds.
What I found, in dollars
Here is the full damage and recovery from that first run, because the specifics are the whole point.
- Software subscriptions: $340/month in forgotten free-trial-turned-paid tools, cancelled that same afternoon
- Travel: 22% under budget, because I had quietly deferred two client trips and forgotten I had
- Marketing: 40% over, because I paid an annual invoice that I had never broken into a monthly budget line
- Net swing from acting on the three recommended actions: roughly $520 the following month
The 'Open Questions' section earned its keep too, which surprised me because I assumed it would be filler. One month it flagged that I had $2,300 sitting in an uncategorized 'Miscellaneous' bucket and asked me to clarify it before it could interpret the month accurately. Those turned out to be three real expenses I had coded lazily under one label. Without the nudge, my reconciliation would have been quietly, confidently wrong, and I would never have known.
I think that 'Open Questions' habit is the most underrated part of the whole prompt, and it is the part that separates it from a dumb spreadsheet macro. A spreadsheet will happily compute a variance off a garbage category and present it to you with total confidence, false precision and all. The prompt instead refuses to pretend it understands a number it cannot actually interpret, and asks. That single behavior has caught messy data three times now, and each time it forced me to fix the underlying bookkeeping rather than building a tidy-looking analysis on top of a quietly broken foundation. An analyst who admits what they do not know is worth far more than one who fakes certainty, and that is true whether the analyst is a human or a model running on GPT 5.2.
The habit it turned into
I now run this on the first Monday of every month. The export takes five minutes, the prompt runs in about twenty seconds, and I start the new month actually knowing what happened in the last one — which, it turns out, is a wildly different feeling from 'vibes, plus hope.' When a category looks suspicious, I hand the raw lines straight to my Bulk Expense Categorisation Assistant prompt to clean them up before the next review. If you run anything with money flowing through it, grab the Monthly Budget Review from Raw Transaction Data prompt on Prompt Dock and run it on last month's transactions. You will find at least one forgotten subscription. I guarantee that the way I guarantee almost nothing else in life.