Prompt Tips I Learned While Cleaning Financial Data with GPT (feat. Field Notes)
I've been putting GPT to heavy use on an in-house financial data analysis project for the past three months. At first it was total trial by fire, so I wanted to share that experience. Since it's confidential company data, I've changed some of the details.
1. Just assigning a role isn't enough. For example, instead of saying "Act as an AI financial expert," giving context like "You're a practitioner putting together quarterly reports. The input data is an intermediate output" makes results far more stable.
2. Define the output format first. If you just say "analyze this," the model responds with a long-winded natural language answer. Presenting a JSON schema or a Markdown table first and asking the model to fill in that framework cut parsing costs in half.
3. Don't ask for corrections all at once. Instead, collect about five wrong cases and say, "Recalculate using these rules" — that dramatically lowers the recurrence rate.
Number 3 helped the most. Now my review time has dropped by nearly half. If anyone has had similar experiences, please share them in the comments.
9 answers
Oh, number 3 is really clever. I used to have them fix each wrong answer right away, but I never thought of bundling them together and handling them all at once.
Yeah, #2 makes the biggest difference. If you don't specify the output format, GPT just writes prose however it wants, and parsing becomes a total nightmare.
Well, the results vary too much depending on the GPT version and model, so those kinds of tips don't really mean much. Financial data ultimately has to be reviewed by a human, doesn't it? Trusting AI too much might actually be more dangerous.
Thanks for sharing your experience. Especially for #3, I think I understand why it's effective—the wrong-answer cases themselves act as explicit feedback.
Got a source? If you've written up anything on a blog or GitHub, please share it.
I found that providing context (Option 1) made the most noticeable difference. When you only assign a role, GPT suddenly starts acting like a dramatic AI expert and blurts out unnecessary explanations.
Right lol. I also tried all sorts of things to solve it with prompts before fine-tuning, and #3 definitely helped prevent it from recurring.
Honestly, I have some doubts about #3. Collecting 5 errors seems to add more management overhead than you'd think. Wouldn't it be faster to just request fixes all at once? Though maybe that's just because the data scale is small.
Giving the JSON schema upfront is really nice. This alone drastically cut down the need to fix parsing code. I just learned about this tip too, lol.