Are you using Responses API or Chat Completions API?
When you say “clean data” what do you mean?
GPT’s have trouble counting. What model are you using? You say “a GPT I’ve created” what does that mean? You mean you saved a prompt in custom GPT, but are you accessing on Webchat app or using the API?
Usually you just say,
I’m going to send you an entire file, perform this process, and return the entire file to me. Ensure that you include 100% of all existing file data, performing only the necessary modifications, and otherwise not removing/deleting any data."
Also, sometimes GPT is better at doing file modifications (especially with large data set) as a providing only-the-set-of-changes. It’s pretty bad even with code files (which look more like natural language than excel files/database sets) at sometimes saying things like "and here, do that other huge thing we talked about doing, I didn’t do it!
The best success I’ve had is either with custom formatting or standardized script interfaces (“patch” utility scripts (i.e. diff files), “ed scripts”), or using other kinds of structured output to essentially say “change cell X to content Y, overwrite cell Z with content X,”, etc. And then you use your middleware to process the response and actually perform the operation.
Rarely will you succeed with getting GPT to provide you back full datasets/large docs with all content intact and only modifications performed. But it will do well if you have a format by which it can provide only the necessary modifications.
Also if you are doing “data cleanup” on an excel file, you most likely should just use native excel scripts/etc. to perform data migrations/formatting changes/“clean up”.
GPT could help you generate those scripts if that’s the kind of cleanup you are doing (i.e. like removing stray whitespace characters, reformatting, doing regular expression extraction, etc.). These are normal coding tasks - NOT gpt tasks (but GPT can give you the code!).