I have tried to use AI to generate a code update of existing program as the data format in pdf changed.
(The program is a data extraction program for pdf file.)

In my first attempt, I fed the model with the new pdf, validation set and some instructions using github copilot with the existing program code. Expecting it gives me a completed runnable program with all my specified requirement. However, it gave me a similar version compare with existing one which is not favourable.

In my second attempt, I stay in the same chat as I have fed many requirement and I am lazy to reopen a new chat for doing the same file insert, I tried to adopt stepwise refinement as AI may often overlook some requirement. Surprisingly, it succeeded and asked for next tasks.

Seems like giving a brief background of the project then guide the AI to complete tasks step by step will be more efficient than asking for a whole solution in development. Texts below is my prompt for this chat.

Prompt:
"The current codebase you can see is base on the old sample receipt of matachana sterilizer, the objective is to extract data of the receipt and insert to db with corresponding information needed, i will provide the doc that contains what value you need, what exception you must implement. As the system currently implementing harsh try and catch logic as it is critical if the program is terminated by error, so we need to log and catch every listed out error.
attached files are information you need for development, pasted image refers to validation set, and the excel has the parameter of receipt to extract, as well as the receipts pdf in the receipt files, you can dig deep in 2026-04 under process file. Please follow the upper instruction or you can refer to the other implementation in the codebase"