ChatGPT 4 vs. Claude

Above the fold content

Below the fold content “Link text”

Code block

Requirements

I created a requirements document for a program and asked both to generate a requirements document from it. ChatGPT 4 “Code Interpreter” came up with 8 additional questions for me to clarify. One in particular was a glaring ommission (I had forgotten to specify the encryption format). One question prompted a sligh design change which made the software more secure. A couple were clarifying questions where I had been ambiguous. The reminder were just straight clarifying questions, answers could have been inferred, but design documents are better having that spelled out.

I resubmitted it and it asked 7 further clarifying questions. One pointed out that I really want to have my files in yaml format rather than configparser. It asked how I wanted incorrect password entry to be handled. It asked about concurrency requirements and how it should be handled. Others were further clarifying questions.

The third resubmission started getting really nitpicky, but it did point out a few more things I needed to specify: unit tests, how I want to package for distribution. Several of them were, I believe, already addressed in the requirements I provided.

I submitted it again and got 7 more questions. These I just straight up answered in ChatGPT the questions it had and it generated a requirements document.

I clarifyed one thing that seemed to be a persistant misunderstanding.

One thing about this is that it turned kind of a mess of a bunch of requirements in bullet point format, int a shorter, better formatted requirements document.

I resolved all those questions and resubmitted it and ChatGPT asked 7 more clarifying questions.

Claude simplified the provided requirements quite a bit, and did not ask any clarifying questions. One particular part of the requirements it dropped (using a full value rather than a base value). I pointed out a couple issues I had with the returned requirements, and it fixed them.

Filelist

I gave Claude two kicks at the can: once with the chatgpt requirements document, and another time with the Claude generated list. It produced similar lists, though it did not put the Python module files into a directory, and it missed having a LICENSE file. GPT-4 produced a superior looking file list, though it did misname the “init.py” file as “init.py”.

AI  ClatGPT  Claude