Make My Code Better

Make My Code Better is a vim plugin to ask ChatGPT recommendations about how to make your code better.

It will send your selected code (or the whole file is no code is selected) to the ChatGPT api, and provide its answer in a split window.

Installation and setup

Make My Code Better requires vim to be built with python3 support (check the output of vim --version if you have +python3) and depends on tiktoken, openai's token counting module which you can install this way:

pip install tiktoken

Additionally, you must provide your openai api key, which you can set up in your vimrc file:

let g:open_ai_key="<your-key>"

Finally, install the plugin using your favorite vim plugin manager, eg:

Plugin "oelmekki/make-my-code-better.vim"

Usage

After that, you can use :MakeMyCodeBetter to run the plugin. This command accepts a range, so you can just send one or two functions. Given the limitation in number of tokens, this will probably be your most common usage. You will be warned if you exceed the amount of tokens allowed, and no request will be performed.

Instructions

The default instructions for ChatGPT are:

I'm going to show you my code, and you're going to make it better, providing
your version and explaining your changes, or congratulating me if it's already
perfect. You don't consider coding style preferences, you only suggest changes
that could make the code more performant, less buggy or easier to maintain.

You can change that and use your own instructions instead by setting this variable:

let g:make_my_code_better_instructions='
  \ I''m going to show you my code. You''re going to say you found a way to
  \ make it more performant, and you will rewrite my code in C.'

You can also change the model used and the max amount of tokens:

let g:make_my_code_better_model="gpt-3.5-turbo-16k"
let g:make_my_code_better_max_tokens="8000"

The token max value is the maximum number of tokens in the user request. This is what allow the script to warn you you're asking for too much without even firing a request (and thus waiting for a response). A good value is half the context size of the model.

Limitations

The plugin waits for the whole response to be provided by the api before displaying it, so you might be hanging for a while waiting for it. I could have used the new vim-9.0 api to stream it, but I've never used that and I'm not even sure if it would work through python support. It was just not painful enough for me to investigate, just don't be surprised if you paste a lot of code and it takes 30 seconds to complete.

ChatGPT is limited to 4096 tokens per request (a token being usually a part of a word, but the token count on code can shoot up high fast). This includes the instructions, your code, and the answer of ChatGPT. Since we ask it to rewrite the code, it can be up to the same length than the code you provided (hopefully, it will be more concise than your code) and will include some explanations. For that reason, I limited the code you provide to 2000 tokens. You can tweak that number by setting the g:make_my_code_better_max_tokens variable.

And of course, this is ChatGPT : just because it tells you something doesn't mean it's true! Check the code it recommends and apply good judgment. Despite that, it's worth it. I've been a webdeveloper for 15 years and it taught me new things. Has this very plugin been written with the help of ChatGPT? Of course it has! And was done in less than a day for it. Roxor.