Ask HN: What is your GitHub Copilot (code LLM assistant) workflow?

8 points·dpflan··4 comments

Are you using a code LLM assistant? How do you use it? How have you changed your coding/software development routine to effectively utilize and integrate this tool into your day-to-day?

This post from today by Martin Fowler is pretty cool: An example of LLM prompting for programming

- https://news.ycombinator.com/item?id=35612494

- https://martinfowler.com/articles/2023-chatgpt-xu-hao.html

View on HN

I use it when I think it would have enough context in the current code to write a new function/class etc, and then defining it in a plain English comment:

// function that does XYZ

Hit enter, see the function suggestion and hit tab to accept it if it looks good. Sometimes I will reference specific variables or methods I want to it to use, again in the code comment. Which is much faster than writing the code block from scratch.

I also accept small autocompletes from it when I'm writing the code myself, it seems like a more powerful autocomplete in general.

Do you try things like make a set of unit tests and then have the LLM create methods that the tests test? It seems that there needs to be tight coupling and verification between what is generated and whether it actually accomplishes the task, and tests can guide this.

I haven't gone too much into testing code, but I'm sure it could look at your code and generate the entire unit test for you. I would just write it out as a comment again:

// Write a unit test for method xyz

Im suggesting the unit tests come first, are made by the coder (maybe with assistance), and the LLM creates the feature code. And the unit tests prove the generated code works or not then you tweak. Like TDD meets LLM

Reply on news.ycombinator.com