Tips for building software with generative AI

[mikethemelanson]

Select Topic Area

General

Body

---

When I first started chatting with Anton Mirhorodchenko (@GreenWizard2015) earlier this year for a featured article I was working on, he explained how it was that we were emailing back and forth.

“Just imagine: I'm responding to you in Ukrainian and omitting about 60% of the letters from the words, but ChatGPT understands what I am saying and provides an answer in English.”

He offered an example of how he composed a sentence using ChatGPT for text expansion:

Mirhorodchenko: h m n s Antn, m f ukr. im dev w cplsy.
ChatGPT: I think you are trying to say: "Hi, my name is Anton. I'm from Ukraine. I'm a developer with cerebral palsy. `

Anton has difficulty typing because he has cerebral palsy, so he’s been exploring how to combine ChatGPT and GitHub Copilot to build software from the ground up: from setting up the environment and designing the architecture to writing the code itself.

The community was so interested in Anton’s use that we asked him to contribute a Guide sharing more.

You can read Anton’s Guide, Harness the power of generative AI for software development, (authored with the help of ChatGPT) in The ReadME Project’s May edition. He writes about his efforts to build software using large-language models (LLMs), and offers his tips for getting the most out of tools like ChatGPT and GitHub Copilot.

I recently conducted research on the effective use of AI for creating quality code, which I published in an open source project called User Country Analytics. Instead of focusing solely on the ability of AI to solve problems, I wanted to determine AI’s ability to use best practices, including testing and other aspects. [...] Just like in real life, I started the project by consulting with an expert: a software architect. In this case, however, the expert was not human, but rather ChatGPT playing the role of software architect.

I highly recommend checking it out. Working with @GreenWizard2015 has definitely changed the ways I've looked at using LLMs for so many things.

❓Also, I'm curious - what tips would you add to the list? Do you have a favorite model? Are you bringing LLMs into your developer workflow somehow, whether GitHub Copilot or a custom open source project? Or maybe you combine GitHub Copilot with Codespaces, like @noahgift does for running MLOps in Rust? Or how about accelerating your test-driven development (TDD) workflow by using GitHub Copilot to write your tests, like @mokagio?

I’d love to hear what y’all are working on! It’s an exciting time for AI.

---

The ReadME Project is a GitHub platform dedicated to highlighting the best from the open source software community—the people and tech behind projects you use every day. To get the latest stories, articles, and best-practices from The ReadME Project delivered directly to your inbox, subscribe to The ReadME Project newsletter.

[100Khushi]

Thanks for sharing Anton's impactful use of ChatGPT in software development! It's amazing to see how AI techniques like voiceover or text to speech is aiding individuals like Anton in overcoming unique challenges.

❓ How are you incorporating large-language models (LLMs) such as ChatGPT or GitHub Copilot into your development workflow? Share your favorite models and tips! Whether it's MLOps, custom projects, or enhancing test-driven development (TDD), let's chat about the exciting intersection of human creativity and AI in coding. Explore more about AI voice over in development.

[anni-anni]

Generative AI is improving at such a fast pace that I doubt whether after sometime, we would simply have to put in proper prompts and the AI itself will be able to code another AI. Just some days back I used the Bard AI to code some simple lines on how I could integrate my external AI dubbing / text-to-speech software with my smartphone at home that my family uses. I used teh codes and didn't work as well as I wanted it too, but nonetheless, it was impressive 👍🏽

[crguezl]

I'm curious - what tips would you add to the list?

@mikethemelanson

In my first interactions with the AI I used to analyze its response and modify it to make it correct.
Over time I have realized that if you need other people to reproduce the chat, this methodology is not the most appropriate.
It is sometimes better to reformulate/refine the questionnaire so that the probability that the AI generates a correct answer increases.

Thus my proposal is: Do the refinement, refactoring, debugging and optimization processes not so much on the initially AI-generated code, but on the question script.

As a consequence of this:

• If you want the questionnaire to be reproducible , save the list of questions on a markdown file
• Use markdown and $\LaTeX{}$ for communication with the AI. It has lot of advantages: you are stating what is code and what is a math formula, what you emphasizes etc. Copilot and Chat GPT answer seems to be influenced by the markdown notation.
• Refine the list of questions until they lead Copilot to a correct solution with high probability
• "the program" is the question script not the IA generated code

[Ahmadraza889]

Very good