GitHub Copilot Chat: This Moment Changes Everything

Last year, I wrote about my experiences using GitHub Copilot. It is a helpful tool for code auto-completion. You get basic functions with clean formatting.

Consider adaptive cruise control on cars. It takes care of the chore of maintaining a safe speed between your car and the car ahead of you. But you still have to steer. This is what GitHub Copilot is for writing code. It takes over some of the chores, but you have to verify it is doing the right thing.

So when GitHub Copilot Chat became generally available, my expectations were set appropriately. Or so I thought.

I have used it for a couple of weeks now. I clearly underestimated the impact of this product release.

Copilot Chat is impressive. Developers now have a personalized, tailored assistant right at their fingertips. It is ready to provide solutions tailored to the project at hand. Imagine having a custom-built tutorial for your project. Now imagine also having access to the author. And asking as many questions as you want, without them getting tired.

Copilot Chat is well-trained, focused, and polished. It is the new standard for AI product user experience.

In many instances, I thought I'd finally stumped it with my most difficult problem. Then it came back with a solution at least 90% correct. Usually, Copilot is no more than one prompt away from a 100% correct solution.?

If I do not understand what Copilot suggests, it can explain it to me. It never gets bored, impatient, or sassy.

Copilot Chat is to developers today what spreadsheets were to accountants in the early '80s. It is a complete redefinition of your job. Considering “what if” scenarios is now fast and easy. It offers a respite from context fatigue. You now focus on asking the questions and testing the solutions.

Buckle up: we’re never going back. But there are some pit stops along the way. Let’s start with what it is like to use Copilot day-to-day.

Practical use of Copilot Chat?

To say that Copilot Chat is impressive is an understatement. But you still have to learn how to use it effectively.

The workflow is Select, Prompt, Verify, Test.

First, you need to select a section of code you want Copilot to consider. You do not have to do this step if you want to ask a general question. But having a piece of code ready for Copilot to read will generally yield better responses.

Next, write a prompt in the chat window. Give an overview of what you would like to do. Also supply what (if anything) has been your approach so far.?

Being terse and commanding Copilot here is less helpful. You will likely get a message starting with “as an AI, I don't have the ability to…”. But it will still try to generate something as a response. Giving as much context as possible will get the best responses.

Once Copilot has come back with a response, it is time to verify. This is where your software engineering expertise comes in. If Copilot missed something, follow up in the chat. You will likely receive a response starting with “I apologize for the confusion,” followed by a correction.

Finally, it is time to test the code from Copilot. It provides a few buttons for copying the code in. One button copies, another skips copying and replaces whatever you have highlighted. After pasting in the code, be sure to test it by running it or using automated testing.

This workflow is natural if you write code on a regular basis. And if you are new to writing code, you can ask Copilot to explain concepts to you.

Where does Copilot Chat stop?

Copilot represents a massive leap forward in your productivity as a software engineer. But it does not do everything. And it also has a few shortcomings.

First, Copilot will not touch your codebase at all. You are having a conversation with Copilot in a separate panel. It generates snippets and suggestions, but you are responsible for implementing them. With generative AI still being new, this is a good guardrail. It ensures you have a chance to review code before it is a part of your codebase.

Next, Copilot doesn't run unit tests or verify the functionality of the generated code. You must test and validate the code to ensure its correctness and reliability. Copilot generates code, but the responsibility for testing and debugging is still yours.

Also, while Copilot excels at providing solutions quickly, sometimes it introduces suboptimal code. For example, I once encountered N+1 database queries in suggested code. If I had not understood what an N+1 problem is, I would not have spotted it. When you point this out to Copilot, it apologizes and regenerates code. But you still have to know.

Copilot will sometimes lead you to a dead-end. This happened to me when I tried to refactor some Next.js code. I wanted to use server-side context instead of client-side state. Only after trying a couple of approaches did I realize I was trying to do the wrong thing in this instance. And Copilot did not stop or warn me that this was not a good idea.

Finally, using Copilot effectively requires breaking down problems into chunks, then asking questions. Copilot's understanding is limited to the context of a single repository. If your project spans multiple repositories, you will need to work with one at a time. And Copilot will remind you that it cannot read your codebase beyond selected snippets. (Even if it seems to behave otherwise.)

Despite these limitations, Copilot is still incredibly valuable. It is a useful sidekick for your development efforts and one you should not write off.

The Race Is On: Envisioning the Future of AI in Development

Today, GitHub Copilot Chat is the most advanced AI development tool available. It is easy to use and has an unparalleled focus on code.

Other platforms are also incorporating AI functionalities as add-ons to their existing products. For instance, you can have Sentry attempt to explain a runtime error based on the stack trace. It does what most other AI tools do: provides immediate and confident responses with the most likely correct information.

But this is all just the beginning.

Imagine a scenario where AI writes all the code for a product MVP. Everything is generated, including the frontend, backend, admin panels, infrastructure, and tooling. All it needs is a description of the product you are building and it can generate the rest.

To truly unlock this potential future, we need tools that go beyond raw prompts. Organizing work and persisting context are going to be the keys. In the current generation, AI is trained on existing frameworks and languages. In the next generation, platforms will be AI-first. The guardrails of not touching the codebase will be removed. Our concept of what a codebase is will transform, or even disappear entirely.

And it all may happen sooner than you think!