How developers spend the time they save thanks to AI coding tools

The introduction of AI coding tools is already changing the way developers work.

Almost all respondents in our current survey said they had used AI coding tools at some point – and the vast majority said these tools help make it easier to write higher quality, safer code. Likewise those Survey on the status of JS 2023 found that only 18% of respondents not Use an AI coding tool regularly.

Generative AI for the IT professional Co-author Chrissy LeMaire, recently told us that LLMs are changing their workflow. “When you start a project, you have to set up all sorts of things,” she said. “It takes a while to get to the exciting parts. Now I’m letting the LLM do it for me. Things start to get exciting with AI coding.”

Research from GitHub suggests that AI tools can increase developer productivity by up to 55% – but what do you do with the time saved by using AI coding tools? According to our survey, respondents are spending more time on system design, collaborating with colleagues, and learning new skills, among other things.

Additionally, AI tools improve their work and create positive feedback loops that not only help them write more code, but also better code. To find out what this looks like on the ground, we spoke to developers about how AI is changing their workflows and what they’re doing with the time saved by using AI coding tools.

In this article we will:

• Hear from developers how they save time using AI coding tools.
• Give us a few tips to get you started.

Less time debugging, more time planning

Many developers report the ability to invest more time in the design and planning phases thanks to AI – and that means more time working as a systems thinker, which is a net benefit. “I spend less time trying to figure things out and more time making sure my code is secure and performant,” open source developer Claudio Wunder told us in our current questions and answers about AI coding tools.

He is not alone. In our survey, 40-47% of respondents said AI has enabled them to spend more time developing systems and customer solutions. Meanwhile, 37-43% of respondents said they are spending more time refactoring and optimizing code.

In other words, developers spend their time improving their code instead of just trying to make it work.

The process begins before you even write your first line of code. For example, Wunder uses the GitHub Copilot Chat to think through projects. This idea of ​​explaining ideas to an inanimate object to clarify one’s thoughts is called “Rubber ducking.” But LLMs add a new dimension to this activity: you can talk back. Instead of just organizing his thoughts, Wunder actually receives feedback on his ideas and goes into new projects with a clearer idea of ​​how he wants to structure his code. LeMaire does much the same thing: “I spend less time grunting and more time just talking.”

Here are some practical tips for getting started designing systems and refactoring code with GitHub Copilot:

• Indicate your preferences. Wunder starts new projects by telling GitHub Copilot Chat that he prefers ES6 built-ins and arrow functions. “These simple statements can usually help you achieve the desired code output and better understand Copilot’s thought process,” he explains.
• Share examples. LeMaire approaches projects the same way she did without LLMs: by finding examples of similar things. “I upload the sample files, sometimes concatenate them into one big file, and tell the LLM what I want,” she says.
• Start with a skeletal function. When it comes to generating inline code, Wunder recommends starting with meaningful parameters, arguments, and comments that explain what the function should be and what each parameter should control.
• Debugging as a conversation. Wunder keeps all the related code open in VS Code and starts a new Copilot chat session with the prompt “Let’s debug some code.” He then asks Copilot questions, such as: B. what it thinks the code is doing and what would happen in response to various user inputs. “I try to give Copilot as much context as possible about what the code is trying to achieve and iterate with follow-up questions until I find the problems and solutions,” he says.

Less time working on documents, more time collaborating

AI is not just used to communicate with machines. It also gives the developers time to talk to each other. 40-47% of our survey respondents say AI helps them have more time to collaborate with team members on projects. Another 39-45% said they are spending more time on code reviews, which is part of it Main methods for developer collaboration And Help each other do better work.

LLMs can automatically generate code comments and documentation, making code easier to understand and therefore easier to contribute to. “I was able to go through some JavaScript and have an LLM create JSDoc-formatted documentation based on function names and parameters with about 95% accuracy,” says LeMaire. “My team really loved it.”

The result: Not only do developers have more time to collaborate, but they can also do so more easily.

Here are some practical tips to get you started using LLMs to improve collaboration:

• Use your favorite existing help text as examples. LeMaire recommends providing LLMs documentation examples in the style you want to replicate. For example, she prompts the LLMs with help text from her favorite PowerShell commands to help the LLMs create documentation that matches the tone and format she prefers.
• Leave comments on each file. When you open a code file, add some comments at the top as a header to help GitHub Copilot understand the code better. “This will increase both your and your team’s productivity as you move past these little tidbits,” says GitHub developer advocate Christopher Harrison. “A digital version of donuts in the break room.”

Less time searching, more time learning and experimenting

Similarly, DevOps architect Alessio Fiorentino used GitHub Copilot to learn Rust. “Rust is a powerful language that offers full control over the flow of execution, but it has many nuances and requires a different way of thinking, especially for those who started with Python or JavaScript,” Fiorentino told us in a previous article. “AI helps me deal with this complexity and ensures that I write efficient and idiomatic Rust code.”

AI coding tools can be helpful, but they still require pilot testing – and they are not a replacement for learning. “Even if LLMs can generate entire applications in the future, you still need to evaluate the code,” says Wunder. He sees the developer’s role changing as LLMs take care of implementation details, and recommends developers focus on understanding high-level computer science concepts and sharpening their communication skills. Fortunately, there is some synergy here: you need to write clear instructions for using an LLM, so AI coding tools actually strengthen the skills that developers need to hone in the future.

Here are some practical tips to get you started using GitHub Copilot as a learning tool:

• Navigate a new or unfamiliar language or technology. Wunder recommends using GitHub Copilot to walk through the syntax and features of a specific language. “I recently started learning Go and asked Copilot, ‘What does adding a type after mean?’ :=< on a variable definition?' It also helped me understand how namespaces and module definitions work in Go.”
• On board with new code bases. Try highlighting a block of code and asking Copilot to explain it, or ask questions about the code, such as: B. which variables relate to certain functions.
• Visualize what you learn. Kedasha Kerr, GitHub developer advocate, has I used Copilot's mermaid charting features to better understand how data flows through an application.

What's next?