GitHub Copilot can now help start a project with AI, not just complete it

GitHub wants Copilot, its code completion platform powered by GPT-4, present throughout the lifecycle of development, including the very beginning of a coding project. 

The company is announcing GitHub Copilot Workspace, a new service that aims to cut down the time engineers and developers spend reading through code and figuring out how to start on a new project. Workspace will only be available in technical preview for a waitlist of developers, but it will be integrated into the larger GitHub Copilot platform after it exits preview. 

GitHub wrote in a blog post that Copilot Workspace will be integrated into GitHub repositories or libraries. Developers can describe to Copilot Workspace, through prompts, what they want to do for the project. Copilot Workspace will then offer suggestions on how to start and provide a step-by-step process. Users can edit the suggestions and, once they’re satisfied with the suggestions, run the code (or even use Copilot to help complete the code) and finish the project. 

GitHub Next head Jonathan Carter tells The Verge they’ve heard from customers how helpful Copilot has been in the middle of a project but that it isn’t that helpful at the start. But adding Copilot to the very beginning of a project lets developers spend more time actually coding than reading the code and documentation around it to figure out how to begin. 

Carter says Copilot Workspace is helpful for reviewing older code because it can scan the code base and understand how it ticks faster than humans can read written documentation explaining it. 

Since the launch of GitHub’s Copilot code writing and completion, it has become ubiquitous and a common benchmarking skill for new AI models. Most lightweight models, or models with a more limited skillset, are optimized for simple tasks like summarization, writing tasks, and code completion. GitHub’s parent company, Microsoft, released a small model capable of code-writing tasks called Phi-3. Companies like Meta also released versions of their large language models for coding projects.