What an Academic Interview Taught Me About How I Actually Use AI

The article discusses the author's experience using AI tools in their software development workflow, including how it has changed their process, what they delegate to AI, and the risks they see for junior developers.

Details

The author describes how their software development workflow has changed with the adoption of AI tools like Claude Code. They now use AI at nearly every stage, from starting new projects to preparing for meetings and mapping out risks. The author estimates that about 40% of the final code for their open-source package 'topiq' was AI-generated, which sped up the process by around 20x. The tasks the author delegates to AI include MVPs, documentation updates, and project boilerplate, as the output is easy to review. However, the author still handles tasks related to security, credentials, and critical architecture themselves. The author also explains that they do not let AI write tests, as it tends to shape the tests around making the code pass rather than solving the actual problem. The author had early issues with AI-automated commits, which led them to start treating constraint-setting as part of the workflow. The recovered time from AI-assisted tasks has allowed the author to work on more projects in parallel, invest more in building and testing new solutions, and think at a higher level about the systems they're working on. The author emphasizes that the most critical skills are now architecture and software design, as they enable setting clear boundaries for the AI model and understanding its actions.