Summary
Highlights
The speaker reflects on how AI's role in coding has evolved significantly since 2023, making him more productive and enjoying engineering work more. He emphasizes that this video is for existing software engineers, not beginners, and acknowledges potential biases due to investments in some mentioned companies.
The speaker introduces ArcJet, a sponsor that simplifies API security, especially for endpoints backed by LLMs. ArcJet provides security as code, offering rate limiting and detailed decision-making for requests, making it easier to secure services without complex dashboards or configuration files.
The speaker demonstrates starting a new project to showcase AI's strengths in scaffolding, planning, and building. He outlines a project idea: a tool where one AI drafts an essay, another reviews it, and the first AI revises it based on feedback, saving all stages as markdown files. He uses Cursor's agent mode for planning and integrates tools like AISDK and OpenRouter.
Work trees are introduced as a way to work on multiple Git branches in parallel. The speaker uses work trees to run three different AI models (Composer, Opus, Gemini 3 Pro) simultaneously to implement the project plan, highlighting Composer's speed.
The speaker reviews Composer's implementation, finds it a good starting point, and tests the generated CLI tool with a prompt. After running it, he observes the quality of the generated essay and review, noting that the Composer version was faster and potentially better than other models due to formatting issues.
The speaker demonstrates creating a GitHub repository and using the GitHub CLI, lamenting its complexity. He then sets up Grapile for AI-powered code reviews, showcasing how it can catch bugs and provide useful suggestions, emphasizing the value of automated review tools.
For more complex tasks, the speaker highlights the importance of tightening the feedback loop for AI models. He explains how adding basic tests, dry run modes, and type-checking commands significantly improved the reliability and quality of AI-generated code by allowing models to self-correct.
The speaker advises treating the prompt, not the code, as the primary artifact to maintain and iterate on. He suggests that if AI models make significant mistakes, it's often more effective to adjust the original prompt or plan and regenerate the code rather than forcing fixes. He also discusses providing context through cloning relevant repositories or including code examples.
The speaker concludes by emphasizing that AI tools are excellent for tasks one already understands, encouraging experimentation with new tools without fear of falling behind. He highlights that using AI agents improves communication skills and ultimately makes one a better coworker by necessitating clear instructions and verification harnesses.