Here's how I'm coding with AI lately, might be helpful! 1. I write code primarily using agents, using models like Opus 4.5 and Codex Max for long-running tasks or tricky bugs, and Composer for frontend changes or fast updates (I still review the code). 2. Most of my web dev work happens inside the integrated Cursor browser. This is similar to using the Playwright or Chrome MCPs. Cursor can access network requests, console logs, and send elements on the page to the agent. It can also control the browser! Which can be fun for having it do automated testing: https://x.com/cursor_ai/status/1999147953609736464… 3. I start projects pretty simple: no upfront rules, commands, or anything. As the project grows, I end up adding the most minimal versions needed. Some examples in this thread: https://x.com/leerob/status/2006043265817071926… 4. I heavily review all the code! Just because I'm using agents, I'm still thinking deeply about the architecture and code quality. I review the code in three passes: first, while the agent is generating it. Second, using the in-editor "agent review" before I push a PR (similar to a custom /code-review command). And finally, using Bugbot (AI code review) on my PRs. This combination helps me fix a lot of silly bugs before I ask other people to take a look at the code. 5. I always start new features with a plan (using Plan Mode). This helps significantly and I would highly recommend planning first regardless of what tool you use. I do like the Cursor UX for visualizing, editing, and saving the plans. You can view some of my plans here: https://github.com/leerob/pixo/tree/main/.cursor/plans… 6. For really hard bugs, I use Debug Mode. It automatically instruments your app with logging, and then asks you to reproduce the issue. The agent then reads the logs and has much more helpful context to pinpoint the root cause. It also comes up with multiple theories on what the issue could be, and works through each one until it's fixed. Has been pretty helpful: https://x.com/cursor_ai/status/1998821350333440133… 7. Always make sure you give coding agents verifiable outputs! They can't fix what they don't know about. For this, I would prefer using typed languages, and set up tools like linters and good tests. These are normal software engineering best practices, but they matter more than ever. There's also newer tools here like tsgo and biome/oxlint and bun which make dev really nice. Worth trying some of those if you do web dev. 8. I use Cursor from mobile! There's really two modes here: quick bug fixes or really big tasks. Quick bug fixes, I just pop open http://cursor.com/agents and fire away, knowing I'll get back a PR that will work 99% of the time and I can merge away. Easier than writing it down on a todo list, my PR queue is now that list. For big tasks, again I start with a plan and then I give the agent an ambitious goal (that is verifiable!). This allows the agent to run for much longer. It will keep going until it hits that goal, and if it gets lazy, you can just say "keep going" and go back to what you were doing. This is all in the cloud, in remote sandboxes, so I don't have to worry about my local machine. 9. Since someone will ask about the theme... yes I'm rocking light mode most of the time, using Cursor Light here Oh and if you're a car person... more on my car below soon