Revisiting my AI compromise

It’s been less than a year since my last AI post and boy have things changed. I haven’t been shying away from all the progress & tooling, but I haven’t been trying every single new model and tool either. I have, however, stepped outside my comfort zone and made considerable changes to the workflow I set out for myself in My AI Compromise.

Gone are the days of working on one thing at a time. My workflow at work now heavily relies on Conductor to orchestrate 2 - 3 agents using the latest anthropic model in separate git worktrees. It takes some getting used to, and claude is still not great at predicting exactly what I want but the workflow of having claude write a plan, thoroughly reviewing the plan, getting claude to execute the plan, and then thoroughly reviewing the diffs works for me.

Conductor does all of this orchestration for me, including keeping track of the PRs associated with each task. It’ll let me know when things have been reviewed, if there’s any comments (from humans or other agents) I should review, and then allow me to merge, all without ever going to GitHub. I’ve also set it up to play this charming French railway jingle when it needs something from me.

Still, it’s not too rare that I have to reach for an editor and make manual changes to guide claude further. For this, I find myself reaching for Zed far more than any AI-first editor. I think this allows me to strike a good balance between all the boring code being generated by claude and then me rapidly organising it or tweaking it using a fast editor. Even though Zed now has AI in its name, I use it with all the AI functionality disabled, just cause it’s so quick.

I’m still concerned about getting lazy, but even though I’m writing a fraction of the code I was before, I’m reading and reviewing so much more code. With all my co-workers also armed to the teeth with the latest AI tooling, there’s been a staggering increase in the amount of code being checked in every day. I hope this, along with remaining steadfast in not using AI for personal projects, continues to keep me sharp.