One of the hardest parts of high-throughput AI coding is juggling context. You may have dozens of parallel threads getting resolved through several streams of work.
I've developed one counter-intuitive habit that actually helps. It involves opening up more threads of work.
Any time I think of something (a package that needs upgrading, a warning in the logs, someone mentioning a bug, a typo or css quirk), I drop it into Claude Code for Web. Then as part of my usual sweep, I merge and close these tasks out.
The aim of this exercise: remove the chores in your work. Not only the chores, but all the extended busywork around the chore: this can easily exceed the work and value delivered. That's often why it ends up in the chore bucket. A standard patching doesn't need to be ticketed, raised at standup, prioritized, raised again, re-proporitized, then picked up when you have a gap. By eliminating the busywork around these minor tasks you create more room. And that room gives you and the whole team the opportunity to take on deeper work[1].
In Resisting AI I talk a little about senior engineers pushing back on AI coding (hard, in some cases). Personally, I think people often start in the wrong spot. Start by getting AI to do the activities you don't enjoy, or where you feel it's busywork. Then ladder up from there.
Here are some specific habits that have helped me land this:
- Launch on your computer, your phone, your tablet[2]. Less friction is better. My notebook-and-coffee morning ritual used to be somewhat sacred, but now I bring my tablet too.
- If it doesn't work, just delete it. We're trained to not "throw code away", but a prompt that didn't land isn't code you owe anything to.
- Don't pre-triage. If you catch yourself wondering whether a task is worth opening a thread for, you've already spent more effort than just opening the thread.
- Batch the sweep, not the starts. Starting is cheap and should happen the moment something lands. Reviewing, merging, and closing out is a dedicated exercise.