4 Comments
User's avatar
Jaden's avatar

Maybe you’ve talked about this somewhere but do you work in a team setting or solo?

- how to you generate tasks for input, and roughly how much of your time is spent on that

- do you have a team for reviewing PRs or all just auto merge

- also if you have a team how do you document and comprehend this volume of work (and at what level of granularity)

Halfcycle's avatar

Minimising human attention is the right thing to optimise. What we found building Halfcycle: the attention you can't remove is the decisions only you can make, and nearly everything else can go. A lot of the "remind me what you were doing" cost is agents sitting on a decision nobody wrote down. Asked once, before any code, each with a recommended answer, those stop pulling you back in, and the high-attention corner gets smaller.

Halfcycle does that inside Claude Code and sizes the process to each change, which is your matrix applied per change rather than per workflow. Worth a run on one real feature: npx halfcycle. How do you decide which quadrant a new task lands in: by feel, or something written down?

Abhishek Ejam's avatar

Subagents are super context hungry though. I always hit usage limit when use them in Claude Code. My approach is to break down the plan into tasks and let a single agent work on them simultaneously and create stacked PR. If the tasks can be done in parallel, I spin up multiple sessions in separate worktrees

Ayush's avatar

How much of the code do you review yourself? And do you still keep track of all the low level details in the code?