If you activity alongside Claude Code or Codex from the terminal, you cognize this one.
You commencement a task. The AI takes period to reply, so you commencement another one. Then the archetypal one replies, but the answer needs one clarification, one prosecute up and one flank project — and all of them wants its own tab. Then the second tab's answer comes in. You juggle. And all of a abrupt you discover yourself hopping between 20 threads, trying to keep all those contexts in your brain RAM.
Until you can't. Your brain gets exhausted, your concentration smashes, and you don't cognize anymore what you were doing, anywhere you started, and anywhere you were expected to finish.
A lot of group are suggesting taking breaks, thinking before typing, limiting scope. I have my own experiment.
I decided this way of operating has two important problems: sequencing and context. So I organised my activity a bit akin a precedence queue.
Here is how it really works.
The queue lives in git. Every ticket is a markdown document in the repo, correct next to the code it talks about, and the province of the commission is a committed snapshot. Nothing crucial lives in a conversation opening and nothing lives in my head. If a meeting dies, the queue is motionless there.
An delegate plant that commission in cycles. It researches new tickets, builds the ones a individual has tagged as harmless to build, runs the tests and a browser check, and verifies what is already deployed. For item touching delicate code it builds on a branch and asks first. It never deploys to manufacturing — that part is mine.
What it cannot do is decide. So everything that needs me lands in one location I call my owner inbox. A manuscript says the entire commission and hands me a prioritised catalog of exactly the items waiting on my judgement — assessment this change, answer this question, regulation on this tradeoff.
The study exterior is a Claude artifact. One page, rebuilt from git on all cycle, mirroring the entire commission affirmative a assessment leaf per ticket. So I don't scroll rear through a terminal trying to rebuild what a alter does. I open one link, peruse what was built and why, and provision my verdict. The verdict goes rear into git, onto the card. The leaf is lone a spectator — git is the record.
Then comes the part that matters most.
I obtain one item. I burden its context. I create a decision or petition an action. And I near it, by putting it rear into the queue.
That decision is the entire trick. The ticket doesn't keep hanging in my recollection as a item to recall about. The environment closes alongside it. So whenever I choice the next item from the queue, I can really focus. One item at a time.
It is motionless exhausting. But I don't awareness distracted or lost.
The difficulty was, and motionless is, the tooling about it. Getting the environment served up in a digestible manner, so loading a ticket takes a infinitesimal and not twenty. And keeping tickets from going into endless loops since they are blocked waiting for a ruling lone I can give.
It's a living organism and I keep improving it. But it felt akin a huge stage onward in my efficiency alongside AI tools, so I idea it was value sharing.
