The 2017 paper that introduced the Transformer was called Attention Is All You Need.
I’m borrowing it. The paper’s attention is a computational mechanism. Mine is the kind needed to stay with a document instead of opening another tab.
Back in April, I wrote about the curse of possibility: AI makes more things possible, and each possibility asks for another decision. I want to get a little more practical this time. Once all these things are possible, what gets a proper stretch of attention today?
Imagine asking an agent for ten approaches to a problem. Getting the list might be easy. Choosing an approach, checking its assumptions, and understanding it well enough to use it still take work.
And the agent can keep going while I do something else. If I start several jobs, though, I also arrange for several results to come back needing a decision. Being able to start them says very little about whether I have room to deal with them.
What goes on the list
The part of David Allen’s Getting Things Done I want to keep is the habit of putting commitments somewhere I can return to, then deciding what each one requires.
GTD has five steps: capture, clarify, organize, reflect, and engage. Capture is only the beginning. Clarifying something can mean deciding on a next action, saving it for later, or throwing it away. Regular review keeps those decisions current.
An interesting project suggested by an agent can go on a someday list. Recording it doesn’t have to turn it into a promise.
Take “write a blog post.” It could mean finding sources or rewriting an opening. If I put both under one vague task, I still have to figure out where to begin every time I look at it.
“Draft the opening paragraph” gives me somewhere to start. The rest can wait on the list without needing to be rehearsed in my head.
But a clear next action still needs time somewhere.
Give it a place in the day
Luke Seavers’s Time-Blocking starts with single focus, deciding what matters, and making room to recharge. Marc Zao-Sanders’s Timeboxing brings tasks into the calendar, with a decision about when to do them and how long to spend.
The terms overlap in practice. I find two questions useful: where in the day will this happen, and how much time am I willing to give it?
For the blog example, I could reserve forty-five minutes for the opening. A meeting already on the calendar has to be accounted for. So does lunch. The list can keep growing; that afternoon cannot.
Inside the block, the task is small enough to recognize: work on this paragraph. Without auditioning a better writing app halfway through.
Nir Eyal made a useful point on X: timeboxing focuses on the time and attention we put in, while outcomes aren’t always within our control.
I can choose to work on an argument for forty-five minutes. I can’t promise it will become good at minute forty-four. Sometimes the useful result is finding the hole in it.
When the time is up, I get to decide whether the task deserves another session. That stopping point matters for research and polishing, where there is usually something else I could check or change.
Giving this paragraph another hour means taking that hour from something else.
Count the work that comes back
Personal Kanban, by Jim Benson and Tonianne DeMaria Barry, has two rules: visualize your work and limit your work in progress.
The second rule is what interests me. A board can make overload easier to see, but moving cards around doesn’t reduce it. The limit has to affect what I start.
Suppose I allow two substantial tasks to be active at once. Starting a third means finishing or deliberately pausing one of them. Two is an example, not a prescription; the useful part is having a limit that forces a choice.
With agents, I would include work that still needs my review. If an agent has generated a draft I haven’t read, that task still has a claim on my time. Sending more jobs off to run can leave me with several drafts waiting to be read.
So I would set the limit around what I can review and finish, rather than around how many agents I can launch. Tasks genuinely waiting on someone else can sit in a waiting column, with enough context to pick them up later.
That leaves a question a task board can’t settle: how much of the day should belong to tasks at all?
Leave some of it unplanned
A day divided into colored rectangles can look settled before it has even started. I can see the appeal. I can also see how it becomes another thing to fall behind on.
If a plan needs every block to go exactly as expected, an urgent request can displace the whole afternoon. For someone on call or caring for another person, interruptions may be part of the responsibility. A useful plan has to leave room for that and make it possible to renegotiate what no longer fits.
I would rather start with one protected block and room around it than allocate every available minute. And when a block gets interrupted, I want to look at what happened before turning it into a verdict about discipline.
The same goes for choosing the task. Forty-five minutes of focused work can still go to something I no longer need to do. GTD’s review step matters here: a task can be removed, not just rescheduled.
I also want room for attention that produces nothing I need to report. I would like to listen through a meal without checking a screen, or go for a walk without using it to catch up on something.
For the walk, I’ll leave the productivity podcast at home.
2017 年介绍 Transformer 的那篇论文，名字叫 《Attention Is All You Need》。这标题实在好，我也想借来用一次。
论文里的 attention 是个计算机制。这里想聊的是另一回事：打开一份文档以后，能不能先别去开下一个标签页。
四月写过一篇《可能性的诅咒》。AI 让很多原本嫌麻烦的事都变得可以做了，每件事又要多做一些决定。这次想往下聊一点：能做的事已经这么多，今天到底给哪件事一段完整的注意力？
比如，让 agent 给一个问题列十种解法，拿到清单可能很容易。可选哪种、前提对不对、自己有没有理解到能拿来用，还是要花心思。
agent 跑着的时候，我当然可以去干别的。但要是同时开好几个任务，也就安排好了几份等着自己判断的结果。开得起来，跟有没有空接得住，是两回事。
先弄清楚，清单里放的是什么
David Allen 的 GTD，Getting Things Done，我最想留下的是这个习惯：把惦记的事记到一个回头找得到的地方，再想清楚每件事该怎么处理。
GTD 有五步：收集、理清、整理、回顾、行动。收集只是开头。理清以后，有些事需要一个具体的下一步，有些先放着，还有些可以直接删掉。之后也要定期回来看看，这些安排还算不算数。
跟 AI 聊出来的新项目，也可以先放进“以后也许想试试”。记下来就好了，不用当场答应自己一定去做。
就拿“写一篇博客”来说，到底是找资料，还是改开头？都塞在这五个字里，每次看见它，还得重新琢磨从哪儿下手。
换成“先写出开头那一段”，就有地方开始了。剩下的事留在清单上，不用一直在脑子里过。
可下一步再清楚，也得找个时间真去做啊。
给它在今天留个位置
Luke Seavers 的 《Time-Blocking》，前面几章就讲一次顾好一件事、想清楚什么重要，也讲休息怎么安排。Marc Zao-Sanders 的 《Timeboxing》，则把事情放进日历，提前决定什么时候做、打算花多久。
这两个词在实际使用里有重叠。我觉得有用的是两个问题：这件事准备放在今天的哪一段？愿意给它多少时间？
还是写博客的例子，可以先给开头留四十五分钟。日历上已经有的会要算进去，午饭也要算进去。清单可以继续加，下午就这么长了。
到了这段时间，就先写这一段。别写到一半，又去找更好用的编辑器了。
Nir Eyal 在 X 上提过一个角度：时间盒让人把心思放在能投入的时间和注意力上，结果却不总能由自己决定。
我可以决定认真想四十五分钟，但保证不了第四十四分钟时，那个论点就突然想通了。有时能发现原来的想法有个洞，这段时间也算花得值。
时间到了，再决定要不要继续给它下一段。查资料、改文字这种事，总还能找到一点可查可改的地方，有个停下来重新看看的机会挺好。
再给这一段一个小时，那这个小时就得从别处拿。
agent 跑完了，不代表事情做完了
Jim Benson 和 Tonianne DeMaria Barry 的 《Personal Kanban》，只有两条规则：把工作摆出来看，限制同时进行的工作量。
我更在意后面那条。看板能让人看见自己摊了多少事，但光挪卡片，事情也不会少。这个上限得真能拦住自己继续开工。
比如，先允许两件需要认真投入的事同时推进。想开第三件，就得完成一件，或者明确暂停一件。两件只是举例，重点是得有个上限，到了就要做选择。
用了 agent，我会把还需要自己审的结果也算进去。一份没读过的初稿，还得等我花时间。再多派几个任务，等它们都回来，自己还得一份份看。
所以这个上限，要按自己能审完、能收尾多少来定。至于真的在等别人回复的事，可以先放进等待区，把上下文记好，回来接着做。
不过，看板管得了任务，还没回答另一个问题：一天里，到底有多少时间非得拿来做任务呢？
没排满的地方，先留着吧
一天切成一格一格，再涂上不同颜色，事情还没开始，看着就挺有秩序了。我能理解那种吸引力。可也很容易多出一件要追着赶的事：自己排出来的日程。
如果每一格都得按预期走，一件临时急事就可能挤掉整个下午。值班也好，照顾家人也好，有些打断本来就是责任的一部分。计划得给这些事留位置，放不下的工作也得能重新商量。
我更愿意先保住一段时间，旁边留点空。不用一下子把所有能用的分钟都分完。真被打断了，也先看看发生了什么，别急着给自己下“不够自律”的结论。
选什么来做，也得回来看看。就算认真做了四十五分钟，也可能回头发现，这事已经没必要做了。GTD 里的回顾，在这里就有用了：一件事可以删掉，不必只会往明天搬。
还有些注意力，我想留给不需要交代产出的事。吃饭时听人把话说完，别惦记着看屏幕。出去走走，也不一定顺手补点什么知识。
散步那会儿，就先不听效率播客了。
Comments留言