Almost every serious productivity method starts in the same place: get it out of your head. Capture everything, immediately, without judging whether it matters. The reasoning is sound — deciding and recording are different cognitive operations, and doing them together means you do both badly.
Then comes step two, and step two is where these systems break for coordination work. Step two is processing: go through what you captured, decide what each item is, give it a project and a next action, file it. Weekly, ideally daily.
Processing is the part that fails
It fails for a reason that has nothing to do with discipline. Processing is a batch operation with no external deadline, competing against work that has both immediacy and other people attached to it. Any week that gets busy, the processing session is the thing that moves. And once the inbox has three weeks in it, opening it feels like an archaeological dig, so it moves again.
The system then enters its terminal state: you still capture, because that's cheap and the habit is real, but you no longer trust the captures to be processed, so you also keep the important things in your head. You now have the overhead of the system and none of the relief.
What if the capture arrived already processed?
This is the part where the technology actually changes the method rather than just decorating it. Reading a messy human sentence and proposing what it obligates — the action, the person, the date, how confident that reading is — is a task language models are now genuinely competent at.
Which means step two doesn't have to be a batch operation you perform. It can be a proposal you glance at. The capture goes in as a voice note in a car park; by the time you next open the app, it's a suggested action with an owner and a Thursday deadline, waiting for a yes.
The habit that survives is the one whose payoff is immediate. Capture that produces a structured proposal within seconds pays off now; capture that produces a queue pays off never.
Practical notes on capturing well
Some things that make the extraction better, none of which require you to change how you work:
- Say the deadline out loud. "Before Friday" extracts cleanly. "Soon" does not, and no amount of model quality fixes that, because the information isn't there.
- Use names. "Ask John to confirm staffing" produces an action linked to a person. "Ask him" produces an action linked to nobody.
- Don't pre-summarise. Capturing the original message beats capturing your compression of it — you'll want the original later, and extraction does better on the real thing.
- Two obligations in one sentence is fine. "Send the schedule to Sarah and ask John to confirm before Friday" should produce two actions, not one vague one. That's a test case, not an aspiration.
- Capture the ambiguous ones too. "Sarah might need the revised schedule" isn't a commitment, and it shouldn't become one automatically — but it should be visible as a possible reading rather than lost.
The part that stays yours
None of this removes judgement, and it shouldn't. Deciding whether a proposed obligation is real, whether it's actually yours, and whether it matters this week is the work — and it's work you were already doing, silently, every time you read a message.
What changes is that you're doing it once, deliberately, against a list, in a moment when you're able to think — instead of thirty times a day in the two seconds before the next notification.