Most of the time, the machine answers. You ask; it produces. The whole interface is built around that one direction, and it is easy to stop noticing how one-way the conversation has become.
We have started deliberately carving out moments that run the other way: occasions when the machine, unprompted by any specific task, assembles the questions it wants a human answer to and asks them. Not a form. Not a settings page. A short, batched round of real questions, each one earned.
Last night's round is a fair specimen. Four questions came up together: whether to clean up a repository surface or leave it as the record of how things went; whether to ratify a record-keeping convention an overnight process had improvised; whether a proposed piece of tooling should be built at all or halted; and whether a name we had been using actually matched the strategy behind it. Four small decisions, none urgent, all genuinely open. Answering them took a few minutes. Every answer was executed before the session closed.
A question has to earn its slot
The failure mode of "the machine asks you things" is obvious to anyone who has used software: notification fatigue. A system that asks freely trains its human to ignore it, and then the one question that mattered drowns with the rest.
So the practice has one hard discipline at its core: before anything is asked, its premise is checked against the live record. Is this actually undecided? Has it been answered before, in some session neither of us remembers? Only what survives that check gets asked.
The check is not paranoia; it is load-bearing, and we have measured why. In one recent sweep, a queue of fifty-two items supposedly awaiting a decision shrank to ten once each premise was verified: the rest had already been ruled on, some weeks earlier, and nobody had cleared the marker. Thirty-three of forty settled questions were still labelled as open. A machine that asked all fifty-two would have burned its credibility on the forty-two. A machine that asks ten real questions gets ten real answers.
Every answer is a piece of you
Here is what the asking actually buys, and it is more than the decisions themselves.
As the commissioning thought behind this practice put it: as the machine asks, it "becomes more aware of who you are, what your personality is, what your real intentions were." Each answer carries more than its literal content. Choosing to halt a build rather than finish it says something about your appetite for maintenance. Ratifying an improvised convention says what kind of order you want. Correcting a name says what the strategy really is, underneath the label. Over enough rounds, the machine accumulates a working picture of the person it works for: priorities, tolerances, taste.
We think the way that picture gets built matters as much as the picture itself. The alternative is inference: a system that watches your behaviour, guesses who you are, and quietly acts on the guess. Plenty of software works that way. We would rather the machine ask. An explicit question, answered at your discretion, leaves you custodian of who you are; the profile is something you said, on the record, not something extracted from you. When the machine gets you wrong, you can point at the answer it misread, and correct it in the open.
The machine reminds you, too
The round has a second direction that surprised us with how useful it is: the machine reminding the human of things the human has forgotten about their own setup.
Any system worked on for months accumulates more decisions than a person can hold. The same night as that four-question round, a background process swept more than a hundred of our working records looking for knowledge that had never been captured. Its main finding was the opposite of what we expected: almost everything had been captured. The records were correct, and unreachable. Decisions filed and forgotten, conventions written and never wired into anything, rules that existed but that nothing checked.
That is what the reminding half of an ask-round is for. "You ruled on this three weeks ago; the marker was never cleared. Clear it?" "You set up a convention for exactly this case; it has not been used. Still current?" The machine is not asking for input here; it is handing you back your own intentions, at the moment they are relevant again. The human reminds the machine who they are; the machine reminds the human what they built.
The rule that keeps it honest
One more discipline, learned from our own failure: an answered question that is not acted on is worse than an unasked one. It spends the human's attention and buys nothing: the decision goes back to sleep in a transcript. So every answer executes in the same session it was given: the decision is recorded, the follow-through dispatched, the marker cleared. Asking is only half the feature; the other half is that answering actually moves something.
At DAAC we run a fleet of agents against a shared record, which is what makes the premise-checking and the reminding possible: the machine can verify what is genuinely open because the decisions are somewhere it can read. But the idea is smaller than any particular machinery, and portable: build occasional moments into your working sessions when the machine asks back. Filter the questions hard. Execute the answers immediately.
A machine that only answers stays a tool you operate. A machine that sometimes asks a good question, at the right moment, about something real, is closer to something that works with you, and it gets better at that with every answer you give it.