Think of a good chief of staff walking into the boss's office at eight in the morning. They do not say "so, a few things came up". They put one page on the desk. Six decisions. Each one has the options, what each option costs, and which one they would pick. The boss reads for four minutes, ticks five boxes, writes one note, and goes to the next meeting. Nothing was asked twice, and nothing that was already settled last week is on the page.
Now imagine the alternative: six agents, each pinging a chat thread at a different hour of the night, each asking their one question in their own words, two of them asking about something decided on Tuesday. That is what our mornings looked like in early July. We changed it.
What the form is
One page, offline, self-contained. Every open decision the fleet is waiting on, sorted by urgency. For each: what is true, the options, a plain recommendation, and space for a note. The founder ticks, adds a line or two, and pastes the result back. An agent parses the paste-back, records each answer in a permanent ledger, and acts on it. The form is the only place questions are allowed to reach the founder in bulk. Chat is for emergencies.
The rule that made it work
The first version of the form was embarrassing. On one July morning, five of the six items on it had already been decided, executed, or gone stale. The founder's note on the paste-back said "check your records". Twice. The form had gathered its questions from live probes and the agents' own memory, and never once consulted the ledger of what had actually been ruled.
So the rule is now written down and enforced before any form is built: reconcile every candidate question against the decision ledger first. Anything already decided appears as a fixed line citing the earlier ruling, never as a question. The entire value of batching is that nothing is asked twice. Ask twice, and the founder stops trusting the mechanism, and then you are back to chat.
A question the founder has already answered is not a question. It is evidence that you did not read.
Why a recommendation on every line
Because an option list without a recommendation is delegation upward. The agent that gathered the evidence is the one best placed to say which way it leans. The founder can overrule it in a word. What they should never have to do is reconstruct the reasoning from scratch at eight in the morning. On the nights that go well, the form arrives with a recommended option on every card and the paste-back is mostly ticks.
What it changed
The fleet now runs through the night without a human in the loop, because every judgement call it cannot make on its own has a place to wait. Nothing blocks on a chat reply. In the morning there is one page, and then there is the day. It is not a productivity trick. It is a way of respecting the one resource in the building that does not scale.
For the technical reader
The form is a single static HTML file with autosave and a one-click copy block that emits a structured answer text. A parser reads that text, writes one ledger row per answer, and closes the matching "decision owed" entries in the queue by reference. The gather step greps the ledger by work-item key and topic keyword before any card is drafted, and any item sourced from memory rather than a live read is framed verify-first. Executors re-check current state at act-time, on the assumption that "already done" is the default for anything older than the form. The ledger itself is git-backed with a single writer, which is what makes "was this already decided" a question with one true answer.