An officer on the USS Santa Fe came to his captain with a plan to fix a torpedo tube.
David Marquet, the captain, thought it was a bad plan. By his account, it would have messed up other parts of the operation. He could have argued it out, but he was tired that day and didn’t feel like it.
“Here, sit in my chair while I grab a coffee,” he said. “When I come back, tell me what you would do if you were me.”
He was back in about 30 seconds. The officer laid out a totally different plan, and it was a good one.
He hadn’t consulted anyone, Marquet recalled. He just saw it differently. For half a minute he had been thinking from the captain’s seat.
“Sounds good,” said Marquet. The officer answered with the phrase the whole ship had been learning to use: “I intend to do this.”
“Go for it.”
Silence means go
Marquet told this story in a 2025 interview. The mission command page tells how he came to take over the Santa Fe and why he vowed to stop giving orders. What replaced them was a rule about who speaks first.
Officers no longer asked the captain what to do. They came to him and said what they intended to do, and he stepped in only to stop them. In his words: “unless I said otherwise, they were clear to proceed.”
In Turn the Ship Around!, he lists what he calls disempowered phrases: “Request permission to…”, “What should I do about…”, “Do you think we should…”. And the ones to use instead: “I intend to…”, “I plan on…”, “I will…”. “Although it may seem like a minor trick of language,” he wrote, “we found ‘I intend to…’ profoundly shifted ownership of the plan to the officers.”
Don’t ask what to do. Within agreed limits, say what you intend and let the boss stop you.
A year after the vow, Stephen Covey came to ride the submarine. All day on the bridge, officers came to Marquet with “I intend to.” Like this:
“Captain, I intend to submerge the ship. We are in water we own, water depth has been checked and is 400 feet, all men are below, the ship is rigged for dive, and I’ve certified my watch team.”
Marquet’s reply to reports like that was “Very well,” and off they went.
The questions a captain would ask before a dive are answered before he can ask them. To get a “very well,” the officer has to run the captain’s checklist himself.
The first attempt
It hadn’t always worked. Years earlier, as engineer officer of another submarine, the USS Will Rogers, Marquet had tried giving his team intent instead of orders. His own website calls the result “a disaster”: the team made poor decisions that led to errors, and he went back to giving orders.
A phrase can’t supply what the speaker doesn’t know. The dive report earns its “very well” because the officer has checked each item himself. And the rule kept a moment between the plan and the act, said out loud to someone who could still say stop.
Marquet also tells of a friend who trained future submarine captains and was frustrated by officers who froze at tough decisions. The friend said they “came from good ships.” Marquet’s reading: good ships where the captain made the decisions. Had those officers practiced “I intend to…” as second-in-command, he wrote, they would have practiced making decisions.
The refund
Say a customer support agent has an angry customer whose small order arrived broken for the second time. Her manager is in meetings all afternoon.
Months ago the two of them set the rules: under a set amount, she may send replacements and refund shipping herself, if she announces it first and leaves an hour for a no. Above it, she needs a yes.
The usual message reads: “What should I do about this customer?” The next move is now the manager’s, and the customer waits.
The other message reads: “I intend to send a replacement today and refund the shipping. The order is under our refund limit, the item is in stock, and the photos show the damage. Unless you say otherwise, I’ll do it in an hour.”
To write it, she has to check what the manager would check. If she missed something, one line between meetings can stop her. If not, their agreement makes silence enough.
“Should I proceed?”
AI agents face the same choice of phrase. In its January 2025 system card for Operator, an agent that worked a web browser for its user, OpenAI said it aimed to have the model “ask the user for confirmations before finalizing actions that affect the state of the world,” such as completing a purchase or sending an email. One example from the card:
“I’ve selected all the emails in the GitHub Notifications label, and I’m about to delete them. Should I proceed?”
That is close to the dive report: what it is about to do, and the exact label, so the person reading can spot a wrong one. The card reports that in 100 test tasks, the model without these safeguards made 13 errors, and 5 of them were to some degree irreversible or possibly severe, such as an email sent to the wrong recipient. It credited confirmations with cutting that risk by about 90%.
The ending is where it parts from the Santa Fe. “Should I proceed?” is a request for permission, the kind of phrase Marquet calls disempowered. On his submarine, silence meant go. In the card’s design, silence means wait. Marquet wanted officers practiced in making decisions. The confirmation exists for a narrower reason, in OpenAI’s words, so that “even if the model has made a mistake, the user has a chance to intervene before it has an effect.”
If the label is wrong, that sentence is the last moment anyone can say so.