David Marquet got command of the USS Santa Fe at the last minute, after its captain quit. By his own account it was the worst-performing submarine in the fleet, and a type of submarine he knew little about.
Early on, he ran a drill. The crew simulated a reactor shutdown, and the boat was running on its backup electric motor, which can only push it along slowly. Marquet wanted to show he wasn’t going to be an easy captain, so he called for more speed: “Ahead two-thirds.”
His most senior, most experienced officer of the deck repeated the order. Nothing happened. Then the helmsman spoke up: “Captain, there is no ahead two-thirds on the EPM!” The backup motor had no such setting.
Marquet turned to the officer. Did you know that? “Yes, Captain, I did.” Then why did you order it?
“Because you told me to.”
The officer wasn’t defying anyone. He did exactly what the system had trained him to do: take the captain’s words and pass them down. “What happens when the leader is wrong in a top-down culture?” Marquet wrote later. “Everyone goes over the cliff.”
He vowed that day never to give an order again. He still had a nuclear submarine to run.
“I intend to”
So the orders started coming from the other direction. Officers stopped asking permission. They came to the captain with “I intend to…” and told him what they planned to do. Marquet’s part was to give intent: the goal they were working toward. He stopped dictating the steps.
By Marquet’s account, the years that followed went well. Retention rose from one in ten sailors to 100%, and ten men from his crew later commanded submarines of their own.
What he had stumbled into is the difference between directions and an address. “Left at the bakery, right at the second light” works fine until a street is closed. Then the driver is stuck, or keeps turning where you said and ends up somewhere you never meant. Tell them “the station, by six, and skip the toll road,” and a closed street is only a detour.
Tell people what you want and why. Let them tell you how.
Step four
Say you hand a colleague a ten-step plan for moving the team’s files to a new system, then leave for a two-day conference. At step four, she finds that finance still opens the old folders every morning. Step four says delete them. She can try to reach you between sessions, or she can delete them, because it’s on the list. Either way, the work now runs at the speed of your phone.
The Santa Fe version starts before the conference. You say what you want and why (one place for every file, so nobody works from an old copy), and what must not happen (nobody loses a file they still use). Then she tells you how she plans to do it. A misunderstanding found at that point costs a minute instead of a week.
None of this was new. In 1871 the Prussian field marshal Helmuth von Moltke wrote that no plan of operations extends with any certainty beyond the first encounter with the enemy’s main force, and the Prussian army learned to command by stating intent and leaving the how to the officers on the spot. The historian Stephen Bungay, who carried the idea into business in The Art of Action (2011), notes that leaders facing the gaps between plans and results usually reach for more detail, and that this makes things worse, because it ties the hands of the people who have to act.
There is a catch. Intent only works when people know the work well enough to choose a route. Hand the address to someone who has never driven, and the address won’t help. That is the other half of Marquet’s story, and the reason Task-Relevant Maturity sits next to this page.
The plan before the code
The same three words now show up somewhere Marquet never served: in software that writes software.
Anthropic’s advice for its coding agent, Claude Code, is to have it read the code and write a plan before it changes anything, a plan you can edit first, because “letting Claude jump straight to coding can produce code that solves the wrong problem.” That plan is the agent saying “I intend to…”
Intent travels into the instructions too. Anthropic’s prompting guide compares two versions of one rule. “NEVER use ellipses.” Or: “Your response will be read aloud by a text-to-speech engine, so never use ellipses since the text-to-speech engine will not know how to pronounce them.” The guide calls the second more effective and explains why: “Claude can generalize from the explanation.” Once the model knows a voice will read its answer, it has a reason to steer around other things a voice can’t say.
The failure looks familiar, too. The officer on the Santa Fe knew the order was impossible and passed it on anyway. Hand an agent a precise list with one wrong step in it, skip the plan, and it may carry out that step faithfully. Because you told it to.