On 9 October 1805, aboard HMS Victory off Cadiz, Horatio Nelson finished his plan for destroying the combined French and Spanish fleet. It filled two sheets of paper, and it was precise.
Forty ships a side. Two British lines of sixteen, plus an advanced squadron of eight. His second in command, Cuthbert Collingwood, would break through about the twelfth ship from the enemy’s rear. Nelson’s own line would go through the centre, with the advanced squadron cutting in just ahead of it to get at the enemy’s commander-in-chief.
Then the man who had counted every ship wrote: “Something must be left to chance; nothing is sure in a Sea Fight beyond all others. Shot will carry away the masts and yards of friends as well as foes.”
A few paragraphs on came one more sentence: “in case Signals can neither be seen or perfectly understood, no Captain can do very wrong if he places his Ship alongside that of an Enemy.”
On 21 October, off Cape Trafalgar, the numbers were wrong. Nelson had 27 ships, and 33 had sailed from Cadiz. At 1.15pm he was shot, and he died at 4.30pm, an hour before the firing stopped. The arithmetic had not survived the day. The sentence about signals was written for days like this.
Tell people what the plan is for, and they can still act when the plan breaks.
Lemons for a tart
Nelson’s captains had not met that sentence cold. According to Royal Museums Greenwich, he laid out his tactics to them over dinner on 29 and 30 September. When he explained what he called “the Nelson touch,” he told Emma Hamilton, “it was like an electric shock. Some shed tears, all approved.” Signed copies of the memorandum then went to every captain.
In the account of the naval historian William James, each captain reaching the enemy line acted on the memorandum and attached himself to “the first Frenchman or Spaniard that crossed his path.” As enemy ships kept shifting position, James judged that latitude “particularly beneficial.” Greenwich credits the British advantage to both Nelson’s planning and better-trained gunners.
Today’s US Army doctrine on mission command, ADP 6-0 (2019), calls this commander’s intent: “a clear and concise expression of the purpose of the operation and the desired military end state,” meant to help subordinates act “even when the operation does not unfold as planned.” Its length test: “If it is longer than a brief paragraph it is probably too long.” Soldiers two echelons down should be able to remember it. When possible, commanders deliver it in person, so people can ask questions on the spot. Nelson’s dinners were that kind of delivery.
A shopping list says two lemons, and the shop has none. If you know they are for a tart, you can come back with limes. If all you have is the list, you come back empty-handed. The list is the plan. What the lemons are for is the intent.
More detail, worse results
The obvious repair for a plan that keeps breaking is a more detailed plan. The military historian Stephen Bungay, who spent 17 years at Boston Consulting Group, argued in The Art of Action (2011) that this usually backfires. He describes three gaps that open between plans, actions and results. Leaders typically answer them with more information, instruction and control, and in his account that makes things worse, because it hampers the people who have to act.
His alternative, which he calls directed opportunism: make no plan more detailed than the situation allows, state the intent as clearly as you can, and leave people free within its bounds. In a 2011 Harvard Business Review article, his briefing starts with what you are expected to do and why, and ends with the boundaries you must stay inside.
The boundaries matter. ADP 6-0 says the intent “establishes the limits within which a subordinate may exercise initiative.” Nelson’s sentence was not “do as you see fit.” It named a default, and the default pointed at the enemy.
The contract on Friday
Say you ask a colleague to get a revised contract to a client by Friday. On Friday afternoon, legal flags a clause and won’t sign off until Monday. If all she has is “by Friday,” she can send it unapproved or let the deadline slip while she tries to reach you.
Now say she also knows why. The client’s board meets Monday morning, and the deal needs to be in front of it. She might send the agreed pages with a note on the open clause, or ask whether the board could approve subject to that one change. Either could still go wrong. But she can decide without you, and you can judge her call against the purpose she was working toward.
A one-line brief
Anthropic met the same problem building Research, a Claude feature in which a lead model splits a question among helper models. Early on, the lead handed out one-line tasks such as “research the semiconductor shortage.” Helpers misread such tasks or ran the same searches as each other. One studied the 2021 car chip crisis while two others duplicated work on 2025 supply chains.
The fix reads like a briefing: each helper now gets an objective, an output format, guidance on which tools and sources to use, and clear task boundaries. Without detailed task descriptions, Anthropic wrote, agents “duplicate work, leave gaps, or fail to find necessary information.” The team also said it aimed to instill “good heuristics rather than rigid rules.”
One difference matters. Nelson’s captains had sat at his table. Anthropic’s helpers run “with their own context windows,” so what they know about the job is what the lead wrote down. The engineers added that today’s agents “are not yet great at coordinating and delegating to other agents in real time.”
A helper agent can’t come to dinner. Whatever it will need when the plan stops fitting has to be in the brief.