On February 24, 2010, Akio Toyoda, the president of Toyota and the founder’s grandson, sat before the House Committee on Oversight and Government Reform. Six months earlier, a Lexus with a floor mat that didn’t fit had sped out of control near San Diego, and the family of four inside had died. Toyoda offered his condolences in his prepared testimony.
Then he explained what he thought had gone wrong. “Toyota has, for the past few years, been expanding its business rapidly,” he said in his prepared testimony. “Quite frankly, I fear the pace at which we have grown may have been too quick.”
Toyota’s priority, he said, had traditionally been the following: “First; Safety, Second; Quality, and Third; Volume.”
He didn’t say the list had changed. “These priorities became confused,” he said.
Second things not at all
Peter Drucker gave this problem a chapter of The Effective Executive in 1967 and called it “First Things First.” It opens with a claim: “If there is any one ‘secret’ of effectiveness, it is concentration.” Then the rule: “Effective executives do first things first and they do one thing at a time.”
Do the most important thing first, one thing at a time, and decide what you won’t do.
The hard part, by Drucker’s account, isn’t the list. There are always more good tasks for tomorrow than time to do them, so the scarce decision is the other one. He called it setting “posteriorities”: deciding which tasks not to take on. The most important thing about priorities and posteriorities, he wrote, is “not intelligent analysis but courage.” One of his rules is to “pick the future as against the past.” Executives who concentrate, as he put it, do first things first, and second things “not at all.”
Read with Drucker in mind, Toyoda’s testimony sounds like a company that hadn’t set its posteriorities. “We pursued growth over the speed at which we were able to develop our people and our organization,” he said. Volume was third on the list. By his account, growth had set the pace.
Drucker granted that a few people work differently. There was Mozart, “of course,” who “could, it seems, work on several compositions” at once.
The feature that waits
Imagine you’re a product manager at a small software company. You’ve promised sales three features this quarter. Then support escalates a sync bug that, for a handful of customers, quietly deletes edits.
Fixing it properly will take your two best engineers about three weeks. If you keep all three features and add the bug, everything ships late and the bug fix gets whatever attention is left. So you move two of the features to next quarter and tell the sales lead the same afternoon, before she hears it from a customer. She’s unhappy, and one deal slips. The bug is fixed in three weeks.
Everyone agreed the bug came first. The hard decision was writing down which two promises you would break.
Five hundred rules at once
AI agents get handed long lists too: system prompts full of rules, style guides, things to always and never do. In July 2025, researchers at the AI company Distyl posted a preprint that measured what happens as the list grows. They asked 20 models from seven providers to write a business report while including up to 500 specific keywords, each one a separate instruction.
Accuracy fell as the list grew. Even the best models included only about 68 percent of the keywords at 500. The researchers also found a primacy effect: instructions in the last third of the list were missed more often than those in the first third. That bias was strongest at middle loads, around 150 to 200 instructions, and faded toward uniform failure past 300. The authors note their test used only simple keyword instructions in English business reports, so it may not carry over to other kinds of rules.
Here the analogy needs adjusting. Drucker’s executives ran out of hours. A model doesn’t get tired, but in this test, the more rules it was given, the smaller the share it met. Putting the most important instructions first “may help,” the authors wrote, but it helps less once the list is very long. So the closer equivalent of posteriorities is cutting the list. Keep a prompt’s rules to the ones you’d check, put the non-negotiable one first, and split a long checklist across separate runs. Then check each rule in the output with a script or by hand, and repeat the run a few times, since one run can differ from the next.
A meeting in October
Four years after the hearing, Toyota admitted in a deferred prosecution agreement with the Justice Department that it had misled US consumers about two safety issues, floor mats and sticking pedals, and agreed to pay $1.2 billion.
The department’s summary of the facts describes a decision on October 21, 2009. Toyota had already sent its pedal maker an “urgent” design change for US cars. That day, engineers and the leadership of its recall decision group cancelled it, “in order to prevent NHTSA from learning about the sticky pedal problem.” Employees were told to pass the word to the plants orally, and not to put anything about the cancellation in writing.