In six logging operations in Oklahoma, every load began at the landing. A full legal load could be anywhere from 60 to 120 logs, every tree a different length and thickness, so how many went on, and which ones, was a judgment call. Then a driver, paid by the hour, took the load to the mill to be weighed.
By Edwin Locke and Gary Latham’s later account, the trucks were arriving at about 60 percent of the legal allowable weight. Every light load was a trip with room to spare.
Latham, whose 1975 paper lists the Weyerhaeuser Company as his address, and J. James Baldes were putting Locke’s theory of goals to work. They gave the drivers a specific, hard target for the net weight of their loads.
Ninety percent
The paper reports that performance “improved immediately upon the assignment of a specific hard goal.” Across 36 trucks, loads rose from about 60 to 90 percent of the legal weight, by Locke and Latham’s account. The company’s cost accounting put the price of the same increase, bought with more trucks, at about a quarter of a million dollars.
“Do your best” lets people define best. A specific, hard target usually gets more.
Locke and Latham explained why in a 2002 review of 35 years of research. A goal to do your best has “no external referent,” so each person defines it, and a wide range of results all count as acceptable. In meta-analyses they cite, specific, difficult goals beat urging people to do their best, with effect sizes (d) from .42 to .80. “In short,” they wrote, “when people are asked to do their best, they do not do so.”
“Drive carefully” leaves the speed up to each driver. A sign that says 50 does not.
Locke set out the theory in a 1968 paper, and critics had asked whether it held in industry. Latham took it into the woods. His 1975 title puts “practical significance” in quotation marks, and the paper says why: statistical significance “is only one step.” In 1990 the two men published A Theory of Goal Setting and Task Performance.
The study compared the trucks with their own record, not with trucks left without a goal. And the theory comes with conditions. One of them was waiting at the mill.
What the drivers already knew
The 2002 review lists them: goals work best when people are committed, which matters most when the goal is hard, and performance levels off at the limits of ability. People need feedback: “If the goal is to cut down 30 trees in a day,” Locke and Latham wrote, “people have no way to tell if they are on target unless they know how many trees have been cut.”
The task matters too. In a meta-analysis the review cites, the edge of specific, difficult goals over do-your-best was .77 on the least complex tasks and .41 on the most complex. In an air traffic control simulation, a performance goal got in the way of learning, and people did better when asked to do their best. Locke and Latham’s remedy is a specific learning goal, such as discovering a set number of strategies.
The truck drivers, Latham later wrote, “drew upon the knowledge they already possessed.” In a study of his on a class-scheduling task that was new and complex for the participants, the same kind of goal hurt performance.
In 2009, Lisa Ordóñez and three colleagues pushed further in Goals Gone Wild. Most goal research, they wrote, ran in “simple, well-specified domains,” counting things like anagrams completed, and managers may have trouble telling when a task is complex enough to call for a learning goal. The log trucks sat at the simple end: one number, read off a scale after every trip.
Forty accounts in week three
Imagine you hand a new account manager 40 customer accounts in her third week. “Do your best on renewals” would leave good up to her, so you write a number: 36 of 40 renewed by the end of the quarter.
It is specific and hard. But she doesn’t yet know why customers leave this product, and a renewal target with no method in hand points at the quickest lever she has, a discount.
So the first month gets a learning goal instead: talk to ten customers who nearly cancelled last year and write down what each one needed. In month two, list in hand, the 36 goes on the board. It is just as hard. Now she has a way to reach it.
The test before the code
A problem statement tells a coding model what to build. Runnable tests can add more concrete feedback on whether it is done. In a 2024 study, Noble Saji Mathews and Meiyappan Nagappan gave GPT-4 Turbo short programming problems from standard benchmarks. It tried each from the statement alone; where that failed, it tried again with the benchmark’s public tests in the prompt. Every answer was checked against much larger hidden test suites.
On one benchmark, MBPP, 69.67 percent of problems were solved from the statement alone, and 82.45 percent once the retries with tests were counted.
Only the failures got a second attempt, so the gain mixes the tests with the retry. The tests also carried information: by the authors’ account, they exposed wrong function signatures and missed edge cases. These were function-sized puzzles, not a codebase, and nothing here tests whether a harder bar matters.
Before you ask an agent for code, write a few tests that say what done means, and keep a few back to run yourself afterward. If the shown tests pass and yours fail, the code may fit your examples rather than the requirement; look closer, and run it more than once before you trust a pattern.
The Weyerhaeuser drivers had a scale at the end of every trip. By Locke and Latham’s account, they modified their trucks so they could estimate the weight before they drove to it.