In the fall of 2009, a large utility company in the American Midwest mailed a flu shot reminder to 3,272 employees, everyone the vaccine was recommended for. Each letter listed the times and places of free clinics at their workplaces.
Some letters carried one extra line: “Many people find it helpful to make a plan for getting their shot. You can write yours here:”. Below it, some had a blank for a date. Others had blanks for a date and a time.
Nobody collected the plans or checked whether anyone filled them in. The researchers, led by Katherine Milkman, counted who got a shot at a clinic.
Without the line, 33.1 percent got the shot. With a blank for a date, 35.6 percent, a gap the study couldn’t tell from chance. With blanks for a date and a time, 37.1 percent.
A moment, not a destination
Tie an action to a moment you name in advance, and you’re more likely to start it.
The psychologist Peter Gollwitzer named this kind of plan in 1993 and laid out the idea in a 1999 paper in American Psychologist, “Implementation Intentions: Strong Effects of Simple Plans.” A goal intention, he wrote, says “I intend to reach X.” An implementation intention has the form “When situation x arises, I will perform response y.”
A goal says you’ll call your mother more often. An if-then plan says that when the train pulls out on Sunday evening, you’ll call her. The second one has a place in the week to live.
In Gollwitzer’s account, choosing the moment in advance makes it easier to notice when it comes, and the response starts with less deliberation. His phrase for it, from that 1993 work, was “passing the control of one’s behavior on to the environment.” That may be why the time mattered on the flu letter: a date names a day and still leaves you to find the moment, while a date and a time name the moment itself. The flu study itself didn’t test why.
The paper also showed where plans add little. In one study, students named two projects for the Christmas break, one hard to get started on and one easy, and said whether they had planned when and where to start. For hard projects, about two-thirds of the planners finished, against about a quarter of the rest. For easy projects, completion was around 80 percent either way. The students chose for themselves whether to plan, so that study shows an association. A follow-up that assigned the plan, for a report due within 48 hours of Christmas Eve, found three-quarters of planners delivered on time, against one-third of the others. When starting is easy, the plan has little to fix.
Strong effects, smaller than they looked
In 2006, Gollwitzer and Paschal Sheeran pooled 94 independent tests and reported an effect of “medium-to-large magnitude” on reaching goals: a d of .65, a standard measure of how far apart two groups’ results are.
The pool has grown, and the number has shrunk. In a 2025 meta-analysis of 642 tests, Sheeran, Olivia Listrom and Gollwitzer found a sample-weighted average d of .36. They also found that small studies with positive results were overrepresented, which met the threshold for “substantial” publication bias. One standard correction barely moved the estimate. A more conservative method put it at d = .15: an effect the authors still judged robust, but “small according to conventional guidelines” and highly variable from study to study. For physical activity, across 64 tests, the average was .14. Plans worked better when written in if-then form, when people were highly motivated, and when they rehearsed the plan.
The flu study fits that picture: a planning prompt that moved about four people in a hundred.
The review nobody starts
Imagine you lead a product team, and the post-launch review has been “next week” for a month. Your intention is fine. It just never meets a moment.
So you write one: when Thursday’s stand-up ends, I open the review document and write the first section before I look at chat. The first Thursday, you write a page. Three weeks later the stand-up moves to Friday, the old cue disappears, and so do the pages. A plan is only as reliable as the moment it hangs on, so you pick one that recurs.
When the cue is a file edit
Coding agents have the same gap between a goal and a moment. Claude Code’s documentation says the instructions you put in a CLAUDE.md file are “context, not enforced configuration.” Claude “reads it and tries to follow it, but there’s no guarantee of strict compliance, especially for vague or conflicting instructions.”
For things that must happen at a specific point, such as before every commit or after each file edit, the docs recommend a hook: a shell command the tool runs at a fixed event, which applies “regardless of what Claude decides to do.” The docs’ example runs a code formatter on every file Claude edits.
Imagine you want every edited file formatted. “Keep the code formatted” in CLAUDE.md is a goal intention. A hook that fires after each edit is an if-then plan, with one difference: your Thursday plan still depends on you noticing Thursday, while the hook is run by the tool, not the model. It covers only the event you named, though. A hook that runs after an edit can’t stop the edit from happening first, and a written rule with no tool enforcing it still depends on the model following it.
You can check which kind of rule you’ve written. Look through a few recent sessions for whether each CLAUDE.md rule was followed every time. For rules that must never be skipped, write a hook, then confirm it ran by its effect, a reformatted file or a line in a log, not by what the agent says it did.
Back at the utility company, nobody ever saw a single filled-in blank. All anyone has is the count at the clinic door.