A ten-year-old has just finished ten picture puzzles. The researcher scores the sheet, says it is a really high score, and adds one more sentence: “You must be smart at these problems.”
Then come harder puzzles, and the news that the child did “a lot worse.” Then a small request: write a note to children you don’t know, in another state, and tell them how many you got right.
In this 1998 study by Claudia Mueller and Carol Dweck, 38% of the children who had been called smart overstated their score. Among children praised instead for working hard, 13% did.
One sentence was about the child. The other was about something the child did. Adults hear the first kind at work all the time.
The wrong price
Picture a team lead hanging up a client call, annoyed. Her new analyst just quoted the client a price nobody had approved. Walking back to her desk, she has the sentence ready: you’re careless with clients.
If she says it, the next twenty minutes are about whether he is careless. He has examples to the contrary, and some of them are good. The price never comes up again.
Instead she says: on today’s call, when they asked about pricing, you gave them the 2023 rate, and now I have to call them back and correct it. What happened?
He tells her the rate came from the sales deck in the shared drive. Nobody had updated it. The careless one, if there was one, was a file, and the next person to open it would have quoted the same number.
Say what they did and what it caused, not what kind of person they are.
Three parts and a question
What she said follows a method the Center for Creative Leadership (CCL), which has trained leaders for more than fifty years, teaches in its guidebook Feedback That Works. It is called Situation-Behavior-Impact.
Name the situation precisely: “This morning at the 11 am team meeting,” not “last week.” Then the behavior. CCL’s instruction is “Describe the actual, observable behavior. Keep to the facts. Don’t insert any opinions or judgements.” Then say what it did to you.
CCL’s own example: “You interrupted me while I was telling the team about the monthly budget,” instead of “You were rude.” For the impact: “I felt frustrated when you interrupted me because it broke my train of thought.”
“You were rude” is a verdict, and all anyone can do with a verdict is appeal it. “You interrupted me during the budget update” is something a person can change by Thursday. CCL says its own research found the method lowers anxiety for the person giving feedback and defensiveness in the person hearing it.
CCL now suggests a fourth step, the one the team lead took: ask about intent. “What was going on for you?” Its argument is that we quietly build stories about why people act, and the only way to learn someone’s intent is to ask. (The ladder of inference is about how those stories get built.)
What the children can and can’t tell you
The puzzle study went further than the note. After the hard set, children praised for intelligence were less keen to keep going, and in one of the studies their scores on a third set, as easy as the first, slipped below where they had started. Children praised for effort improved.
It is also one study of fifth graders, and it has not replicated reliably. In a 2019 paper, Yue Li and Timothy Bates tested 624 children aged ten to twelve and found that praise for intelligence had no significant effect on their performance.
The broader evidence on feedback is larger and older. In 1996 Avraham Kluger and Angelo DeNisi pooled 607 effect sizes from feedback studies. Feedback helped on average, but in over a third of cases it made performance worse. Their reading of the pattern: feedback works less well the more it pulls attention away from the task and toward the self.
A bug report is feedback too
The analyst could explain the sales deck because his boss asked. When the one doing the work is an AI model, the burden of being specific sits mostly with you. Anthropic’s guide to its coding tool, Claude Code, puts it this way: “Claude can infer intent, but it can’t read your mind.”
Its advice has something in common with SBI: describe what happened, concretely, instead of labeling it. Don’t write “fix the login bug.” Write: “users report that login fails after session timeout. check the auth flow in src/auth/, especially token refresh. write a failing test that reproduces the issue, then fix it.” The guide sums this up as giving “the symptom, the likely location, and what ‘fixed’ looks like.” That isn’t SBI word for word. The guide asks for no feelings and no question about intent. What carries over is the refusal to hand over a verdict (“it’s broken”) in place of a description.
One difference shows up when corrections pile up. The guide suggests that once you’ve corrected Claude more than twice on the same issue in one session, the conversation is cluttered with failed attempts, and it’s better to clear it and start over with a more specific prompt. In its words, “A clean session with a better prompt almost always outperforms a long session with accumulated corrections.”
With a colleague, the third conversation builds on the first two. With the tool, the guide’s advice is to write a new first message, with the timeout and the failing test in it from the start.