In 1990, a Stanford psychology graduate student named Elizabeth Newton sat pairs of undergraduates back to back. One was the tapper, the other the listener.
The tapper picked three tunes from a list of 25 well-known songs: “Yankee Doodle,”“Twinkle, Twinkle Little Star,”“Silent Night.” The job was to tap out the rhythm of each with a finger. The listener’s job was to write down the name of the song.
After each tune, the tappers were asked a question. What are the chances your partner got that one?
Their answers averaged 50%. The most pessimistic tapper said 10%. The most hopeful said 95%. These were simple songs, and the tappers had just played them.
Then Newton counted what the listeners had written. Out of 120 tunes, they had named three.
A tap is just a tap
That is 2.5%, below the lowest guess any tapper made.
Newton’s explanation, in her dissertation, is about what was going on inside the tapper’s head. While you tap “Yankee Doodle,” you are hearing it: the melody, the phrasing, maybe a voice singing along. You can’t switch that off. What you hear as a held note, she wrote, is to your listener “merely an absence of tapping.” And “a tap, outside of the vivid musical context into which you have incorporated it, is just a tap.”
Once you know something, it’s hard to imagine not knowing it, so you explain too little.
Hearing the taps from the outside didn’t cure it. When the experimenter tapped each tapper’s first tune back to them, they still guessed that about half of an audience of 100 would name it.
Economists had already given the effect its name. In 1989 Colin Camerer, George Loewenstein and Martin Weber wrote that “better-informed agents are unable to ignore private information even when it is in their interest to do so,” and called this “the curse of knowledge.” In their experiments it was in people’s interest to think like the less-informed side, and they still leaned on what they knew. In market experiments the bias shrank by about half. It did not go away.
In 2006 Chip and Dan Heath carried the idea into the boardroom, to explain why strategy slogans like “Unlocking shareholder value!” fall flat. When top executives speak abstractly, they wrote, “they are simply summarizing the wealth of concrete data in their heads.” Frontline employees, who aren’t privy to the meaning underneath, “hear only opaque phrases.”
The new hire nods
Picture a finance manager, eleven years at the company, teaching a new hire the month-end close. “Accruals go in before the flash. Unless it’s a Brightline invoice, then you wait for Dana.” To her, each phrase is a whole song: the report the CFO reads first, the supplier who always bills late, the colleague in Lisbon who signs off. To him, it is three knocks on a table. He nods, because he has picked up enough to nod.
She asks if it makes sense. He says yes. In Newton’s study, the number furthest from the truth was the tappers’ own sense of how clear they had been.
She can’t fix that by trying harder to imagine his view; the tappers who heard their song tapped back still couldn’t. What she can do is let the listener speak. “Walk me through Friday as if I weren’t here.” (Asking for the plan back is the heart of a backbrief.)
He starts out fine. Pull the invoices, book the accruals. Then he slows down. The accruals go in before the flash, he says, and stops, because he has no idea what the flash is or who reads it. The Brightline invoice he would book with the rest. He remembers there was something about Dana, but not what.
None of that showed when he nodded. It shows now because he has to use each word himself, and he stalls on exactly the ones that had a tune only in her head.
She fills in each gap and asks him to go again. The second time through, he gets to the end of Friday without stopping.
Don’t expect it to just know
A language model is a listener in a stranger position than the new hire.
Anthropic’s prompting guide asks you to think of Claude as “a brilliant but new employee who lacks context on your norms and workflows.”
Brilliant and new is a precise pairing. Brilliant covers the general: what an accrual is. New covers the local: who Dana is, and why you wait for her. The curse bites on the local part, because to you it feels too obvious to say.
The guide’s remedy is the one Newton’s tappers lacked: an outside ear. “Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they’d be confused, Claude will be too.”
Where the two listeners differ is what carries over. The new hire who learned about Dana on Thursday still knows on Friday. Claude Code’s documentation says each session “begins with a fresh context window.” What you explained in yesterday’s conversation isn’t in it, unless it was saved somewhere the new session loads. The same page names two such places, both loaded at the start of every session: CLAUDE.md files, which you write, and auto memory, notes Claude saves for itself from your corrections and preferences when it judges them worth keeping. The docs call CLAUDE.md “the place you write down what you’d otherwise re-explain,” and one of their tests for adding a line is whether “a new teammate would need the same context to be productive.”
Dana reaches the next session only if she made it into one of them.
At Stanford in 1990, the tappers heard a full song with every knock. Behind each of them, back to back, someone heard knocking and wrote down a guess.