The Curse of Knowledge. Why experts explain badly, and prompts fail the same way

In 1990 a Stanford graduate student, Elizabeth Newton, asked people to tap the rhythm of well-known songs on a table while a listener guessed the tune. Tappers predicted listeners would get about half. Listeners got roughly one in forty.

The tappers heard the melody in their heads. The listeners heard knocking. Neither could feel the other side.

Colin Camerer, George Loewenstein and Martin Weber named the effect in 1989: better-informed people struggle to reason as if they knew less, even when it pays them to.

Steven Pinker later called it the main cause of bad writing. Experts skip steps they no longer notice, use terms they forgot they had to learn, and leave out the one example that would make it click.

Writing for a model is writing for a listener who hears only knocking. You know the codebase, the customer, the goal and the ten things you already tried. The model knows the prompt.

Most bad agent output I see is not a model failure. It is the curse of knowledge, transferred.

Test on a real outsider, not a colleague. Show a concrete example before the abstraction. Write the step you think is obvious, because it is obvious only to you.

The curse does not lift with effort. It lifts with feedback. You find out what you failed to say only when someone fails to understand it.