Commissioned for this museum · after Brown et al., 2020

Show, Don't Tell

English·2020·25 lines·469 bytes

Curator’s note

The label set is never defined. It is demonstrated, five times, and the demonstration is the whole instruction.

A prompt written the other way would list the categories and explain each one, and it would be longer and worse. "Refactor: a change that alters structure without altering behaviour" is a definition somebody has to interpret. Extract the retry logic into its own module is an example somebody can pattern-match against, and pattern-matching is what the machine on the other end is built to do.

Notice the trailing Kind: on the last line. It is not decoration. The examples establish a shape — message, newline, kind — and the unfinished final line puts the reader in the middle of that shape with only one plausible way out. The technique is older than any model: it is the same reason a form has boxes rather than a paragraph asking you to describe yourself.

Notice too what the examples are chosen for. They are not five features. They cover the whole label set once each, in the order the list gives them, with no overlap and no ambiguous case. Examples that all point the same way teach nothing except that direction.

This is the only room in the museum whose works cannot be run and checked. The build reports them as unverified and that is the honest description: the same prompt produces different words every time, and a different answer again on a different machine. What is exhibited is the technique, which is stable, not the output, which is not.