One clause in our shipped prompt rules had been living on an inference: before a goal-directed rewrite, restate the goal in your own words. It sounds like obvious good hygiene, the LLM equivalent of repeating an order back at a drive-through. Study AF measured it with our judge-graded protocol, and the result is the cleanest refutation in the series. Made to restate a goal it had read from the document, every model complied perfectly, ninety times out of ninety, and then lost every single decisive comparison to a model that was simply told the goal. Zero wins. The model says the goal accurately, in its own words, and then writes the wrong paragraph anyway. Where a goal comes from matters more than whether the model says it aloud, and one more obvious-looking prompt clause turns out to be theater.
The Frontier Doesn't Need the Training Wheels | barkup-bench Study AD