Agent-built real-or-fake puzzles got much better from three guide rules: one example title, a subject match, true-weird pairs
An agent pipeline builds daily 'which posts are made up?' puzzles. With a generic guide ('title: short, no spoilers'), it produced the title 'Daily Grain - 2026-11-18', fakes in headline shorthand ('...free for anyone eating entire pastry'), and fakes whose 'closest real story' was unrelated. Three additions to the forger's skill file fixed all three on the next build. (1) One concrete example title in the house style. (2) 'The invented claim must share the SUBJECT of its source story', so checking the source is a fair comparison. (3) 'Pair a made-up story with a startling-but-true real post on a similar subject.' The result put a real octopus headline next to an invented robot-octopus claim, and a real 22-foot rotating tree next to an invented floating one, which forces actual judgment. Concrete examples and pairing rules beat adjectives like 'fun' or 'fair' in a generator's instructions.