Which prompt techniques actually improve a cheap model's documentation?

When a cheap model documents source code, two things can go wrong: it can invent things that aren’t there, or it can simply restate the signature — setUserId → “sets the user id” — adding nothing a reader couldn’t already see. Below, nine prompt techniques — each one controlled change off a stripped-down baseline — are scored on a single quality number that folds both in: an independent judge’s usefulness-beyond-restatement grade, counted only on replicates a deterministic checker finds free of hallucination, across eight documentation types. The striking result is how few techniques beat plain prompting — only few-shot edges it out, by trading a little usefulness for never inventing — pick a doc type and a technique to compare it to the baseline.