What it checks, and what it cannot
It checks structure: patterns that reliably correlate with wasted tokens or unreliable output, drawn from the failure modes that recur across production prompts.
It cannot check correctness. A prompt can pass every check here and still ask for the wrong thing. Treat it as a review checklist rather than a verdict, and evaluate output separately.
The three findings worth acting on first
Missing output format is the most consequential. It is the most common cause of a parser failing in production, and specifying the exact shape you want costs almost nothing.
Boilerplate persona text is the most common. "Helpful, harmless and honest" describes every instruction-tuned model already; it changes nothing and is billed on every request forever.
Stacked tasks joined by "also" is the most damaging to quality. Splitting into separate calls almost always outperforms asking for three things at once, and often costs less because each call can use a cheaper model.
On prohibitions
The check flags heavy use of "do not" and "never". The reason is that models follow positive instructions more reliably than negative ones — stating what to do gives the model a target, while stating what not to do leaves the space of acceptable answers undefined.
Prohibitions are not banned, and some are necessary. Group them in one place rather than scattering them, and prefer a positive statement wherever one exists.