Nice breakdown. One thing that helped me with the RISEN-style cold email example: add a "don't" line and a source line. For instance, "Use only the facts I paste below about the prospect; if something isn't there, leave a [MISSING] placeholder instead of guessing." That stops the model inventing a pain point or a stat, which is the most common way these emails go wrong. Also, rather than asking for one result, ask for three variants with a one-line reason for each, then pick and edit. Finally, keep a short "good outputs" note with two or three emails you actually sent and liked, and paste them as examples next time. Tweaking one variable per run, as you suggest, makes it easy to see which change helped.