I've spent the last few months building a prompt library for my own marketing work, and the biggest lesson wasn't finding "secret" prompts — it was fixing the structure.

Here are 3 changes that made the single biggest difference in output quality:

1. Role + Constraint, not just Role

Everyone knows "act as a [X]" prompting. But without a constraint (word count, tone, audience sophistication), the model defaults to safe, generic language. Add: "Write for someone who already knows the basics — skip the 101 explanation."

2. Give it the failure mode, not just the goal

Instead of "write a cold email," try "write a cold email that doesn't sound like a cold email — no 'I hope this finds you well,' no generic flattery." Naming what to avoid works better than only naming what you want.

3. Ask for 3 variations, not 1

Single-shot prompting gets you the model's "average" answer. Asking for 3 distinct angles (e.g., "one direct, one story-driven, one contrarian") forces actual creative range, and you pick the best one instead of settling.

Small structural changes like these compound fast once you're running dozens of prompts a week for content, outreach, or research. Happy to share more of what's worked if people find this useful.