Most ad-copy prompts fail for the same reason: they describe the product, not the reader. The patterns below all push the task back toward a specific audience and a specific emotional shift, which is where good copy actually lives.
The "before / after / bridge" prompt
Give the model three beats to fill in: what the reader is experiencing right now, what they want to be experiencing, and what single shift takes them from one to the other. Ask for five variants of the bridge — that's the part that becomes the headline.
Tip
The trap is over-specifying the "before." If you hand the model the reader's exact pain in your voice, it will just rephrase you. Describe the situation neutrally; let the copy find the tension.
The voice-sample prompt
Paste three to five examples of your existing copy that you think work, then ask the model to infer the voice rules — sentence length, vocabulary, stance — and write new copy against those inferred rules. This works better than "write in our brand voice" because the model is reasoning from concrete samples rather than an abstract guide.
The constraint list
Headlines live or die on word count. Build the constraint into the prompt: "Under 8 words. Must contain a verb. No em dashes. No questions." Constraint-first prompts produce far less slop than open-ended ones.
A/B variant generation
Once you have a winner, ask for five variants that change exactly one dimension — tone, specificity, length, angle — and label which dimension each one flexes. This gives you testable hypotheses instead of five indistinguishable rewrites.
What to still do yourself
- Verify claims. Models will cheerfully invent statistics.
- Decide the promise. AI is good at packaging a promise; it cannot make the strategic call about what you are promising.
- Read each variant out loud before you ship. Ear catches what eye misses.