You need about five patterns, not a discipline. The gap between a bad result and a good one is almost always missing source material rather than missing technique: a model cannot know your customers, your numbers, or your voice unless you supply them. Fix the inputs and most prompting problems disappear.
Why the output is generic
The default output is roughly the average of everything written on a topic, and average is precisely what fails to persuade.
That is not a flaw to be prompted around. It is what the model has, absent anything from you. “Write a landing page for a bookkeeping service” has exactly one reasonable answer in the absence of information, and it is the answer everyone else gets too.
So the useful move is not a better instruction. It is more evidence.
The five patterns
1. Give examples, with evidence attached
The highest-leverage thing you can do. Not “write in our brand voice” but three real pieces that performed, with their numbers, and an instruction to match the register.
This changed my social output more than anything else. Asking for “a viral hook” produces a parody of one; supplying openings that actually did over a million plays, with the mechanic behind each, produces something usable. Pattern application against supplied examples is near the model’s best skill; pattern invention from a vague brief is near its worst.
2. Constrain the shape
Word count, sentence count, forbidden phrases, required elements. Models drift toward comfortable lengths and comfortable structures, and a constraint is a cheap way to stop that.
“Four sentences, under 110 words, no adjectives before nouns in the first line” produces something quite different from “write a short email.”
3. Make it a critic, not just a writer
Underused and unusually effective. Give it your draft and a stated criterion, and ask which parts fail.
Models are markedly better at judging against a criterion you supply than at generating one. “Here is my draft and here are seven tells of AI-sounding copy — which did I commit?” gets you a genuinely useful edit list. “Make this better” gets you a rewrite that is different rather than improved.
4. Ask for the plan before the artefact
On anything substantial, get the outline or the angle first. Correcting an approach costs one line; correcting 1,200 finished words costs a rewrite, and you will be tempted to keep paragraphs because they already exist.
5. Ask for variations, then choose
One output invites you to accept or reject. Five outputs invite you to compare, which is a much better decision. This is especially true for subject lines, headlines, and hooks, where the difference between candidates is obvious side by side and invisible alone.
What to supply, by task
| Task | Supply | Or you get |
|---|---|---|
| Landing page copy | Real objections from real calls | Generic benefit claims |
| Cold email | A specific observation about them | “Love what you’re building” |
| Social hook | Openings that performed, with numbers | A parody of a viral post |
| Ad copy | Competitor ads that ran a long time | The category average |
| Brand voice | Three pieces you are proud of | Corporate neutral |
| Positioning | Who you lose to and why | “Innovative solutions” |
The pattern: in every row, the input is something only you have. That is the actual work, and no prompt substitutes for it.
Things that do not help
- Elaborate role-play preambles. “You are a world-class copywriter with 30 years of experience” does very little on a modern model. Say what you want instead.
- Politeness and threats. Neither improves output.
- Enormous prompt-library templates. Good for discovering a format, poor as production assets, because anything that works is specific to your inputs.
- Asking for “viral” or “high-converting.” The model has no feedback loop on either. Describe the mechanic you want instead.
The verification habit
The one that matters most in marketing specifically: a model will invent facts about your business and your customers with complete confidence.
Statistics, case study details, customer quotes, competitor pricing — all of it can arrive fluent and false. In marketing copy that is not an inconvenience, it is a claim you are publishing.
The rule I hold to: every number and every named claim in published copy has to trace to something real. If I cannot point at where a figure came from, it does not ship. That is also why the pieces on this site that carry numbers name their method — so you can check them.
FAQ
Do marketers need prompt engineering?
About five patterns, not a discipline. Most bad output is a missing-input problem.
What is the biggest single improvement?
Real examples of what you want, with evidence attached. Application beats invention.
Why does AI copy sound generic?
The default is the average of everything on the topic, and average does not persuade. Specificity comes from you.
Should I use a prompt library?
For discovering formats, yes. For production, no — anything that works is specific to your inputs and voice.
Related: using AI for social content, cold email that doesn’t read like AI, and AI content that ranks.