AI answers come back generic because the prompt describes a task instead of a situation. Add four things and the output changes immediately: who the message is for, what just happened, what the constraints are, and what shape the answer should take. That structure is the whole difference between "write a follow-up email" and something you can actually send.
Why "write me a follow-up email" fails
A prompt like that gives the model almost nothing to be specific about. It does not know your trade, your customer, the last conversation, your prices, your tone, or whether the message is going to a person who owes you money or one who is still deciding.
So it writes the average of every follow-up email it has seen. That average reads like a template, and customers can smell a template.
The fix is not a cleverer phrase or a magic word. It is information. Prompts that produce usable output are mostly context, with a short instruction at the end.
The five-part prompt structure
Use this order every time. It works in any chat tool, and it survives switching models.
- Role and audience. "You write short client emails for a residential electrician in Ohio." Not "you are a world class copywriter".
- Situation. The facts of this specific case: what was quoted, when, what the customer said, what you want to happen.
- Constraints. Length, tone, what must not be promised, what to avoid ("no exclamation marks", "do not offer a discount", "under 90 words").
- Output shape. A subject line plus three short paragraphs, or five bullet points, or one SMS under 300 characters.
- One example. A message you have actually sent that sounds like you. This single element does more for voice than any adjective.
The same prompt, before and after
| Element | Weak prompt | Working prompt |
|---|---|---|
| Role | none | You write client emails for a two person landscaping company |
| Situation | "follow up on a quote" | Quoted $4,800 for a patio job nine days ago, no reply since, install window closes in three weeks |
| Constraints | "make it professional" | Under 100 words, no pressure tactics, one clear next step, no discount |
| Shape | none | Subject line plus two short paragraphs |
| Example | none | One past email pasted in, so the voice matches |
The right-hand column is not longer because it is fancy. It is longer because it carries the facts the model cannot guess.
What can I do when I do not know what to say?
Most business messages fall into a small set of situations: quote follow-up, price objection, schedule change, late payment, bad review reply, referral ask, scope change. You write each one a few times a year and start from a blank page every time.
That is the case for keeping a prompt library rather than improvising. PromptPack Pro is built on exactly this: pick your trade (contractor, realtor, coach, freelancer) and get 50 prompts with the situation, tone and constraints already written in, plus 10 swipe files for the messages you send most. The 5 free prompts, no email required, show whether the style matches how you write before you spend $19.
Whether you buy or build your own, the shape is the same. Write the prompt once, save it with placeholders like [customer name] and [quote amount], then fill the brackets.
How long should a prompt be?
Long enough to remove ambiguity and no longer. In practice, most business prompts land between 60 and 150 words. Beyond that, you are usually repeating yourself or adding background the model does not need.
Three things to cut:
- Compliments to the model ("you are the best copywriter in the world"). They cost tokens, not accuracy.
- Backstory that does not change the output. Years in business, awards, mission statements.
- Contradictory instructions. "Be brief but cover everything" is a coin flip.
OpenAI, Anthropic and Google all publish guidance that says roughly the same thing: give context, be specific, show examples, state the format. Anthropic's prompt engineering overview and Google's prompt design strategies are both worth twenty minutes if you write prompts weekly.
Should I ask for a draft or for options?
Ask for options when the decision is yours and the cost of a bad draft is high. Ask for a draft when you know what you want and just need it written.
A practical pattern for client-facing messages: ask for three variants, each labelled with its angle (direct, warm, no-pressure), then pick one and ask for a rewrite of that one with your edits. Two turns, and you get a message that sounds like you rather than a message that sounds like a model.
Do not ask for ten options. You will not choose well, and each one will get less specific.
How do I stop the output sounding like AI?
Four fixes, in order of effect:
- Paste a real example of your writing. Voice is copied, not described.
- Ban the tells. Say plainly: no "I hope this email finds you well", no "delve", no three-item lists in every paragraph, no em dashes.
- Give a word count. Vague length instructions produce padded output.
- Ask for the message, not for a message about the message. "Write the email" beats "explain how you would approach writing the email".
Then edit. Every prompt in a pack still needs your judgement, because you know the customer and the model does not.
Where prompts should not be used
Be careful with anything that carries legal, financial or safety weight: contract terms, insurance advice, tax positions, permit requirements, medical claims. A model can draft a reminder about a deposit, but it should not be the source of what your contract says.
Two habits keep you out of trouble:
- Facts come from you, wording comes from the model. Never let a prompt ask the model to state a rule, a fee or a deadline you have not confirmed.
- Say what must not be said. Add lines like "do not promise a completion date" or "do not offer a discount" to the constraints.
This is general information, not professional advice. For contract, tax or licensing questions, check with the relevant authority or a professional in your field.
A starting library for a service business
Build these six prompts once and you will use them monthly:
- Quote follow-up, nine days after sending, no pressure.
- Price objection reply that holds the price and restates the value.
- Late payment nudge, first and second tone.
- Five-star review reply that sounds grateful rather than robotic.
- Referral request, sent after a completed job.
- Scope change message that names the extra work and the extra cost.
Each one takes about ten minutes to write well the first time. After that, filling in brackets is faster than typing a fresh prompt, and the output stops sounding like everyone else's.



