Why most sales prompts give you mush
You type "write a follow-up email to a prospect" and get back six paragraphs that open with "I hope this email finds you well". The AI had nothing to work with, so it wrote the average of every follow-up it has seen.
The fix is giving the AI what only you know, asking for one thing, and reading the result like an editor. That works in ChatGPT, Claude, Gemini or the assistant your company allows.
If you want the bigger picture first, AI in sales covers where AI helps a team and where it gets things wrong.
What to put in a prompt
A prompt that works has four parts:
- Who you are and who you sell to, in one line.
- The one job you want done.
- Your facts: what they said, what you sell, what happened on the call.
- The shape you want back: how long, what tone, what to leave out.
Anthropic's own prompting guide puts it well: think of the model as a brilliant but new employee who lacks context on your norms and workflows. Its test is a good one for sales prompts too. Show the prompt to a colleague who knows nothing about the deal, and if they'd be confused, the AI will be.
Here's the same job asked two ways.
Weak: Write a follow-up to a prospect.
Strong: I sell office cleaning contracts to clinics. I met Dr Lim today. She said the current cleaner misses the weekend and she wants a trial in March. We agreed I'd send a quote by Thursday. Write a follow-up under 100 words that repeats her problem in her words and confirms Thursday. No "hope this finds you well".
The strong version takes a minute longer to write and saves you five minutes of rewriting. Most of that minute is facts you already have in your notes, so paste the notes rather than typing from memory.
Which jobs suit AI, and which don't
AI is quick at turning your facts into words and slow to admit it doesn't know something. So the jobs that suit it are the ones where you hold the facts, and the jobs to avoid are the ones where it would have to supply them.
| Job | Good fit? | What you give it | What you check |
|---|---|---|---|
| Summarise a company before a call | Yes, from text you paste | Their website and news page | Every claim against the page |
| Prepare call questions | Yes | Who they are, why they booked | That the questions fit this buyer |
| Follow-up after a call | Yes | Your notes and the agreed next step | Dates, names and promises |
| Rehearse objections | Yes | Your product, price and buyer | That the answers sound like you |
| Reply to a WhatsApp message | Yes, kept short | Their message and your rough points | Length, tone, their question answered first |
| Friday review of your numbers | Yes | Your counts for this week and last | The arithmetic |
| Facts about a company it hasn't been given | No | Nothing it can check | It will fill gaps with things that sound right |
| Prices, terms or delivery dates | No | Only you know them | Write these yourself |
Research is the row most reps start with, and AI for sales prospecting turns it into a weekly routine.
Monday to Friday: one prompt for each job
Fill the brackets with your own facts. If your week's problem is the calendar rather than the words, copy a prompt from the AI prompts for sales template instead: it covers clearing an afternoon, blocking focus time and finding a slot for a 1:1.
Monday: research the week's calls
I sell [what you sell] to [type of buyer]. Here's what I found on [company]'s website and news page: [paste]. In five bullet points, tell me what they do, what's changed in the last six months, and one problem they're likely to have that I can help with. Put a question mark after anything you're guessing at.
The question marks matter most. Ask for them and you'll get them, and you'll know which bullet to check before the call.
Before each call: questions and a rehearsal
I have a 20-minute first call tomorrow with [role] at [type of business]. They booked because [reason]. Give me six open questions that get them talking about how they do it today and what it costs them, in plain language with no jargon.
Play the buyer. You're a [role] at a [type of business] who's busy and a bit sceptical. I'll pitch you [product] in two minutes. Push back the way a real buyer would, then tell me which part of my pitch was weakest.
Ten minutes of rehearsal before an important call costs nothing and shows you where you ramble. Check its pushback against the real objections in common sales objections and your own lost deals, because AI is good at listing objections and weak at knowing your buyer.
After each call: the follow-up
I just had a call with [name] at [company]. What we discussed: [three bullet points]. What we agreed: [next step and date]. Write a follow-up email under 120 words that repeats their problem in their words and confirms the next step, without "hope this finds you well" or "just circling back".
For WhatsApp, say so in the prompt and cap it at 50 words, or you'll get an email pasted into a chat.
Friday: the review
Here are my numbers for the week: [calls, conversations, meetings, proposals, deals]. Last week: [same]. My weekly target: [numbers]. In five lines, tell me where I'm behind, which step lost the most people, and one thing to change next week.
To pick which numbers to count, sales activity metrics lists the six that predict revenue. Check the AI's arithmetic before you act on it; models make simple maths slips with confidence.
The Friday review works better when the AI can see your week. ActivityTracker's Ask AI answers questions about your day from your own tasks, goals and calendar, and one tap turns the answer into tasks, a goal or time on your timetable.
See Ask AIHow to check what comes back
Treat every output as a draft from a new colleague who's quick but doesn't know your clients. Before anything goes out, run five checks:
- Check every name, number and date against your own notes, because AI fills gaps with things that sound right.
- Look for anything it promised on your behalf, like a start date, a discount or a cheerful "we can do that".
- Make sure it answers what they actually asked, and answers it first.
- Cut it by a third.
- Read it out loud. Anything you wouldn't say on a call, you shouldn't send in writing either, so change it before it goes.
You can make the AI do the first two checks on itself:
Check this draft for anything that's a guess, a promise I can't keep, or a number I haven't given you, and list each one: [paste]
Anthropic's page on reducing hallucinations suggests the same moves: give the model explicit permission to say it doesn't know, and have it back each claim with a quote from what you gave it. A claim it can't back with a quote comes out.
Where prompts go wrong
Most bad output traces back to one of six habits:
- Asking for facts instead of giving them. "What's new at [company]?" gets a confident answer it can't support. Paste the page instead.
- Five jobs in one prompt. Research, email and objections in one go gets you a weak version of each. Split them.
- A long chat that drifts. After twenty messages it forgets your first instructions. Start a fresh chat for each new job and paste the four parts again.
- Asking it to agree. "Is this a good email?" gets a yes. Ask "What would make the buyer ignore this?" instead.
- Trusting its sums. Totals, percentages and conversion figures need checking with a calculator.
- One message for a whole list. A prompt that writes the same text for 40 people writes 40 messages nobody answers. Write per person, from their own words.
Keep a prompt library
The prompts that work for you are worth more than any list, because they carry your buyer and your voice. Keep them in one note:
- The prompt, with the brackets left in.
- The output you actually sent, after editing.
- The date, and whether it got a reply.
Once a month, delete the ones you stopped using. If you run a team, share the note so a new rep starts with ten prompts that already sound like your business. Add the edit prompt below to the top of it:
Rewrite this so it sounds like me. Here are two messages I've written myself: [paste]. Match the sentence length and the words I use. Remove anything I wouldn't say.
What never to paste in
Before you paste anything, check what your company allows and what your AI tool does with what you type. In ChatGPT, OpenAI's data controls page says that when "Improve the model for everyone" is off, your new conversations won't be used to train its models, and that by default it doesn't train on content from ChatGPT Business, Enterprise or Edu workspaces. Other tools have their own settings, so read yours.
Even with training switched off, keep these out:
- Clients' ID numbers, bank details, health information or passwords.
- Full chat or email threads with phone numbers, addresses and names. Summarise them in your own words instead.
- Contracts or anything marked confidential, unless your company has approved the tool for it.
- Anything a client told you in confidence.
A simple habit covers most of it: write "the client" instead of the name, and describe the situation instead of pasting it.
Made-up example: a rep who sells office cleaning contracts uses AI prompts for one week.
Monday: she pastes her list of 40 prospects and asks for a sort into "call this week", "later" and "not a fit". It puts two clinics in "not a fit", because she forgot to say clinics are her best customers. She adds that line and runs it again.
Wednesday: the follow-up prompt writes her email in seconds. It includes a start date the client never agreed to. She catches it on the read-through and adds the check prompt to her routine.
Thursday: a buyer who'd been quiet for eight days replies to a two-line WhatsApp follow-up she drafted from his last message.
Friday: she pastes her week's counts into the review prompt. It points out that she booked 6 meetings from 55 calls last week and only 3 from 60 this week, which she hadn't noticed, so next week she moves her calls back to the mornings. She checks the counts herself before she changes anything.
She keeps five prompts in a note on her phone and drops the rest.
What to do first
- Pick the one job you most dislike: research, follow-ups or the Friday review.
- Write your "who I sell to" line once.
- Use one prompt for that job for a week on real work, and run the five checks on every output.
- Save the prompt and the version you sent. Delete prompts that didn't save time.
- Add the next job the week after.
Keep reading
Questions people ask
The ones that give the AI your facts and one job, such as a follow-up built from your notes or a weekly review from your own numbers. A generic prompt gets a generic answer back.
Say who you sell to, name the one job, paste the facts it needs and describe the shape you want back, such as "under 100 words, no jargon". Then check the result before you use it.
Anything where it would have to supply the facts: news about a company you haven't pasted in, prices, terms and delivery dates. Write those yourself, and use AI for turning your own facts into words.
Yes, as a first draft. Give it what the buyer said and the agreed next step, then check every fact and rewrite it so it sounds like you.
Client ID numbers, bank or health details, passwords, full chat threads with names and numbers, and anything confidential. Check your company's rules and your tool's data settings first.
Sources
- Anthropic, Claude docs, Prompting best practices: be clear and direct, read 27 October 2026
- Anthropic, Claude docs, Reduce hallucinations, read 27 October 2026
- OpenAI Help Center, Data Controls FAQ, read 27 October 2026

