ChatGPT sales prompts and scripts that work

A ChatGPT sales prompt works when it carries the facts, the goal and a ban on inventing details, and when you test its output on a small group. Seven prompts cover most needs: an observation opener, a follow up, an objection reply, a call brief, a summary of a thread, a breakup note and a rewrite in another language. On our own LinkedIn outbound, observation openers beat question openers by 10 to 15 points.

"Scripts that work" is a bold promise, and most pages that make it cannot back it up. We will not promise that a prompt gives you a winning script. We can give you prompts that produce usable drafts, a way to test them, and the one pattern that our own data supports.

Why do most ChatGPT sales prompts disappoint?

Because they ask for too little. "Write a cold email for my software" gives text that could be sent by anyone to anyone. It has no fact about the person, so there is nothing to react to, and the recipient sees it for what it is.

The model is not the problem. It writes what it is asked to write. A prompt that supplies the person's role, something they did, your goal and your rules gets a message that fits. A prompt with none of that gets filler.

The other failure is trust. A model asked for a fact it does not have will supply a plausible one. A message that cites a funding round that never happened costs you the reply and your credibility.

Which seven prompts are worth saving?

Write each once, keep it in a note, and change the details.

1. The observation opener. "Write a LinkedIn message of under 60 words to [name], [title] at [company]. Use only the text below. Open with one specific observation about their work. Then say in one sentence why I am writing: [reason]. End with one easy question. No flattery, no pitch, no invented facts. If the text gives you nothing to observe, say so."

2. The follow up with a new angle. "Here is my first message and the thread. Write a follow up under 40 words that adds something new: a result, a question or a useful link from the list below. Do not say 'just checking in' or 'bumping this'."

3. The objection reply. "The prospect wrote: [text]. Write two replies under 40 words. The first accepts the answer politely. The second asks one question that could change it. Neither argues."

4. The call brief. "From the notes and thread below, list what the person cares about, what they have not said, and five questions for the first ten minutes. Quote the line each point comes from."

5. The thread summary. "Summarise this thread in five lines for a colleague who has to call this person tomorrow. Include the last thing they asked and anything I promised."

6. The breakup note. "Write a last message under 40 words. Make it easy to say no and easy to forward to the right person. No guilt."

There is a ready breakup message template and a forwardable last note if you want a base.

7. The language rewrite. "Rewrite this message in [language], the way a native speaker in business would write to a stranger. Keep it under the same length. Do not translate word for word."

What does the data say works?

One pattern, measured. On our own LinkedIn outbound, openers that state an observation out-replied openers that ask a question in every month we compared: 43.5% against 33.5% in February 2026, 45.2% against 33.6% in March, and 51.9% against 37.3% in June. That is a gain of 10 to 15 points, at equal or better meeting rates. Prompt 1 is built on it.

A caution the same data holds: only 5 of about 30,000 openers mentioned the post that put the person on the list. Most openers were observations about the profile, not about the reason the person was on the list, so there is room to do better. See observation openers beat questions.

The other published finding that matters for prompts is from the Lamothermic study: the agent wrote as the president and asked one short engineering question, never a pitch, and 293 of 879 engineers and buyers replied (33%). The question was specific to the reader's work, which is what prompt 1 forces.

These are results from our accounts, with our audiences. Test them on yours.

What does a finished prompt look like?

This is a made up example with an invented person, to show how the pieces fit. It is not a real prospect.

Input pasted below the prompt: "Dana Voss, Head of Revenue at a logistics software company. Posted last week: 'Our reps spend more time updating the CRM than talking to buyers.' Reason for writing: we help small teams keep the CRM current from their conversations."

Draft the model might return: "Dana, your line about reps updating the CRM more than talking to buyers is one I hear a lot. We work on keeping the CRM current from the conversations themselves. Is that a problem on your team too, or just a general complaint?"

Check it against the rules. It uses only the pasted text. It makes no claim about numbers. It ends in one easy question. It is under 60 words. If the model had written "I saw you raised a new round", and the text said nothing about it, you reject the draft.

How do you review a draft before you send it?

1. Every fact traces to the text you pasted. Delete any that do not.
2. Read it aloud. If you would not say it to a stranger, change it.
3. Cut a third of the words. Most drafts are too long.
4. Remove the compliments. Praise with no detail reads as a template.
5. Check the ask. One question, easy to answer in a line.
6. Check the language. If the recipient writes in another language, send in theirs.

How do you test a script?

1. Pick 100 people from one list. Split them at random into two groups of 50.
2. Send script A to one group and script B to the other, from comparable accounts and on the same days.
3. Count replies, then positive replies, then meetings per person messaged. Fifty people per group is small, so look for large differences only.
4. Keep the winner and change one thing. A new first line, not a new everything.

If you cannot reach 50 per group, read every reply instead and learn from the words.

What are the rules to put into every prompt?

• Facts only from the text I gave you.
• If a fact is missing, write "not found".
• No numbers I did not provide.
• No flattery, no "I hope this finds you well".
• A length limit.
• The language of the recipient.

These are the same rules our agents run on. The post on cold email rules we enforce lists the ones built into Sellinger's email agent.

When should you move from prompts to a platform?

When the work is no longer writing. Sending at volume needs limits per account, quiet hours, a suppression check before every send, stop on reply and a way to answer at night. Pasting drafts into LinkedIn by hand does not scale and is easy to get wrong.

Sellinger connects to ChatGPT and Claude through a hosted MCP connector with 43 tools, so you can keep a chat as your cockpit and let the platform do the sending. Writes ask first. See ChatGPT or Claude: can an assistant run your outreach for the full picture.

A checklist

• [ ] Every prompt has the facts, the goal, a length and a no inventing rule.
• [ ] I saved seven prompts and change only the details.
• [ ] I read every draft and cut a third of the words.
• [ ] I test two scripts on 50 people each before I trust either.
• [ ] I count meetings per person messaged, not only replies.
• [ ] I know where a prompt stops and a platform has to start.