How to use ChatGPT and Claude for sales, with prompts that work

Sales teams can use ChatGPT or Claude, and most sellers already use general AI tools, for researching one account at a time, preparing calls and drafting messages a person then edits.

BCG’s survey of 400 US sales representatives and managers (October 2025) found that roughly seven in ten sellers rely on general-purpose AI tools for tactical tasks such as drafting emails, summarizing calls or automating follow-ups. The editing matters: in Hunter’s State of Email Outreach 2026, manually edited emails got a 5.2% reply rate, against 4.4% for fully automated ones.

ChatGPT and Claude handle the tasks in this guide in similar ways, and both change with each release. The prompts below work in either; what matters more is the context you give and the checks you make. For the wider picture, see our guide on how to use AI in sales.

A chat assistant is good at reading, structuring and drafting from material you can check, and bad at keeping lists, guessing contact details and stating current facts without a source.

Give the assistant standing context once, so every answer starts from your offer and your rules rather than from generic sales advice.

1. Write a one-page brief: what you sell, the problem it solves, your ideal customer profile, three customer outcomes you can prove, and claims you must never make.
2. Add your writing rules: under 90 words for a first email, one question, no superlatives, US spelling.
3. Save the brief as standing instructions or in a project, so you do not paste it every time. Both assistants offer a way to do this.
4. Turn on web search when you ask about current events, and ask for a link and a date for every fact.
5. Use only the accounts and plans your company has approved for business data.

Prospecting prompts work when they ask for evidence, not opinions: dated events with links, fit judged against your written rules, and roles rather than guessed names.

1. Account research: “Research [company]. List events from the last 12 months that could affect [the problem we solve]: new executives, new sites, acquisitions, funding, job posts naming a project. For each, give the date and the link. If you cannot find a source, say so instead of guessing.”
2. Fit check: “Here is our ideal customer profile: [paste]. Here is what we know about [company]: [paste]. Is it a fit, not a fit, or unclear? For each judgment, name the rule it rests on.”
3. Buying group: “At a [size] [industry] company, which roles usually own [problem], approve the budget, and could block a purchase like ours? Explain what each role cares about.”
4. Signal list: “List events that usually come 3 to 12 months before a company buys [what we sell]. For each, say where it shows up in public sources and how long it stays relevant.”

Outreach prompts work when you supply the facts and the assistant supplies structure and wording, so nothing reaches a buyer that you have not checked.

1. First email: “Using only these facts: [paste research], write a first email to [role] under 90 words. Name the event and the problem it creates for them, ask one easy question, and do not use superlatives or claim anything not in the facts.”
2. Follow-up: “Write a 50-word follow-up to this email: [paste]. Add one new, useful detail and keep the same question.”
3. Call brief: “From these notes: [paste], write a one-page call brief: what changed at the company, who is on the call and what each cares about, three questions to ask, and two likely objections.”
4. Discovery questions: “Write eight discovery questions about [problem] for a [role]. Order them from easy to hard, and make none of them answerable with yes or no.”
5. Objection practice: “Play a skeptical [role] at a [industry] company. Raise the objections you would really have to [our offer], one at a time, and wait for my answer before the next.”
6. Call summary: “From this transcript: [paste], list what the buyer said about the problem in their own words, what we promised, and the next step with a date.”
7. Reply sorting: “Sort these replies into interested, not now, wrong person and not interested, and suggest a short answer for each: [paste].”
8. Win and loss review: “Here are our last 20 won and lost deals with notes: [paste]. What do the wins have in common that the losses do not? Quote the notes behind each point.”

A good prompt gives the assistant the facts, the reader and the limits. A weak one asks it to invent all three.

Sales teams can use Claude Code and other AI agents if someone on the team is comfortable with code. Claude Code is Anthropic’s agentic coding tool: it reads files, edits them and runs commands.

Sales operations teams use coding agents for small jobs such as cleaning and merging exported lists. The person running it still owns the checks on the output. For most sellers, a chat assistant with good prompts, or a tool built for AI prospecting, is the simpler route.

Never send what you have not checked, and never let the assistant supply facts, contact details or claims about your own product.

• Sending output unread. In KPMG and the University of Melbourne’s 2025 study (more than 48,000 people in 47 countries), 66% of employees who use AI at work said they had relied on its output without evaluating it.
• Inventing personalization. A made-up detail about a buyer’s company ends the conversation.
• Guessing contact details. Wrong addresses bounce and harm the sending domain.
• Pasting confidential deal data into tools your company has not approved.
• Letting it describe your product. It will happily promise features you do not have.
• Sounding synthetic. In Hunter’s 2026 report, 69% of US decision makers said it bothers them if AI was used to write an email, unless the result feels genuinely human.

You need more than a chat assistant once you research more accounts a week than one person can check, or once several people work the same list.

Our review of AI account research tools compares the options for the second column.

SalesOne is AI account research and qualification for B2B teams. Where a chat assistant answers one prompt about one company, S1 agents work your whole market against your written profile: they record each buying signal with its source and date, keep or reject every company with a reason, and map the people who own the problem.

In SalesOne’s production research runs in October 2026, all 375 buying signals behind kept accounts link their source, and 98% are dated to the day. A person approves every message, and every approval is recorded with who gave it.