Most “AI prompts for sales” lists are prompts for writing emails. Those have limited value now, because the models write adequate cold email with almost no prompting and the output converges anyway.
These are different. They assume your assistant is connected to your outreach platform and can act on it. Each one does work rather than producing text.
A note before the list: be specific about criteria. Vague instructions get confidently wrong interpretations, and the fastest way to lose trust in this workflow is to watch an assistant enrol the wrong 200 people because you said “my best leads”.
Monday: fill the funnel
1. Build a list against explicit criteria
Find companies in [industry] with [headcount range] headquartered in [regions], then pull contacts with the title [role] or more senior. Show me the count and a sample of 10 before doing anything else.
The last sentence is the important one. Always ask to see a sample before committing credits or enrolling anyone.
2. Score and cut
Score those leads against my ICP and show me the distribution. Then list the ones scoring above [threshold], with the reasoning for the top five so I can check the scoring is sensible.
Asking for reasoning on a sample is how you catch a scoring model that has latched onto the wrong attribute.
3. Find the lookalikes
Look at my closed-won accounts from the last 12 months. What do they have in common on industry, headcount, region and tech stack? Then find 100 companies matching that pattern that aren’t already in my workspace.
This one regularly surprises people, because the pattern in closed-won is often not the ICP they wrote down.
Tuesday: build the campaign
4. Create the sequence
Create a multichannel campaign called [name] targeting the list from Monday. Email on day 1, LinkedIn connection request day 2, email follow-up day 4, LinkedIn message day 6 if connected, break-up email day 11. Don’t enrol anyone yet — show me the structure first.
5. Check the merge fields
Go through every step of [campaign] and list any merge field that doesn’t have a fallback. For each one, suggest a fallback that would still read naturally.
Broken personalisation at scale is visible and permanent. This takes ten seconds and prevents it.
6. Pre-flight the list
Before I launch [campaign], check the enrolled list for: anyone already in another active campaign, anyone on the suppression list, anyone at a current customer account, and any contacts whose company has changed since the record was created.
The last check catches the departed-contact problem that produces bounces.
Wednesday: work the replies
7. Triage the inbox
Summarise every reply that came in since yesterday, grouped by category: interested, question, objection, referral, not now, negative. For each one, tell me who it is and what they actually said in a line.
8. Draft, don’t send
Draft responses to the interested replies and the referrals. Use my calendar link for anyone asking for a time. Hold everything for my approval — don’t send.
Keep “don’t send” in the prompt until you have read a few weeks of drafts and trust them.
9. Find what is being missed
Show me any conversation with an inbound message and no response in over 48 hours, across every channel. Sort by how interested the reply sounded.
This surfaces the referrals and product questions that quietly go cold, which is the least visible loss in outbound.
Thursday: act on signals
10. Work the job changes
Which contacts in my lists have changed companies in the last 90 days? For anyone who moved to a company matching my ICP, draft a short congratulations-and-reconnect message referencing where we knew each other from.
People who bought from you before, now with new budget, is the highest-converting signal in B2B and almost nobody works it systematically.
11. Harvest an engaged audience
Pull everyone who engaged with [competitor or influencer] LinkedIn page’s posts in the last two weeks. Drop anyone already in my workspace, score the rest against my ICP, and show me who clears the bar with the post they engaged with attached.
Friday: look at what happened
12. The weekly review
Give me last week across all campaigns: sends, replies, positive replies and meetings, broken down by channel and by campaign. Then tell me which sequence step produced the most replies, and flag any campaign where the reply rate dropped more than 30% against the previous week.
The step-level breakdown is the one worth having. Most teams find a disproportionate share of replies come from touches four through seven, which is precisely where manual outbound gives up.
Four habits that make these work
Ask for a sample before committing. “Show me 10 first” costs nothing and catches misinterpretation before it costs credits or reputation.
Say what you do not want. “Don’t send”, “don’t enrol”, “don’t spend credits yet”. Assistants act on what they think you meant, and negative constraints are the cheapest protection available.
Name things explicitly. Campaign names, list names, thresholds, date ranges. “My recent campaign” is ambiguous; “the FinTech CFOs campaign” is not.
Check the workspace. If you run several clients, confirm which workspace the assistant is connected to before asking it to do anything that writes.
Where these break down
These prompts assume the assistant can act on your platform. Without that connection they produce text describing what should happen, which is considerably less useful.
They also assume a clearly defined ICP. Prompt 2 and prompt 3 are close to useless if your ICP is “companies that might need us”, and no amount of prompt engineering fixes that. The specificity has to exist somewhere, and if it is not in your head it will not be in the output.
Setup is covered in the MCP guide, and the broader question of what to delegate in what agentic outbound means.
Frequently asked questions
What are good AI prompts for sales outreach?
The useful ones operate your tools rather than write copy: building lists against explicit criteria, scoring against an ICP, checking a campaign before launch, triaging replies by category, and weekly reporting with step-level breakdowns.
Can ChatGPT or Claude actually build a campaign?
Yes, if connected to your outreach platform through MCP or an equivalent integration. Without that, the assistant can only describe what to do rather than doing it.
How do I stop an AI assistant doing the wrong thing?
Include negative constraints in the prompt — don’t send, don’t enrol, show me a sample first — and name campaigns, lists and thresholds explicitly rather than referring to them vaguely.
What should I not ask an AI to do in outbound?
Decide your ICP or your offer, send to prospects without review in the early weeks, or answer pricing, contract and security questions.