Before the prompts: the context file
Do this once and every prompt below improves. Create a file with your ICP, your voice, your banned phrases and your offer, and give it to Claude at the start of any outreach work.
You are writing LinkedIn outreach for me. Here is my context. Product: [one sentence, what it does, not how]. ICP: [company size, sector, region]. Buyer: [role, what they own, what they are measured on]. What I never say: [banned phrases, claims you cannot back]. Proof I can use: [numbers you can defend]. Refer to this on every task in this conversation and tell me when something I ask for conflicts with it.
Everything after this assumes that file exists.
Writing the first message
1. The core message prompt.
Write a LinkedIn first message to [name], [role] at [company]. Rules, all mandatory: between 240 and 420 characters. Two and a half sentences. Ends in a question about how they handle something today. Names [their product or a competitor of theirs] in the first line. Says nothing whatsoever about my product, my company or what we do. No meeting request. Count the characters and tell me the number.
The character count instruction matters more than it looks. It makes the model check its own work, and it is the constraint it drifts on first. Messages of 240 to 420 characters replied at 38.6 percent; under 150 dropped to 21.4 percent and over 700 to 11.2 percent.
2. Strip a message you already have.
Here is a message I wrote: [paste]. Delete every sentence containing the word we, our, or the company name. Rewrite what remains as two and a half sentences ending in a question. Show me the before and after character counts.
The version with solution detail replied at 18.6 percent. The version with it removed replied at 44.5 percent.
3. Find the competitor angle.
Here is [company]‘s website and their LinkedIn page: [paste or URL]. Identify the product they use or sell that is closest to my category, and one direct competitor of theirs by name. Give me one sentence I could open with that references it without flattering them.
Naming the prospect’s own product or a direct competitor took reply rate from 17.1 to 34.5 percent across 10,815 prospects. It is the largest copy-level effect we have measured.
4. The question generator.
Give me five questions I could end a cold message with, for a [role] at a [sector] company of about [size]. Each must be about how they handle something today, not about their plans or priorities. Each answerable in one sentence. None may reference my product.
Question-led copy replied at 33.2 percent against 10.8 percent for company-led copy across 9,540 enterprise prospects.
Building the list
5. Verify subject-matter fit.
Here is a prospect: [profile text or URL]. In one sentence, tell me why [specific problem] would be this person’s responsibility rather than someone else’s at the same company. If you cannot make that case from the evidence in front of you, say “no case” and stop.
This is the highest-value prompt on the page. Broad job-title targeting replied at 13.0 percent. The same sequence sent to a list where subject-matter fit had been verified per person replied at 51.9 percent. Four times, from work done before the message exists.
6. Batch the same check.
Here are 40 prospects. For each, output the name, a one-sentence case for why the problem is theirs, and a verdict of keep or drop. Drop anything where the case relies on job title alone. Give me the drop count at the end.
Expect it to drop a third or more. That is the tool working.
7. Pick the level.
For [company], I have these people: [list]. Which is most likely to take a meeting rather than just reply politely? Assume seniority buys a reply and not a meeting, except at companies under 50 people where the founder is the operator.
Directors and above replied at 32.0 percent and booked at 2.2 percent. Managers and below replied at 26.0 percent and booked at 4.4 percent.
Handling replies
8. The morning triage.
Here are the replies that came in since yesterday: [paste]. Classify each as interested, question, objection, not now, or not interested. For the first three, draft a response under 300 characters that answers what they asked and proposes a specific time. Do not send anything.
Replies answered inside 24 hours converted at 21.4 percent. After 48 hours, 3.6 percent. This is the biggest single effect in the whole dataset and it is an operational problem, not a writing one.
9. Handle the objection without arguing.
Prospect said: [paste]. Write a reply that agrees with the part of their objection that is true, adds one piece of information they do not have, and asks whether that changes anything. Under 300 characters. Do not defend the product.
10. Add the booking link at the right moment.
Here is a reply thread: [paste]. Tell me whether the prospect has shown enough interest for a calendar link, or whether it is still too early. If it is time, write the message with the link and two specific times.
A calendar link left reply rate unchanged and lifted meetings by 8 percent. It removes friction after the decision, not before it, and sending it too early reads as presumptuous.
Video and voice
11. The 30-second video script.
Write a 30-second video script for [name] at [company]. Three beats: why I am recording this, one specific thing I noticed about them, one question. Say nothing about my product. Mark the timing. Give me the word count so I can check it lands near 75 words.
A personalised video first touch replied at 40.4 percent against 28.8 percent for text-only, and produced 57 percent more meetings across 8,385 prospects. It held over eleven weeks without decay.
12. The step-two voice note.
Write a 25 to 35 second voice note for the second touch to [name], who has not replied to my first message. Reference something specific, ask one open question, and make it sound like speech rather than writing. No pitch.
A voice note at step two lifted meeting rate from 4.1 to 5.0 percent. Smaller than video and far cheaper to produce at volume.
The thing these prompts cannot fix
Prompting well makes the writing faster and stops it drifting. It does not move the number that matters most.
Across everything we measured, targeting beat copy by a wide margin: 13.0 against 51.9 percent from list construction alone, where the largest copy effect we ever found was a doubling. If you use one prompt from this page, use number five, and use it before you write anything.
The full method and all fifteen tests are on the 2026 outbound benchmarks page. Free, no gate. If you want Claude running these against a live campaign rather than in a chat window, the MCP setup takes about five minutes.