LinkedIn Cold Outreach Prompt
A first message and its follow-up, built from a stated hypothesis, with every claim traceable to something you actually gave it. The flagship outreach prompt. Most "personalisation" prompts produce a compliment about the prospect’s career and a pitch underneath it; the prospect has read forty of those this month. This one requires a hypothesis about what is true for this person, writes the message that tests it, and refuses to invent a reference when you have not supplied one.
You are writing a first LinkedIn message from one working person to another. You are not writing marketing copy. The test for every sentence: would a busy person read it and think "this is about me", or "this is about them"?
## Objective
{{OBJECTIVE}}
## Me
{{ROLE}} at {{COMPANY}}
What we do: {{PRODUCT}}
Proof I may cite: {{PROOF}}
Tone: {{TONE}}
Constraints: {{CONSTRAINTS}}
## Them
{{PROSPECT_NAME}} — {{PROSPECT_ROLE}} at {{PROSPECT_COMPANY}}
{{PROSPECT_PROFILE}}
## The ask
{{OFFER}}
## Method
1. State the hypothesis first, before writing anything: one sentence about what is probably true for this person that makes my product relevant. If the evidence does not support one, say so and stop — a message with no hypothesis is a template and should be sent as one, not dressed up.
2. Choose the single reference. It must be checkable by the recipient and specific to them.
3. Write the message in this shape:
- Line 1: the reference. No greeting beyond their first name.
- Line 2: the hypothesis, stated as something you have observed elsewhere, not as a claim about them. "The pattern I keep seeing is X" invites correction; "You clearly have problem X" invites a defence.
- Line 3: the question. Answerable in one line, genuinely open, and useful to me whichever way they answer.
4. The ask goes in the follow-up, not here — unless my stated objective is explicitly a meeting request, in which case put it in one short final line.
5. Write the follow-up for five business days later. It must ADD something: a number, a counter-example, a link to something useful that is not mine. Never "just bumping this".
6. Self-check before returning: remove every sentence that would be true of a different prospect.
## Hard rules
- Under 700 characters for the message body. Print the character count.
- One reference. One hypothesis. One question.
- No compliments, no "I hope you don't mind me reaching out", no "we help companies like yours", no three-adjective descriptions of my product.
- Never state a fact about their company that I did not give you. If you want one, ask me for it in the "what would make this better" section.
- If they are the wrong person, say so instead of writing the message.
## Output format
### Hypothesis
One sentence, plus the evidence and a confidence rating.
### Message
The message. Then (n characters).
### Why each line is there
One line per line of the message. This is how I learn to write the next one myself.
### Follow-up, day 5
The message. Then (n characters).
### What I would cut if it has to be shorter
The specific sentence and what is lost.
### What would make this better
The one piece of information you did not have.- You have research on a prospect and need the message that follows it.
- Your messages get read and not replied to.
- You are writing the first touch for a sequence and need it to be the strong one.
- You want to give an agent an instruction it can run per prospect without supervision.
- Run the Prospect Research prompt first and paste its value hypothesis into {{PROSPECT_PROFILE}} along with the raw profile text.
- Fill {{PROOF}} honestly, including "none yet". A model with no proof writes a more honest message than one that invents a case study.
- Read "Why each line is there". After five runs you will stop needing the prompt for straightforward prospects, which is the correct outcome.
- Send the follow-up. About half the replies to a well-aimed first message arrive after the second one.
The output schema it demands
The single highest-leverage part of any prompt. Without a declared shape you get prose you then have to reformat; with one, you get sections you can act on — and you can tell at a glance when the model skipped something.
### Hypothesis (+ evidence + confidence) ### Message (+ char count) ### Why each line is there ### Follow-up, day 5 ### What I would cut if it has to be shorter ### What would make this better
What it looks like filled in
Illustrative, not a promise — this is the shape of a good answer rather than a guarantee of one. The scenario is the same fictional company used across the whole toolkit.
Prospect: Dana Whitfield, Group Finance Director, Northbeam Logistics. Evidence: her post on a five-week close after an acquisition. Objective: book a 20-minute call.
Hypothesis — the close got longer because the acquisition added a fifth entity to a four-entity process, not because her team slowed down. Confidence: high — she wrote "since the acquisition" herself.
Message, 384 characters — "Dana — the line in your post that stuck: five weeks, since the acquisition. The pattern I keep seeing at 4-to-5-entity finance teams: manual inter-entity matching grows faster than headcount does, so the close stretches even when nothing else changed. Close to what's actually eating the time for you, or something else?"
Follow-up, day 5 — adds a number rather than repeating the ask: "One thing I didn't say last time — the two customers with this exact problem got their close back to 3 days with no headcount added. Still curious whether the fifth entity is the actual driver for you."
Do not do these
Each one is a real failure mode, not a disclaimer.
- Adding a second reference "to show you did the work". Two references reads as a dossier and changes the tone from interested to transactional.
- Putting a calendar link in the first message. It converts a conversation into a funnel step before they have agreed there is anything to discuss.
- Keeping a hypothesis the model rated low confidence. That is the message that gets the "who is this?" reply.
Attach these once and stop retyping your business
This prompt works pasted into a blank chat. It works considerably better inside a Project with these attached, because then the ICP, the product and the limits are already in context and every answer inherits them.
OUTREACH.md
Every message an assistant writes comes out at the right length, in the right shape, with the openers you hate already banned.
Free downloadBRAND-VOICE.md
Drafts arrive in your voice on the first attempt instead of after four rounds of "less corporate".
Free downloadPRODUCT.md
Messages stop describing features and start describing what changes for the person receiving them.
Free downloadOBJECTIONS.md
The assistant answers the objection you actually get, in the way that has worked, instead of improvising.
Free downloadThis is one step in a sequence
The output of the step before is the input to this one, and this one's output is the next one's input. That is the whole reason these are not independent pages.
Before you run it
Why does it refuse to write the message sometimes?
Because when you supply no specific evidence, the only thing the model can do is invent a reference or write a template. Both are worse outcomes than being told to go and read the profile properly, and the invented one carries a real risk of you asserting something false to a prospect.
Is 700 characters a LinkedIn limit?
No — LinkedIn messages allow far more. It is a reading limit. A first message longer than about 700 characters is scrolled past on a phone, which is where most of them are read.
The reference behind this prompt
If this message is going out as an InMail rather than to a connection, it is costing you a credit worth between $2.40 and $8 depending on your plan.
Message volume is capped by the invitation ceiling and the InMail allowance, not by how many good messages you can write.
Written to a fixed structure and run through Claude and ChatGPT to confirm each one returns the output format it declares. That is not the same as a measured reply-rate benchmark, and nothing here claims one — if you run these at volume and have numbers, we want them.
Take the whole folder instead
This prompt, the files it expects and the rest of the sequence, as a zip you can drop into a Claude Project, a Cursor workspace or an agent repo. Free, no account.