A first message gets replies when it proves you looked: one specific observation, the pain it implies, the outcome you create, and one question. Six short sentences, no deck, no calendar link, and no number you have not confirmed.
A first message gets replies when it earns the opening line: one specific observation about the company, the pain that observation implies, the outcome you create, one low-friction question, and nothing you cannot prove.
Every bad cold message shares one trait: it could have been sent to anyone. The writer did not learn anything, so there was nothing to say, and the reader can tell from the first line.
The six-part structure
Part | What it does | Example shape
Greeting | Uses the first name only | Hi Ada,
Signal | Proves you looked | I saw that you opened two depots this month.
Pain | States what that costs, hedged | That may mean the dispatcher queue is already behind.
Outcome | What you create, in their terms | We find the customers that work leads to.
Offer | The outcome, not the feature | No extra hire and no list to buy.
Question | One low-friction ask | Worth a short look?
One sentence each. Put the signal and the pain in one paragraph, the outcome and the offer in the next, and the question on its own. The whole thing should fit on a phone screen without scrolling.
Why the signal is the whole message
The first sentence decides whether the rest is read. A signal is something the company published that creates work: two operations roles open, a new location, a launch, a funding round. It is not a compliment about their website, and it is not a guess about their strategy.
A hiring post is the strongest version, because a company only pays for a role when the work has already outgrown the team. It is also easy to verify, which means you can quote it.
Never state a number you have not confirmed
This is the rule that separates outreach that builds a brand from outreach that damages one. A hedged inference, that the team may be stretched, reads as honest. An invented statistic reads as spam, and the reader knows it immediately, because it is usually about their own industry.
A hedged inference keeps the reply. An invented number loses it and the next four messages too.
The practical test: for every sentence that claims something, write down where you read it. If you cannot, cut the sentence. If the message is empty after the cuts, the research was not done.
The same message on a social channel
A connection note is not an email with the subject removed. It is one sentence about the signal and one question, because that is all the format allows. A reply in a thread should answer the person first, then earn the right to say what you do.
Email: six short parts, one question, no attachment.
LinkedIn: two sentences, and a reason you looked at their profile.
X: answer the post, then one sentence of context. Never paste a pitch.
The mistakes that cost the most replies
A paragraph about your company before their problem.
A calendar link in the first message, asking for a meeting before there is a reason.
A subject line that names the product instead of the pain.
Three questions in one message, so none of them gets answered.
Attachments, which look like a proposal and get archived.
Claims about their industry that are a year out of date.
Test it before you approve a hundred
Write one message to one real company you know well enough to be specific about.
Read it out loud. If you would not say it to their face, rewrite it.
Send ten by hand. Count replies, not opens.
Change one part, the signal or the question, and send ten more.
Only hand it to the agent once the hand-written version replies.
Where AI helps and where it hurts
A model is good at finding the signal in a page and turning it into a sentence that sounds like a person. It is bad at knowing which claim you are willing to stand behind. That is why the useful pattern is an agent that proposes and a person who approves, with every factual sentence tied to a saved quote.
Snaut labels each sentence as a fact, an inference, an offer or a question, and refuses a fact whose quote is not in the snapshot it cites. You can read the source for any line before the message goes out.
What is a good reply rate for a first message? There is no single number, because it depends on the list, the offer and the channel. What you can judge is whether the message is specific enough that it could not have been sent to another company. A high rate on a tiny list usually means tight qualification, not better copy.
How long should a cold email be? Under about 90 words. Greeting, one signal, one pain, one outcome, one offer and one question, each a single sentence. Long enough to be specific, short enough to answer from a phone.
Should a first message mention the recipient's company? Yes, with something they published that proves you looked: a hiring post, a launch, a page on their site. Naming the company without naming something real about it reads worse than not naming it at all.
What subject line works best? Two to five words that name the pain, not the product. A subject that describes their problem is opened; a subject that describes your company is deleted.
Does the same message work on LinkedIn or X? The parts are the same, the length is not. A connection note is a sentence or two, and a reply in a thread should answer what the person actually said. Copying an email into a direct message is the most obvious way to be ignored.
Is it okay to use AI to write the first message? Yes, if every factual sentence is tied to a saved source and someone is accountable for the claim. The risk is not that a model writes the message, it is that nobody can show where a number came from.