Cold Email With AI: Research First, Then a Claude Draft
Mass AI email blasts are dead. How to write a cold email with AI that gets replies: research, a tight brief, a Claude draft, a human edit. Prompt included.
A couple of years ago, AI-generated cold emails still got replies. Today the recipient spots the template by the second line: "I noticed your company has been growing", "I'm impressed by your product". These emails get deleted unread, because ten of them arrive every week and they all look the same.
The tool is not the problem. AI is bad at "write an email for everyone" and good at "write an email to this specific person, here is what I know about them". The difference is what you feed it. Here is the mechanic of a research-first cold email: a quick research pass on the prospect, a brief, a Claude draft, a human edit.
Why mass AI blasts died
When text generation became free, cold outreach turned into a flood. Everyone uses the same tools with the same default settings, so the output is a crowd of clones: same structure, same compliments, same meeting request. Spam filters learned to catch them, and people learned even faster.
Three things give a template away:
- Empty flattery. "I admire your achievements" is written by someone who cannot name a single one.
- Generic pain. "Companies of your size often struggle with..." fits everyone, which means it fits no one.
- A disproportionate ask. A stranger requesting 30 minutes of your time in the very first email.
And the more templated emails go out, the worse each one performs. Meanwhile a single email built on real knowledge about the prospect stands out immediately: against a flood of clones, it is the only one that reads like a human wrote it.
What works instead: a research-first cold email
A research-first email. The value of a cold email is created before you open a chat window. Claude does not know that this company has a marketing role open, that the founder complained about the sales team on a recent podcast, or that they shipped a new product last week. You find that. The AI then turns your findings into a clean draft in your voice, much faster than you would from a blank page.
The division of labor is simple: research and decisions are yours, packaging belongs to the model.
The step-by-step plan
Step 1. A quick research pass
Ten minutes per prospect: their website, the company's and the founder's social accounts, job openings, news, customer reviews. You are not collecting information in general, you are hunting for one hook: something happening right now that connects to what you offer. One strong hook is enough, three are unnecessary.
Step 2. The brief
Condense your findings into five lines. The company and what it does. The specific fact you noticed. Your hypothesis about the problem that follows from it. Your offer, in one sentence. The proof you can back it with.
If you cannot fill the brief, it is too early to write. Go back to research or cross this prospect off the list.
Step 3. A Claude draft
A prompt to start from:
``` You help me write cold emails. I am [who you are], selling [what and to whom].
Prospect brief: - Company: [name, what they do] - What I noticed: [a fact from their site, socials, job posts, news] - Hypothesis: [the problem they likely have] - My offer: [one sentence] - Proof: [a short fact about your track record]
Tone: [how you normally write]. Here are two of my real emails as a voice sample: [paste them]
Write a draft: - under 120 words - first paragraph about them and the fact from the brief, not about me - one sentence about what I offer - end with a light question, not a meeting request - no compliments, no flattery - use only facts from the brief, add nothing ```
If you write these emails regularly, set up a dedicated Claude project and load it with your product description, your best emails, and your tone rules. Then you will not have to repeat the same preamble in every chat. We covered how that works in the post on Claude Projects.
Step 4. The human edit
The draft is raw material, not a finished email. Read it out loud and rewrite anything you would never say to a person's face. Check every claim about the prospect against your brief. Add one detail only you could know: a mutual contact, a line from their talk, a link to a specific page on their site. Details like that cannot be generated, and they are exactly what separates your email from the flood.
Where AI gets it wrong and what to check by hand
- Invented facts about the prospect. When the brief is thin, the model fills the gaps: "you are a market leader", "you recently expanded". Check every claim: is it from your brief, or did it appear on its own?
- Flattery. Compliments leak through even when the prompt forbids them. Cut every line where you praise instead of observe.
- Length. The model happily stretches one paragraph into three. Trim until the email reads on a phone screen without scrolling.
- Corporate filler. "I would like to inquire", "as part of a potential collaboration". No human talks like this, and the recipient can tell.
- Promises made in your name. Drafts can include timelines, prices, or guarantees you never gave. Before sending, make sure you can stand behind every word.
Data privacy
You are working with information about real people, so keep boundaries. Use open sources only: the website, public social accounts, news, job posts. Do not paste unnecessary personal data into the prompt; a name, a role, a company, and public facts are enough. If your prospects live in a CRM, do not dump the whole base into a chat: take one prospect and only the fields you need. And check your company's policy on passing data to external services; many teams have rules about this.
Ethics: precision, not spam
The mechanic in this post pushes you toward fair play on its own: research takes time, so you physically cannot blast everyone. The rest comes down to three rules. Write only to people your offer can genuinely help. Introduce yourself honestly, with no fake "we spoke last year". After one or two unanswered emails, stop: silence is an answer too. And know the rules that apply in your market, such as CAN-SPAM in the US or GDPR in Europe.
The short version
The formula: research by hand, a brief, a Claude draft, a human edit. AI speeds up the packaging, while the knowledge about the prospect is still yours to earn, and that knowledge is what decides the outcome. Once a prospect replies, the same mechanic builds the next document in the deal; we covered it in the post on writing a proposal with AI.
If you want to practice this on your own prospects, take a look at AGINE Academy. It is a game-style academy for Claude: you run missions in the real tool, every lesson ends with a working result, and an AI mentor helps along the way. The starter block of 4 lessons is free and open without signing up.
Questions
It can write the words, but not the substance. Everything that makes a cold email land comes from what you know about the prospect: what is happening at their company, which problem is urgent, how you can help. You gather that in research and condense it into a brief. Claude turns the brief into a clean draft and holds your tone, but do not send without your own edit: verify the facts and cut anything that does not sound like you.
Shorter than feels natural. Aim for something that fits on a phone screen without scrolling, usually 80 to 120 words. One paragraph about them, one sentence about you, one light question at the end. Language models tend to inflate text, so put a hard length limit in the prompt and trim further during the edit.
Read it out loud. Templates give themselves away with vague flattery, generic pains of "companies your size", and corporate filler. If the email contains a fact you could only know about this specific prospect and it sounds like the way you normally talk, it passes. Feeding Claude two or three of your real emails as a voice sample helps a lot.
A relevant, personal email to someone's work address is normal business practice. Trouble starts with volume: scraping personal inboxes, thousand-recipient blasts, five follow-ups in a row. Write only to people your offer genuinely fits, introduce yourself honestly, stop after one or two unanswered emails, and check the rules in your market, such as CAN-SPAM in the US or GDPR in Europe.