Claude for Cold Email

You write cold email that gets replies by giving Claude the prospect's real context and asking for an opening line built on a specific observation, not flattery. The first line is most of the result, so the work is to make one true sentence per lead, not one template sprayed at two hundred people. The prompts and the installed skills below both do that, the difference being whether the depth survives scale.

Personalization is the whole game. Hunter.io's analysis of cold-email data found that real personalization lifts reply rates by about 52% (Hunter.io), while the average cold email reply rate sits around 8.5% (Backlinko, 2025). The gap between those two numbers is where the replies live, and almost all of it is decided by the opening line. For the system that personalizes a whole list at once and remembers what it sent, see Claude as a business operating system.

• Real personalization lifts reply rates by about 52% (Hunter.io).
• The average cold email reply rate is about 8.5% (Backlinko, 2025).
• Most cold email fails because line one is about the sender, not the prospect.
• One true line per lead beats one template sprayed at two hundred.
• The two skills install in about five minutes; the setup interview runs about thirty.

Why Does Most Cold Email Get Ignored?

Because the first line is about you: your company, your product, your "quick question." The reader can tell in two seconds that the same email went to a thousand people, and the moment they can tell, the send is dead. A line built on something specific and true about their business earns the next sentence. Everything else in the email, the body, the proof, the ask, matters less than that one sentence, because nobody reads the body of an email whose first line already told them it was a blast.

This is not a copywriting problem you fix with a cleverer hook. It is a context problem. A good opener requires knowing something real about the recipient, and the only way to know something real at scale is to read each prospect before you write to them. Most cold email skips that step because reading two hundred prospects by hand is slow, so the sender reaches for a template and hopes volume covers for relevance. It does not. Volume of a line that reads like a mail-merge just burns the list faster.

Why Does Most "Personalization" Still Fail?

Because most of what gets called personalization is a merge field, not an observation. "Loved what you're building at {Company}" is technically personalized, it has the prospect's name in it, and it is worse than sending nothing. It signals a mail-merge in the first five words, which is exactly the tell you were trying to avoid. The reader has seen that opener a hundred times and reads it as the form letter it is.

Real personalization is an observation a human could only make after spending thirty seconds on the prospect's site or profile: a specific change they shipped, a role they just filled, a claim on their pricing page, a gap between what they say and what they show. The test is simple. If the line could be sent to any other company on the list, it is not personalization, it is filler with a name slot. The whole point of personalize-at-scale is to pass that test two hundred times without a human writing two hundred lines by hand, and without the quality collapsing into "Loved what you're building."

The failure mode worth naming is the tradeoff most people assume is unavoidable: you can have depth or you can have scale, not both. Hand-write ten lines and they land; automate two hundred and they read like a robot. That tradeoff is real for a template and false for a system that reads each lead's actual context before it writes. Depth surviving scale is the entire job, and it is the line between an 8.5% reply rate and a 52% lift on top of it.

What Does A Real First Line Actually Use?

A first line that earns a reply is built from a single concrete fact about the recipient, tied back to the reason you are writing. It uses something that exists in the prospect's own material, not an inference and not a compliment. The structure is small: name the specific thing you noticed, then bridge it to the problem you solve, in one sentence that could not have been sent to anyone else.

The inputs that make this work are the prospect's site copy, their recent activity, a role or hire, a public number, a launch, a stated priority. The inputs that ruin it are adjectives ("impressive," "innovative"), generic praise ("loved your post"), and anything you cannot point to a source for. A line that references a fact the writer cannot cite is a hallucination, and a confident hallucination in line one is worse than a blank, because the prospect knows their own business better than you do and will catch the invented detail instantly.

That last rule, every claim traces to a source field, is the difference between personalization that scales and personalization that embarrasses you at volume. One wrong "I saw you just raised a Series B" to a company that did not raise anything undoes the whole send. The discipline is to write only from what the data actually says, and to leave a line blank when the data is too thin to say anything true.

Which Prompts Should You Paste?

Here is the manual version. Open Claude, paste the block below, and fill the brackets with one prospect's details. This is the one-off taste of what the installed skill does for a whole list. It works for a single lead and is exactly why the skill is worth it the moment you have more than a handful.

The "name the exact fact" instruction is doing the heavy lifting. It forces every line to point at a source, which both improves the line and lets you catch an invented detail before it goes out. The final instruction, leave it out rather than inventing, is the same rule the installed skill enforces programmatically. A blank beats a tell.

What Does Good Look Like Next To Bad?

The same prospect, written two ways. The before is the merge-field default. The after is built on a fact from the prospect's own site, tied to the offer.

Neither line is longer or cleverer. The after simply references something true and specific, then points it at the thing you help with. That is the whole move, repeated once per lead. The hard part is never writing one good line. It is keeping that quality at line number two hundred.

How Do You Do This For A Whole List Without It Going Generic?

The prompt above personalizes one prospect well. Doing it for two hundred by hand is where quality dies, because by lead forty the writer is tired and reaches for the template. That is the exact problem the installed skill solves, not by writing faster, but by reading each lead's context and writing one line against that context, then checking every line before it ships.

The /email-personalize skill writes one QC'd icebreaker per lead for a cold list. It reads an enriched lead file plus your saved business context, calibrates the tone on two sample leads so the voice is yours, fans out writers across the list in parallel, then runs every line through a quality gate: no banned openers, no filler, person-not-company focus, the observation must connect to your offer, and no fact that does not trace to a source field. A line that cannot pass cleanly is left blank with a note, not faked. The skill needs a vault, the small folder of context files that holds your offer, your voice, and your customer profile, because that is what keeps the lines on-brand and tied to what you actually sell.

The /email-sequence skill is the layer underneath the opener. It designs a full per-email follow-up, welcome, or sales flow in your voice, with one job per email, the timing, the subject and preview text, and a metric plan tied to a real target. It also needs a vault, for the same reason: the copy comes out in your voice, aimed at your audience, pointed at your ask, read from your context rather than guessed. The two pair cleanly. The sequence is the body of the send; the icebreaker is the opening line per recipient.

One-Off Prompt Or The Installed Skill?

Both write a good line. The difference is what happens at scale and over time. The prompt is a single read with no memory. The skill reads context, checks its own work, writes the result back, and on the next run leads with what changed.

For one prospect, the prompt is plenty. The moment you are writing to a list, the skill is the difference between depth that holds and depth that decays around lead forty. The setup is what buys you that, and the setup is what most people skip.

What Can It Do, And What Can It Not Do?

Honest limits beat a feature list. The skill is good at reading a lead, writing a true opening line, checking it, and saving the result. It is not a scraper, not a guarantee, and not a substitute for a clean list.

The right mental model is a careful writer who reads every prospect, never gets tired by lead two hundred, refuses to make up a detail, and writes only in your voice. You still pick the list and you still hit send. The skill removes the part that does not scale by hand, which is reading and writing two hundred true lines without the quality sliding.

How Do You Set It Up?

Both skills ship inside Conversion OS, a free, open set of skills that run in Claude. Installing it takes about five minutes and no coding. The one-time setup interview that builds your vault, the context these two skills read from, runs about thirty minutes.

1. Open Claude Code or Cowork, then Plugins, then Marketplace.
2. Add the Conversion OS repository and install it.
3. Run the setup command. Claude interviews you and writes your context folder, your offer, your voice, your customer, in one session of about thirty minutes.
4. Point /email-personalize at an enriched lead list, calibrate the tone on two samples, and let it write and QC the rest.

You do not have to do the whole list on day one. Run it on ten leads first, read the lines, dial the tone, then let it loose on the full file. The point of the first run is one batch of icebreakers you would actually send, saved to memory so the next run starts from a locked-in voice instead of a cold guess.

What Are The Common Mistakes?

Calling a merge field personalization. "Loved what you're building at {Company}" is a tell, not a touch. If the line could go to any company on the list, cut it.

Skipping the source check. A line that references a fact you cannot point to is a hallucination waiting to embarrass you. Make every observation trace to something real, and leave it blank when it cannot.

Trying to scale without a vault. Without saved voice and offer context, the lines drift generic by lead forty. The context file is what keeps two hundred lines on-brand.

Spraying the template anyway. Volume of a line that reads like a blast does not beat the odds, it burns the list faster. One true line per lead is the only version that moves the 8.5% baseline.

Writing the opener without the body. A great first line into a buzzword body wastes the open. Pair the icebreaker with a sequence that earns the click, which is what /email-sequence is for.

Methodology

The reply-rate figures are from named third-party research: Hunter.io's analysis of cold-email data for the roughly 52% personalization lift, and Backlinko's 2025 cold-email statistics for the roughly 8.5% average reply rate, both linked above. Conversion System publishes no client outcome numbers, so none appear here; the proof is the named research and the shown method. The skill descriptions reflect what the open Conversion OS /email-personalize and /email-sequence skills do, which you can read in the public repository: one QC'd icebreaker per lead for a cold list, and a per-email sequence in your voice, each reading from a vault you build in the setup interview. The aim is replies, and replies start with one true line per lead, not one template sprayed at two hundred.

Related: Claude as a business operating system · Score your leads with Claude · Run a free revenue audit with Claude · Win clients with a revenue audit