The principles of cold copy — good vs bad, annotated.

TL;DR

• Until $1M ARR, founder-written emails outperform SDR-written emails. Use that pattern interrupt.
• Length: 50 to 125 words. Under one minute to read. Plain text. No images, no links, no HTML.
• Structure: hook (specific observation about them) → value prop (the outcome) → credibility (one line) → CTA (binary yes/no).
• Write to a friend, not to a target. If you would not say it out loud to someone you respect, don't send it.
• What kills emails: pitching features, sounding like an SDR, links and images, generic "Hi {first_name}", "just bumping this" follow-ups.

The premise

A cold message gets read by someone who did not ask for it, in a triage context — the first 8 to 15 seconds where the question is open, archive, or delete. Most operators write from inside their own context. The message reads sensibly to them and unintelligibly to the recipient. Sender-oriented copy replies at 0.3 to 0.7%. The same lists and infrastructure, written against four structural principles, reply at 3 to 8%. Infrastructure is the precondition. Copy is the conversion.

The four principles below are not proprietary. They are restated because the gap between operators who apply them and operators who recite them while violating them is roughly an order of magnitude in reply rate.

One more upstream point: until you hit roughly $1M ARR, you should be writing these emails yourself. Founder voice replies at roughly 2x the rate of SDR voice on the same offer. Hire an SDR before that and you are paying someone to make your unfair advantage harder to deploy.

The four principles

1. Specificity. Named companies, named features, specific numbers — anything that could not have been written about a different recipient.
2. Brevity. 50 to 90 words for a first-touch. Reading cost on mobile is per-line, and every line past the third costs reply rate.
3. Recipient-orientation. The message is about the recipient's context, not the sender's. The test: delete every sentence about the sender and see whether the message still functions.
4. Explicit ask. One specific action the recipient can take in one click. Not a request to construct the next step themselves.

The principles compound. Three of four typically clears 2 to 4% reply rate on a clean list. All four clears 4 to 8%. None — the modal B2B cold message — sits below 0.5%.

Specificity

The correlation between concrete specificity and reply rate is the strongest single signal we have observed across audited campaigns. A first-touch that names the recipient's company, a specific product or feature, and a specific number — headcount, funding round, recent shipment, hiring signal — outperforms its de-specified equivalent by 2 to 3x.

The failure mode is the generic claim that could apply to any recipient. “I noticed your team is scaling fast” applies to every Series B company that has ever existed and signals to the recipient that the sender does not know which Series B company they are writing to. “You're probably dealing with the usual challenges around X” signals the same thing. The recipient's pattern-match on these phrases is fast and reliable — a sub-second classification that the message is templated, the sender has no specific reason for reaching out, and the message can be deleted without cost.

The operational test for specificity: take the message and replace the recipient's name and company with a different recipient's name and company. If the message still reads sensibly, it has failed the specificity test. The mark of a specific message is that the substitution would break it.

Brevity

Reading cost is paid per line on mobile, where 65 to 75% of B2B mail is first opened. A message that runs four screen-heights on a phone costs more to read than it costs to delete. The reply curve bends down sharply past 120 words. A 90-word message and a 150-word message differ by 30 to 40% in reply rate on the same offer.

The target is 50 to 125 words — enough for a hook, a why-you-specifically sentence, a credibility line, and an explicit ask on a single mobile screen with no scroll. The recipient should be able to read it in under a minute. Below 40 words reads as low-effort. Above 130 reads as a pitch deck pasted into mail.

The failure mode is the operator who believes the recipient needs more context to understand the offer. They don't. A 200-word first touch is you solving your problem of feeling under-explained at the cost of the recipient's decision speed.

Recipient-orientation

The single most useful diagnostic for cold copy is the deletion test: cross out every sentence in the message that is about the sender — their company, their product, their accomplishments, their previous customers — and read what remains. A message that still functions after this deletion is recipient-oriented. A message that collapses to a fragment is sender-oriented and will reply at a sub-1% rate regardless of how well the infrastructure is configured.

The failure mode is the opener-pitch-CTA structure where the opener is a perfunctory line about the recipient, the pitch is three sentences about the sender's product, and the CTA is a generic meeting ask. By word count this is typically 80% sender-context. The recipient reads the opener, recognizes the pivot to pitch, and stops reading. Reply rate across the campaigns we have audited: 0.4 to 0.6%.

Recipient-orientation does not mean the message contains no information about the sender. It means sender information is structured as a function of the recipient's situation. “We work with three companies in your exact stack and the consistent finding has been X” is recipient-oriented. “We are the leading X for Y” is sender-oriented even when Y describes the recipient's segment.

Explicit ask

The CTA is a binary distinction. Either the recipient can take the requested action in a single click without constructing the next step themselves, or they cannot. Messages in the second category reply at roughly half the rate of messages in the first category.

“Open to a 15-minute call next Tuesday or Wednesday afternoon?” is an explicit ask. The recipient's response is a one-word reply. “Would love to learn more about what you're working on — let me know if there's a good time” is not. The recipient is being asked to propose a time, draft a calendar invite, and construct the framing themselves, against a context where they do not yet know what the meeting is for. The cost of the second formulation is roughly 1 to 2 percentage points of reply rate.

The failure mode is the operator who believes the soft, deferential ask reads as polite and therefore as effective. The soft ask is read by the recipient as work — work the sender is offloading to the recipient. The polite framing is irrelevant; what is being measured is the cost of replying. The lower that cost, the higher the reply.

The “good copy” annotated example

A 71-word first-touch, written against a Series B fintech head of revenue operations, illustrating the four principles in their operative form.

Subject: routing rules for inbound from Plaid signups

Hi Maria —

Saw the November SOC 2 announcement and the four CRE roles you opened
last week. Three teams in roughly your post-Plaid-integration stage
hit a routing problem around month nine — inbound from product
signups outruns the rep capacity that was sized for outbound.

Curious if that pattern matches what you're seeing. If yes, happy
to send a one-page write-up of how the three teams solved it.

— Phillip

The annotations: specificity sits in “November SOC 2 announcement,” “four CRE roles,” and “post-Plaid-integration stage” — three concrete details that could not have been written about a different recipient. Brevity sits in the 71-word total, which fits a single mobile screen. Recipient-orientation sits in the structure: the deletion test removes only one phrase (“happy to send a one-page write-up”) and the message still functions. The explicit ask is a yes/no question — “Curious if that pattern matches what you're seeing” — that costs the recipient a single-word reply.

The “bad copy” annotated example

The same offer, written by an operator who has internalized none of the four principles. 184 words.

Subject: Quick question

Hi Maria,

I hope this email finds you well! I noticed your company is doing
some really exciting things in the fintech space, and I wanted to
reach out because I think there might be some synergies between
what we're building and what your team is focused on.

We're a fast-growing platform that helps high-growth companies like
yours optimize their go-to-market motion by leveraging AI-powered
insights and best-in-class automation to drive efficiency across the
revenue org. We've worked with hundreds of companies and have helped
them achieve incredible results, including some of the largest names
in fintech and SaaS.

I'd love to set up a quick 30-minute call to learn more about your
priorities and explore whether there's a fit. We could also discuss
how some of our customers have approached the challenges you might
be facing as you continue to scale.

Would love to learn more about what you're working on — let me know
if there's a good time that works for you.

Best,
Alex

The annotations: specificity is absent — every phrase could have been written about any of ten thousand recipients. Brevity is absent — 184 words across roughly six mobile screen-heights. Recipient-orientation fails the deletion test catastrophically; removing every sentence about the sender leaves the salutation, the closing, and approximately twelve words of recipient-context, mostly the false familiarity of “I noticed your company is doing some really exciting things.” The ask is non-explicit — the recipient is being asked to propose a time, against zero context for what the meeting is for. The expected reply rate of this message on a clean list is in the 0.2 to 0.5% range.

AI-slop failure modes

A specific set of patterns now reliably signals AI-generated cold copy to recipients trained, over the past two years, to identify and dismiss it. The signal is approximately as reliable as the templated-message signal was in 2018.

• Uniform sentence cadence. Three sentences of roughly equal length, each beginning with a different subject. Human writing varies cadence — a five-word sentence next to a twenty-five-word sentence is a human signature; three sequential fifteen-word sentences is a model signature.
• The “I noticed your recent…” opener at scale.When a model is asked to write a unique opener for each recipient from a scrape, the openers rhyme — “I noticed your recent post about X,” “I saw your recent announcement about Y” — and recipients who get four of these in a week recognize the pattern.
• Hallucinated specifics. A model asked to insert a concrete detail will, at a rate of roughly 5 to 10%, insert one that is wrong. Recipients who fact-check one message and find the error assume every message from the sender is similarly fabricated.
• Em-dash overuse without rhythm. Models trained on web text overproduce em-dashes — the punctuation appears between every clause. A human uses an em-dash for emphasis; a model uses it as a default comma. The accumulation is detectable.
• The closing pivot to “Let me know your thoughts.” Appears across roughly 40% of AI-written cold messages on lists we have audited and is structurally indistinguishable from spam classifier training data.

What kills cold emails

Five patterns produce the largest reply-rate collapse, in our observation, independent of infrastructure quality:

• Pitching features instead of problems. “We use AI to…” loses to “your reps spend two hours a day on…”. The recipient cares about the outcome, not the mechanism.
• Sounding like an SDR. Smooth, polished, structurally uniform copy reads as commercial mass-send. Awkward, specific founder copy reads as human.
• Links and images. Plain-text outperforms HTML on every metric. Each image, logo, and embedded calendar link is a small deliverability tax and a large authenticity tax.
• Generic “Hi {first_name}” personalization. First-name token without specific context is the marker of a sequencing tool. Either personalize meaningfully or skip the salutation.
• “Just bumping this” follow-ups with no new value. A follow-up that carries no new angle, peer reference, or specific observation produces roughly half the reply rate of a follow-up that does, and the slot is spent.

Personalize to the recipient, not the company

The lowest-effort personalization is the company-level token — industry, headcount, funding stage. Every other sender on the recipient's list has access to the same data, so the result reads as scraped enrichment, not research. The personalization that works references the recipient as a person: something they wrote, a talk they gave, a hire they made, a post they engaged with. The threshold is whether a different person at the same company would receive a different opener — if not, the personalization is at the company tier, and the recipient classifies it accordingly.

A practical corollary: do not trust AI to personalize the pitch. AI-generated openers that pull from public profiles read like a LinkedIn DM from someone who skimmed the About page. The model is fine at varying greetings and structure for deliverability (spintax on “Hi/Hey/Hello”). It is not fine at producing the specific human observation that earns the read-through. That part is yours.

The “personalization that looks like AI” problem

A particularly costly failure mode: the operator who layers AI personalization on top of a templated structure produces a uniqueness signature that recipients detect more reliably than a fully-templated message. The mechanism: the static template structure is constant, the variable insertion is high-variance, and the resulting message reads as a templated frame with an awkwardly grafted personalization graft. The recipient's classification is “templated, with effort to disguise” — a more damning classification than “templated.”

A fully-templated message and a fully-bespoke message both have legible voice signatures, and recipients respond to legible voice. The hybrid has neither. Reply rates on the template-with-AI-personalization structure run 20 to 30% below either pure approach.

The voice problem — founder vs SDR

Copy that reads like a sales development rep replies at roughly half the rate of copy that reads like a founder. The recipient does not check your title before deciding to reply — the mechanism is voice. Founder voice tends to be specific, occasionally awkward, and structurally uneven. SDR voice tends to be smooth, generic, and structurally uniform. The recipient responds to the former at roughly 2x the rate.

Practical implication: the person who actually understands the offer — the founder, the head of the function, the engineer who built the product — outperforms the person whose job is to write outbound. This is your unfair advantage as a founder. Use it until you hit $1M ARR. The handful of SDR hires who outperform founders are the ones who have been embedded in the offer long enough to write in the founder's voice. That is rare, and it is the central qualification for any early outbound hire.

What you want to say is: write each email like you would write it to a friend who happens to be the buyer. If a sentence sounds like it came from a sequencing tool — “I hope this email finds you well,” “just wanted to circle back,” “let me know your thoughts” — cut it. If it sounds like something you would say out loud, keep it.

The grammar-and-typo question

Meticulous grammar is a signal. The signal depends on the rest of the message. Meticulous grammar inside a templated structure reads as templated; meticulous grammar inside a specific, recipient-oriented message reads as a writer who cared. The calibration is contextual.

A single small typo — a missing capital, a misplaced apostrophe, a comma in the wrong clause — in a message that is otherwise specific and recipient-oriented reads as human authenticity and modestly lifts reply rate, on the order of 0.2 to 0.4 percentage points. Multiple typos read as low-effort and harm reply rate. The calibration that works in practice: one human-looking imperfection per first-touch, no more. Operators who deliberately insert typos as a personalization signal typically overdo it and end up in the “low-effort” bucket.

The negative-priming question

Should a cold message acknowledge that it is cold outbound, or should it pretend to be something else. The empirical answer, observed consistently across campaigns we have audited: acknowledgment lifts reply rate by 0.5 to 1.5 percentage points on first-touch.

The mechanism: the recipient knows the message is cold outbound within the first two seconds of reading. A sender who pretends otherwise — “I've been following your work for a while” from a sender the recipient has never heard of — is read as dishonest, and the message is deleted on that basis alone. A sender who acknowledges the cold nature directly — “cold outreach, won't be offended if not relevant” — is read as honest, and the message is read on its merits. The acknowledgment lift is largest on senior recipients who have been the target of cold outbound for the longest and whose pattern-match for dishonesty is fastest.

The acknowledgment is not a license to skip the other three principles. The empirical ceiling is a specific, brief, recipient-oriented message that acknowledges its cold nature without apology.

Common operator failures observed in production

• Writing inside the operator's context. The operator writes what they want to say, sends it, observes a 0.3% reply rate, and concludes that cold email does not work. The message was structured around the sender's offer rather than the recipient's context.
• Over-correcting to a 200-word “personalized” message. The operator reads that cold email needs personalization, writes 200 words with three variables and an embedded pitch deck, and observes a 0.4% reply rate. Length killed the message; the variables were correct.
• Using a model without a voice prompt. The output is grammatically correct, structurally smooth, and recognizably AI-generated. Reply rate sits in the 0.5 to 1.0% range — meaningfully below disciplined human-written copy.
• Refusing to A/B test on the grounds that the operator “knows their customer.” A 4-arm A/B across a 2000-recipient cohort produces, in our observation, a 2 to 3x reply-rate spread between best and worst arm. The operator's pre-A/B prediction of the winner is correct in roughly 30% of cases.
• Optimizing the CTA before fixing the body. A heavy CTA on a recipient-oriented body outperforms a light CTA on a sender-oriented body by a wide margin. CTA tuning on a broken body produces negligible movement.
• Treating the first-touch as the only touch. First-touch reply rates of 1 to 2% become cumulative sequence reply rates of 4 to 6% across a disciplined 5-7 touch cadence. Operators who judge the experiment on first-touch alone systematically under-estimate the channel.

Pre-write copy checklist

• One specific detail per message that could not have been written about a different recipient — named company, named feature, specific number, recent shipment, public signal
• First-touch word count between 50 and 90
• Deletion test passed — removing every sender-context sentence leaves a message that still functions
• One explicit ask, phrased as a yes/no question or a binary choice between two specific options
• No generic openers — no “I hope this email finds you well,” no “I noticed your company is doing exciting things,” no “I wanted to reach out because…”
• No more than one em-dash per first-touch unless the rhythm explicitly demands it
• Voice check — read aloud, does it sound like the operator or like a sales development rep
• One acknowledgment line that the message is cold outbound, framed without apology
• Fact-check the specific detail — hallucinated specifics damage reply rate by more than they help
• Subject line under 35 characters, specific, lowercase, no emoji, no question mark (Chapter 2)

Where this fits

The principles in this chapter are the foundation for Chapters 2 through 5. Subject lines (Chapter 2) operationalize specificity and brevity at the 35-character mobile preview. Opening lines (Chapter 3) operationalize recipient-orientation in the first sentence. Value-proposition framing (Chapter 4) operationalizes recipient-orientation across the body. CTA architecture (Chapter 5) operationalizes the explicit-ask principle across a per-touch progression.

The campaign chapters that follow (Chapters 6 through 9) compound this per-touch performance into a multi-touch, multi-channel sequence. The compounding is multiplicative. Operators who skip Chapter 1 and tune the sequence cadence are tuning the wrong variable.

Related chapters

• Cold Email Subject Lines That Get Opened — applies specificity and brevity at the 35-character mobile preview.
• Cold Email Opening Lines That Get Replies — operationalizes recipient-orientation in the first sentence.
• Cold Email Calls to Action — Heavy vs Light Asks — operationalizes the explicit-ask principle per touch.
• Cold Email Sequencing — The 5-7 Touch Cadence — compounds per-touch performance across the multi-touch sequence.