There is a specific frustration that runs through most cold email agencies: the humans spend most of their time on work that doesn’t require human judgment. Researching prospects. Sorting replies. Writing the same email in 40 slightly different ways. The actual hard part—figuring out what a person cares about and saying something they’ll respond to—is buried under hours of mechanical work.
A cold email agency owner on Reddit shared how they rebuilt their operation around Claude. Not to replace their SDRs. To give them leverage.
The Framing That Changes Everything
The headline insight: Claude is a leverage tool, not a writer.
Most people who try AI for sales use it to generate emails. The agency owner tried this early. The result was reply rates that looked the same as before—or worse—because mass-AI-generated copy is detectable and forgettable.
The approach that worked was different. Humans write the emails. Claude handles everything that surrounds the email: research compression, reply sorting, quality checks, and campaign planning. With that division of labor, each SDR went from handling roughly 80 accounts per day to 220.
Six Workflows That Actually Moved Numbers
Pain point extraction from job postings. Paste a job description into Claude with a prompt asking for the single biggest operational pain the role was created to solve, which tools the company is betting on, and one specific outreach angle. The instruction “be specific, don’t give me generic stuff” is built into the prompt. This becomes the research foundation for the first email—done in two minutes instead of twenty.
The de-AI-ify pass. The SDR writes the email. Then Claude does one pass with strict rules: no em dashes, short fragments allowed, no words like “leverage” or “streamline,” the CTA must be a real question, max 65 words. The prompt also asks Claude to explain what it changed and why—which turns each pass into a lesson the human can absorb and apply without Claude next time.
Reply classification at scale. Batches of 30–50 replies classified into seven buckets: INTERESTED, REFERRAL, NOT_NOW, HARD_NO, UNSUBSCRIBE, OOO, QUESTION. Each with a confidence level. The NOT_NOW bucket is where the real money lives. Tag every one with the month they mentioned, set a calendar reminder. The agency closed a $24,000-per-year deal in November from a February reply that said “circle back in Q1.”
High-value account research. For priority targets, a research prompt asks Claude for three signals this account needs the offer, three signals they don’t, recent strategic moves, the specific person most likely to feel the pain, a non-obvious angle competitors wouldn’t notice—and critically, “a reason NOT to pursue them.” That last question catches accounts with red flags (recent layoffs, just bought a competing solution) before an SDR burns hours chasing them.
Variant generation by hypothesis. Instead of “give me 20 subject lines,” the prompt asks for five variants each testing a different psychological mechanism: curiosity gap, pattern interrupt, specificity, social proof, direct value claim—with an explanation of what behavior each optimizes. This is A/B testing with intent rather than random variation.
Pre-flight campaign quality check. Before a sequence goes out, Claude roleplays as a skeptical VP of Marketing receiving 40 cold emails a day. For each email: what would you actually do (delete / ignore / reply / read fully), the single weakest sentence, what screams “automated,” what claims you’re skeptical of. “Be harsh. I’d rather you tear it apart now.”
The Finding That Surprised Them Most
The highest-reply campaign from this agency wasn’t built on individual personalization. It was built on writing one email per segment that addressed what everyone in that segment had in common.
Mass AI personalization—one email per LinkedIn profile, each slightly different—performed like noise. One email written for a specific shared pain in a specific role, sent to 100 people who had that pain, performed at 11%. The insight is counterintuitive: personalization at scale is less effective than specificity at the segment level.
Who This Is For
This isn’t a playbook for a solo founder sending 20 emails a week. The leverage compounds when you have SDRs already doing research and reply management at volume—and you want to multiply what each person can handle without hiring more people.
But the individual workflows are usable immediately. The de-AI-ify pass, the NOT_NOW classification, the job-posting research prompt—any of these can be added to a sales process today with nothing but a Claude subscription and copy-paste.
