# The 2x Reply-Rate Teardown > Doubling reply rate on a live campaign. Part of The Seqnced Archives: https://archive.seqnced.com/ i gave myself 7 days to send 250,000 cold emails, handing a big part of the process to Claude Code, to find out once and for all whether the hype is real. i send 1.5M+ cold emails a month and booked 3,000+ sales calls last year, so cold email isn't new to me. handing the ops to AI was. this is the whole run, start to finish, exactly as it happened, including the deliverability disaster that nearly sank it. what's inside: → the 7-day infra plan at a glance → the AI Arc look-alike list method (5 companies to a 10,000-person list) → the /offers workflow that picked the front-end offer → the /copy workflow that wrote and de-risked the emails → the .info disaster and the 5-minute recovery → the real end numbers → the volume lesson that carried the whole thing we didn't hit 250k. we'll get to why that didn't matter. ## the 7-day infra plan at a glance 250,000 in 7 days is 50,000 per day for 5 days. that much volume from one or two accounts gets flagged instantly, so you spread it across a network of fresh domains and inboxes. the plan at a glance (Outlook via Azure for high volume, cheap): → 1 Azure tenant needs 1 domain and gives you 100 inboxes. → roughly 5 cents per inbox. → to hit 50,000/day: about 100 tenants, 100 domains, 10,000 inboxes. the setup itself is one command. i ran /inbox kit inside Claude Code (connected to their MCP), told it the scale, the persona, and the domain angle, and it found look-alike domains of our main domain, set the redirect, and spun up the inboxes. the provider handles the DNS setup and runs the warmup automatically before anything sends. that's the whole infra step from the operator's seat. this is a plan-level view on purpose. i'm not walking you through DNS records or warmup mechanics here. that's a rabbit hole, and it's not the point of this teardown. ## the AI Arc look-alike list method infra ready, now you need people to email. this is usually the slowest, most expensive step. Claude Code collapsed it. instead of jumping between a scraper, an enrichment tool, and a verifier, i connected Claude Code to the AI Arc CLI (a database like Apollo or Sales Navigator, we like it for data freshness and full API access) and told it who to reach. the method: → run a look-alike search off 5 ideal-fit customers. → feed it 5 domains similar to the company you want to emulate. → start with the company list, then build the prospect list off it. what one segment returned: → 4,000 records at the founder level → 3,200 at the marketing level → 3,500 at the business development level → 10,000 total leads that's one segment of several you'd stack to feed a 5-day send at this volume. one prompt, no tool-hopping. ## the /offers workflow a list with no offer is dead on arrival. this is another step where Claude Code did the heavy lifting. i opened the /offers command inside the same project, so it had context on exactly who we were targeting (shipping and logistics companies). it has all my context on front-end offers that get replies. i asked it for something desirable i could give as an outbound agency, free or low cost. it ran through our offer docs (Million Dollar Offer Formula, Offers That Win) and our winning campaigns, did market research on the pain angles, scored them psychologically, and produced 3 front-end offers: 1) an ICP lead pull (a list of their most ideal customers) 2) a competitive outbound breakdown 3) a 10-day pipeline audit the value-equation pick was offer 1. run it through the levers: → dream outcome: the exact ideal customers they want to reach. → perceived likelihood: high, you're handing it straight over. → time delay: low, they get it right away. → effort and sacrifice: near zero, you do all the work. the prompt landed on the same answer: offer 1 wins because it's deliverable in about 20 minutes on infrastructure we already have, and almost nobody else can match it. a strong intent-signal list is genuinely hard for a competitor to replicate. ## the /copy workflow offer set, time to write. i ran the /copy command, trained on the script strategy that's been working, and had Claude do two things at once. first, it added two personalizations to the AI Arc list: the prospect's ideal customer (the exact person they sell to) and the service they provide. those two variables carry the whole copy. the base line: 'first name, if i could connect you with 50 of your ideal customers who are actively looking for {service they provide} by the end of the week, would that be worth the 15-minute call?' it produced 7 variations, a mix of one-liners, two-liners, and one-liner-plus-PS. a few that painted the picture: → 'first name, finding ideal customers actively sourcing {service} takes weeks of manual research. if i could do that in 48 hours and hand you 50 of them, verified and ready to go, would that be worth a quick call?' → 'most logistics companies we talk to spend 6 to 8 hours a week building prospect lists by hand. if i could hand you 50 ideal customers actively sourcing {service} by end of week, would that be worth a quick conversation?' three things made the copy work: 1) live data, not a stale prompt. it pulled the current top-performing copy straight from our Email Bison instance, so the writing referenced what's winning right now, not what worked two months ago. outbound moves fast. 2) implied relationship over CSV. 'if i could connect you with 50 people actively looking...' implies you'd facilitate the connection. that's worth far more than 'here's a lead list,' because a CSV on its own reads as low value. 3) an anti-spam pass. a script runs on every piece of copy, scans for spam words, strips them, and spins variations. between that and the two personalizations, each message lands nearly one-to-one. on split-testing: with a 10,000-person list and 5 variations, you put each variation in front of about 2,000 prospects, enough to validate the direction. and when you test, test big. 3 to 4 totally different angles, not tiny word tweaks. find the winning angle first, then hyper-optimize. one opinion on social proof: lines like 'four meetings booked in two weeks' aren't needle-moving, and they can raise sales resistance by signaling you're selling. if the list and the person are right, a clean low-friction ask usually beats stacked proof. ## the deliverability disaster and the 5-minute recovery here's the part most case studies edit out. the good news at the mid-point: about 100,000 emails out, 20-30k a day. 32+ interested responses, 1 to 4 meetings booked every single day. the bad news: a 0.36% reply rate. that's rough. even a bad campaign usually sits at 0.8 to 1%, average is 1 to 2%, good is 3 to 4%. the cause: most of our inboxes were on .info domains. the exact week we launched, one of the biggest blacklists wiped a mass of .info domains and listed them. .info had been a cost hack (about a third the price of .com or .co) and, like every hack, it ran out. the twist: normally you flip from your active infrastructure to a backup set the moment deliverability drops below 1%. our backups were on .info too, so both sets were compromised at once. the recovery, in about 5 minutes: back into Claude Code, one instruction, 'we need 50 more Azure tenants, set them up on .co domains for our outbound.' it came back with 50 fresh domains to purchase and stand up via Azure. one 'go ahead' replaced 25,000 daily sends of infrastructure. the lesson, not the how-to: → keep a genuine backup set that doesn't share the same weakness as your primary. same TLD, same everything, is not a backup. → don't build your whole operation on the cheapest shortcut. it works right up until a blacklist decides it doesn't. ## the real end numbers where it landed by the end of the week. → the reply rate roughly doubled, from about 0.36% to almost 0.66%, once we got off the dead infrastructure and nailed a couple of campaigns. → 98 interested responses total. → the best campaigns carried it: one hit 1.13% reply at 31% positive, and the very first campaign (a ~5,600-lead subset) hit 2.04% reply at 22.41% positive. → about 20 booked meetings in just over a week, 3 to 4 on the good days. → about a 70% show rate, so roughly 14 qualified conversations that actually happened. → about a 20% close on those, which is roughly 3 deals closed in a week. we did not hit 250,000. we got roughly 100,000 out before the reset ate into the volume. and that's the honest headline: worst-case deliverability, a mid-run infrastructure fire, and it still produced about 20 booked meetings and 3 closes in seven days. on a high-ticket or multi-month offer, that math works. ## the volume-vs-precision lesson the approach is what carried this, and it matters more than any single number. at this volume you only need one or two strong campaigns to carry the results of all the rest. the first campaign alone drove a big share of the interested responses. more volume means more campaigns, which means more shots on goal, and you only need a couple to hit. that's why volume beats precision early: → low-volume sending has almost no margin for error. one bad week and you've got nothing. → high-volume sending gives you room to operate through the ebbs and flows. a blacklist event, a weak angle, a dead campaign, none of them sink you, because something else is carrying. so the order is volume first, precision second. get enough shots up to find the one or two angles that work, then tighten. two honest caveats: → cold email is far from dead, and you don't need a huge team or budget to run it. → but don't blindly hand your whole operation to Claude Code. it does the heavy lifting on infra, lists, offers, and copy, and you still need human oversight and input on top. i gave myself 7 days to send 250,000 cold emails, handing a big part of the process to Claude Code, to find out once and for all whether the hype is real. i send 1.5M+ cold emails a month and booked 3,000+ sales calls last year, so cold email isn't new to me. handing the ops to AI was. this is the whole run, start to finish, exactly as it happened, including the deliverability disaster that nearly sank it. what's inside: → the 7-day infra plan at a glance → the AI Arc look-alike list method (5 companies to a 10,000-person list) → the /offers workflow that picked the front-end offer → the /copy workflow that wrote and de-risked the emails → the .info disaster and the 5-minute recovery → the real end numbers → the volume lesson that carried the whole thing we didn't hit 250k. we'll get to why that didn't matter.