Learn / AI and automation
6 ways we use Claude Code for cold email, end to end
We send over 2 million cold emails a month and booked 3,000+ sales calls for clients in the past year. There is no version of that where a person does the admin by hand. Every Claude Code cold email workflow on this page exists because some step in the pipeline was quietly eating an afternoon a week and producing nothing.
Six of them, in the order the pipeline runs: pulling campaign analytics, loading campaigns into the sequencer, generating offers, writing copy, building lead lists, and everything that fires after a call books. This is the written version of the training below, and the hub for every Claude Code build we publish.
The six, and what each one plugs into
Setup is the same shape every time: take the API key from the tool you want to reference, load it into Claude Code, ask in plain language.
| Use case | What it connects to | What you hand it |
|---|---|---|
| 1. Campaign analytics | Your sequencer API | A question about a month, a workspace, or an angle |
| 2. Campaign uploads | Your sequencer API | Campaign name, lead list, sender emails, copy, schedule |
| 3. Offer generation | Offer frameworks plus every campaign you have run | The website and the industries you are targeting |
| 4. Email copy | Sequencer API, your winning scripts, a spam check | The offer and research from step 3 |
| 5. Lead lists | A data provider CLI | Titles, seniority, company size, location, tech stack |
| 6. Post-booking chain | n8n, calendar, CRM, agreement tool, Slack | API keys and a description of what should happen |
1. Campaign analytics in minutes, not afternoons
Reporting is where momentum dies. You tab between the sequencer, the CRM, and a spreadsheet, copying numbers across to work out what is working. Most people never reach the comparative version, so the data sits there while you fund segments that were never going to land and pull the ones that needed two more weeks.
With the sequencer key loaded, the prompt was one sentence: pull the campaign analytics for everything that ran in August 2025 under this workspace, and name the top performing campaign.
A few minutes later: about 100,000 emails sent, a 2.15% reply rate, 113 interested replies, roughly a 10% interested rate for the month. Top campaign was to influencer marketing agencies, sent August 12, 12,000 emails, 19+ interested replies. Runners up were podcasting, accounting, and SaaS. From there you can go as granular as you want: why that angle beat the others, which vertical replied fastest. All of it was always possible. It just cost hours, so nobody did it.
The numbers were always there. What changed is the cost of asking.
2. Loading campaigns without touching the sequencer
Think about what loading a campaign really takes. List ready, copy written, sequence built, then an hour uploading contacts, mapping fields, and checking everything landed where it should. None of it requires your brain. It is a transfer of information from one place to another, and every hour of it is an hour not spent growing the business.
Ask Claude Code what it needs and the list is short: campaign name, lead list, sender emails, sequence copy, schedule. The schedule and settings are presets, identical on every campaign we launch, so the only moving parts are the list and the copy. In the run it built the campaign, mapped the leads, dropped in the subject line and both emails, set the sending accounts, and left one thing to do: press launch.
Then it compounds, because the same connection answers better questions. Pull the leads nobody has contacted in 3 months, re-verify them through something like MillionVerifier, load them into a fresh campaign running the top angle from step one. Same API key, smarter question.
3. Offers, generated against every campaign you have run
Your copy is only as strong as the offer behind it. The cleanest email in the world gets ignored if the offer does not speak to a pain the prospect has right now. Most people reuse one offer across every campaign without testing it, or burn a week reverse engineering what their buyers care about.
The offer prompt has real context behind it: the frameworks from the handful of people who genuinely understand offers, plus every winning campaign we have run across every client. It is not writing from the general internet. It is writing against what has already converted.
The input can be as small as the client website and the industries you want to hit. In the run it did the market research on its own first: what those buyers care about, what has failed them before, the pain points, the dream outcome, cross-referenced against past campaigns. Out came three front-end offers for a CRM platform selling to sales leaders. The tool consolidation play: we pull every tool your reps touch daily and show you where you are paying twice, and you keep the one-page map whether we speak again or not. The ramp time killer: what your top three reps do differently in discovery and deal progression. And pipeline risk scoring: every open opportunity scored against six risk factors.
You are not picking the winner at your desk. Run them, let the market do the talking, and feed the reply data back into the context so the next round starts smarter.
4. Email copy that is written from your own campaign data
This is where the hours actually go. Done manually you start near scratch every campaign, and writing, testing, and iterating one takes an afternoon even when you are good at it. The usual AI shortcut is a prompt with no idea what has worked for you, which produces slop you can smell from the subject line down.
Context is the whole difference. The copy prompt reads your best performing scripts through the sequencer API, the offer and research from the previous step, your buyer, and your tone, all at once. Then a second pass runs a spam guard over the output so nothing ships carrying words that keep you out of the inbox.
The run produced nine scripts: three offers, three angles each. It pulled the top framework from adjacent verticals, the top offer types, the average word count of winners, the subject line patterns AND the anti-patterns from losers, then spintaxed the bodies so a few thousand sends are not a few thousand identical emails. One of the nine, in full, under the subject line "quick audit":
{{first_name}}, if I could pinpoint exactly where your revenue team's tech stack is duplicating functionality and quietly draining rep selling time, and hand you a one-page consolidation map in 20 minutes, would that be worth a look?
Short, specific, one micro-action, and it answers "am I about to get sold something" inside the same sentence. The other eight run the same shape at 30 and 25 minutes against different value props. Personalization ideas come last, layered in only after the offer proves it works. The evidence behind why those shapes win is what 1.5 million sends say about copy.
5. Lead lists through a CLI instead of a credit meter
A few years ago everyone swore by Apollo, and you were capped by its data and the honesty of its filters. Then Clay, a real upgrade that also costs real money and real setup before you get a usable output. Claude Code connects to everything in the stack instead, so you pull, scrape, verify, and enrich in one workflow without paying a credit per row.
For this one we drive the AI-Ark CLI: find people by title, seniority, skills, and location; find companies by industry, tech stack, revenue, and headcount; bulk export with verified email discovery; enrich from a LinkedIn URL; reverse lookup; check credits. All of it from the command line, all of it promptable.
The prompt started deliberately vague: sales leaders at enterprise-size companies. It pushed back, because "sales leader" is a category, not a filter. Narrowed to VP of Sales and Director of Sales, 500+ employees, verified emails, and a tech stack filter for the incumbent CRM, which quietly turns the list into a competitor takeout campaign. Back came the full record on each person and company, a top 50 by fit, and its own observations flagging which results looked off target. Then it exports through enrichment into the sequencer, no CSV shuffle. The step-by-step version of the list half is the Apollo and Claude Code build.
6. The post-booking chain nobody builds
Most people stop thinking at the booked call. Reply came in, invite went out, lead confirmed. Then a human opens the CRM, updates the deal stage, sets the pre-call reminder, creates the agreement, logs the contacts off the email thread. At 50 or 100 calls a month that is hours of data entry done by the people who should be closing.
Be clear about the division of labor: n8n does the automation. Claude Code builds the workflows, the part that used to need a developer or an agency. I knew nothing about automations. I handed over API keys and described what should happen.
What runs now: the calendar webhook fires, contact, company, and deal details get written into the CRM, a deal is created at discovery call booked, and a task series spawns behind it. Pre-call text the morning of, LinkedIn connection, then post-call tasks for the notes, the recording, and the follow-up booking. When the follow-up books, the deal moves itself to strategy call booked, and Slack gets a notification on every stage change.
The favorite is the agreement flow. Drag a deal into agreement sent and it reads the call transcript and notes, updates the agreement to the terms actually agreed on the call, drafts it, and pushes it to Slack for review before anything reaches the prospect. A person still approves it. Nobody retypes anything.
What this does not do
Two honest caveats. Claude Code has a real learning curve, and standing these up from scratch takes hours, more likely days. And none of it decides anything: point it at a weak offer and it produces a beautiful campaign for something nobody wants, aimed at people who were never going to buy. The tooling compresses the admin half of outbound. The strategy half is still yours.
Two neighbors, since each is its own build: the sending infrastructure underneath all of it is the infrastructure walkthrough, and one campaign taken from blank page to live sends is the end-to-end AI campaign build.
tl;dr: give Claude Code the API keys for your sequencer, your data provider, and your CRM, plus real context on what has already worked. It runs the admin half of outbound in minutes. You still own the half that decides whether anyone replies.
FAQ
What is Claude Code and why use it for cold email?
Claude Code is an AI coding agent that runs in your terminal. For outbound it matters because you can hand it an API key and it can then read and write in that tool directly: your sequencer, your data provider, your CRM. Instead of tabbing between five dashboards, you ask for the thing you want in plain language and it does the clicking.
Do I need to know how to code to run these workflows?
No. Every workflow on this page was built by describing the outcome and handing over API keys, including the automations, which I knew nothing about before starting. What you do need is patience: there is a real learning curve, and building these from scratch takes hours, more likely days, before the first one runs cleanly.
Is Claude Code cheaper than Clay for building lead lists?
For sourcing, usually yes, because you are not paying credits on every row you pull. Clay still earns its keep on enrichment and personalization once the list exists. The setup we run drives a data provider CLI for the pull and the filtering, then pushes into enrichment only where something is actually missing.
Can AI write cold email copy that does not sound like AI?
Only if it is writing from your data instead of the general internet. A generic prompt with no context produces slop you can smell. The copy prompt we run reads past winning campaigns through the sequencer API, the validated offer, the buyer research, and the tone rules, then runs a spam pass over the output before anything ships.
What can Claude Code not do for cold email?
It cannot decide anything. It will happily generate a beautiful campaign for an offer nobody wants, aimed at people who never buy. The tooling compresses the admin half of outbound: pulling, formatting, uploading, wiring. The strategy half, what the offer is and who the buyer is, is still yours, and it is the half that decides the reply rate.
Which tools do you connect Claude Code to?
The sequencer for analytics and campaign uploads, a lead data provider CLI for list building, a verification tool before any send, the CRM for deal stages and tasks, the calendar for booking triggers, the call recorder for transcripts, and Slack for notifications. n8n sits in the middle of the post-booking chain and Claude Code builds those workflows.
If you would rather have the whole machine built and run for you, infrastructure, lists, copy, and replies, that is what we do: 3,000+ sales calls booked for clients in the past year. Or keep building it yourself: every training by topic is on the trainings hub, and the full video is on YouTube.
PS - none of this is what produced 240+ positive replies in a single week. The offer and the list did that. The tooling just meant nobody lost an afternoon uploading it.
