8M Emails Monthly: The Frameworks Worth Stealing

Two Videos, One Clear Pattern

I watch a lot of cold email content. Most of it recycles the same three tips. This week Eric Nowoslawski put out two videos that are genuinely different from the noise. One is a deep operational breakdown of how his agency sends 8 million cold emails per month. The other is a framework for a specific campaign type he calls the creative ideas campaign, which he claims outperforms everything else across every client they run.

Both videos are worth watching. But they are also dense, and Eric talks fast. So I watched them, pulled the parts that matter, and I will tell you where I agree, where I have reservations, and what you should actually implement if you are not running anywhere close to 8 million emails per month.

The throughline between both videos is the same: systematize the inputs so you can focus creative energy on the one thing that actually moves reply rates. Eric has built an entire infrastructure layer to remove the manual grunt work. Then he spends his remaining attention on prompt engineering and campaign design. That is the right order of operations. Most people do it backwards.

Video 1: How to Send 8 Million Cold Emails Per Month

What Eric Is Actually Building

Eric runs Growth Engine X, and the numbers he opens with are real enough to pay attention to. 8 million emails per month, 27.8 million Clay enrichments in a single week, 2 million contacts processed per day. He also mentions that they generated 62,000 positive responses for clients in the past year. These are agency-scale numbers. If you are a solo founder or a small team, you are not replicating this system this week. But the architecture is instructive.

The core problem he is solving is one I have seen kill agencies that try to scale: manual CSV uploads. He had three full-time employees whose entire job was downloading lists from data providers and uploading those CSVs into Clay templates. Every morning his team would start at 8am, begin hitting API endpoints, slam into rate limits, and scramble to keep client campaigns topped off with leads. It was chaotic and it did not scale.

His solution was to use Claude Code to build an internal application that automates the entire pipeline. The system works like this: you set up a search in Prospeo, click one button to get the JSON request, drop it into their internal app, and Claude Code automatically creates the Clay table, pulls the contacts, and stores everything in Supabase. The pipeline then runs at 2am so by the time the team is in at 8am, everything is processed and queued.

The rate limit problem that was strangling them gets solved because Clay's queuing system handles all of that, and by running at 2am instead of 8am they stop competing with their own peak traffic.

The Email Waterfall That Actually Matters

The part of this video I found most practically useful is the email finding waterfall. Eric's team does not just use one email finder. They run a three-step waterfall, and the logic behind each step is worth understanding.

Step one is their internal cache. Every email they have ever found across any client is stored in Supabase with a validation date. If they have found that person's email before and it was validated within the last 90 days, they use it without spending any API credits. If it is older than 90 days, they revalidate it with Million Verifier before trusting it. This is smart and most agencies are not doing this. They are burning API credits on contacts they have already enriched.

Step two is Prospeo's API. Eric is using this as his primary email finder because of speed, quality, and one specific feature: Prospeo validates every email through BounceBan. That means catch-all emails, which normally require a second validation pass, come back with an extra layer of confidence. He says they have stopped double-validating Prospeo emails because of this and have not seen bounce rates increase.

Step three is Smartlead's email finder API. He puts this last because it is slower, but he loves the underlying logic: Smartlead's email data is based on actual sending history from their platform. If another Smartlead user sent an email to that address and it landed, that is a real signal. It is validated by behavior, not just by lookup.

This three-layer approach to email finding is something you can implement at any scale. Even if you are sending 10,000 emails a month, building a cache of previously validated addresses saves money and improves deliverability. If you want to explore email finding options at the tool level, ScraperCity's email finder is one option worth looking at, and for validation ScraperCity's email validator integrates cleanly into similar workflows.

What I Agree With

The forecasting piece Eric mentions is underrated. He points out that when you pull 100,000 contacts from a data source, you are not getting 100,000 sendable emails. After validation you are typically working with 50 to 70 percent of that. So you actually have 50,000 to 70,000 emails. If you are running campaigns with volume targets for clients and you have not accounted for this, you are going to miss your numbers and scramble to fill the gap at the worst possible time.

I have seen this exact problem destroy client relationships at agencies. The agency promises a certain lead volume, does not forecast properly, and then either sends lower quality contacts to hit the number or comes back to the client with bad news. Eric's system solves this by finding and validating emails before they ever enter the sending pipeline, so the number you see in your sequencer is close to the number you will actually send to. That is the right way to operate.

The move to 2am processing is also something I would encourage any agency doing volume to adopt. You are not paying more for it. You are just scheduling your automations to run when API rate limits are less contested. The same work happens, it just happens without your team watching it panic.

What to Skip If You Are Not at Scale

Most of the Claude Code infrastructure Eric describes is overkill if you are under 500,000 emails per month. You do not need to build a custom Supabase database, edge functions, cron jobs, and an internal app to manage your Clay tables. You need a good list, a validated email, and a strong sequence. The infrastructure becomes necessary at volume. Before that it is a distraction.

If you are building out your cold email tech stack and want a clear picture of what tools actually matter at different stages, the cold email tech stack breakdown I put together covers this without the enterprise complexity.

What you can steal from this video right now, regardless of your scale, is the email waterfall logic and the caching concept. Even a simple spreadsheet tracking which emails you have validated and when is a step up from hitting the same API endpoints repeatedly for the same contacts.

Video 2: The Creative Ideas Campaign That Beats Everything

The Framework

This video is more immediately actionable and I think it is the better watch for most people reading this. Eric describes a campaign type called the creative ideas campaign that he discovered while running outreach for Instantly. He has since rolled it out to YC-backed companies, bootstrapped founders, and Fortune 500 businesses. His claim is that it consistently outperforms every other campaign type they run for every client.

The structure is simple. Instead of leading with a generic value proposition, you lead with three specific, personalized ideas for how the prospect could use your product or service. Each idea is constrained to a specific offering, and AI personalizes each bullet point based on the prospect's company description.

Here is the example he walks through using Clay as the hypothetical client. If you are doing outreach on behalf of Clay, you might constrain the three ideas like this: the first bullet is always about building lists of new hires in their market, the second is always about how they could use Clay Agent to surface data points competitors cannot get, and the third references a specific integration like Built With or Predict Leads that seems relevant to their business. The AI reads the company description, figures out who that company sells to, and then applies those three constrained templates to their specific situation.

The result is an email that feels like you did real research on the prospect's business without actually requiring a human to do that research for every single contact.

The Three Rules He Lays Out

Eric is specific about three rules that make this work, and all three are worth taking seriously.

First, constrain the creativity. He made the mistake of rolling this out to a software development client without constraining what the AI could suggest. The AI started fabricating features the agency could build for prospects. The emails got great response rates. The clients could not actually deliver what was promised. That is a fast way to burn client relationships. Constrain each bullet point to something your business actually does.

Second, give the AI real context about your business. A one-sentence description of what you do is not enough. You need to front-load the system prompt with enough information that the AI understands your offerings, your tone, and the range of problems you solve.

Third, and this is the part where most people cut corners: write the examples yourself. Eric says he hand-writes the few-shot examples in the system prompt every single time. He is not having an intern do this. He is not asking the AI to generate the examples. He picks companies that represent the range of his audience, one consumer, one B2B, and he writes what a great version of the output looks like for each. Then the AI learns from those examples and replicates the pattern.

He mentions that when output quality is not where he wants it, he does not change the prompt. He adds more examples. That is the right instinct. Prompt engineering for email personalization is 80 percent examples and 20 percent instructions.

The Cached Inputs Trick

There is a brief but important tactical note buried in this video. Eric mentions that he deliberately overloads the system prompt and keeps the user-facing prompt extremely simple. The only thing that changes between rows is the company description input. He does this to trigger cached inputs on OpenAI's API, which gives you a discount when the prompt is identical across requests. If you are running this at scale, that discount adds up quickly. It also speeds up response times because OpenAI does not have to process the same system prompt from scratch every time.

He does not go deep on this in the video but the implication is worth noting. If you are using GPT-4o Mini or GPT-5 Nano for personalization at volume, structure your prompts so the system prompt is static and the variable input is as minimal as possible. You get lower costs and faster processing.

What I Think About This Campaign Type

The creative ideas campaign works because it solves the core problem with most AI-generated personalization: it is too generic to feel real and too weird to feel human. The three-bullet format with constrained creativity hits a specific sweet spot. The prospect reads it and thinks you understand their business well enough to have opinions about how they should use your product. That is very different from reading a generic pitch that could have been sent to anyone.

I have seen variations of this in some of the best-performing campaigns I have reviewed over the years. The underlying mechanic is show don't tell. Instead of saying we help companies like yours do X, you show them specifically how they would do X. The specificity does the persuasion work.

If you look at what the best-performing templates in our killer cold email templates have in common, this is it. The emails that perform are the ones that make the prospect feel seen. Three tailored ideas does that more efficiently than a paragraph of social proof.

What Eric has done is systematize a creative approach so it can run at scale without sacrificing the quality that makes it work. That is the hard part. Anyone can write a great personalized email manually. The challenge is getting that quality out of an AI across thousands of contacts without it turning into generic mush. The combination of constrained creativity, front-loaded context, and hand-written examples is how he is doing it.

How to Apply This If You Are Starting From Scratch

You do not need Clay to run a version of this campaign. You can set up a similar prompt structure in any AI tool and manually apply it to a list. Here is the process stripped down to its core:

• Pick three specific things your business can do for a client, one per bullet, and define them tightly so the AI cannot improvise
• Write a system prompt that explains your business in detail and instructs the AI to apply those three things to the prospect's specific situation based on their company description
• Find five to ten real companies in your target market and hand-write what great output looks like for each one, then paste those into the system prompt as examples
• For each prospect, input their company description and let the AI generate the three bullets
• Review the first fifty outputs manually before you send anything

That is it. You can run this with a spreadsheet and the OpenAI API directly if you want to keep costs down. If you are using Clay, the setup Eric walks through in the video is fast, and the column structure makes it easy to review outputs before they go into your sequencer.

For the sequencer side, Instantly and Smartlead are both solid choices depending on your volume and how much infrastructure you want to manage. Smartlead gives you more control over the technical deliverability settings if that matters to you.

The Pattern Across Both Videos

Here is what I think is actually going on across both of these videos and why I think Eric's agency is performing at the level it is.

He has made a clean separation between two types of work: infrastructure work and creative work. The infrastructure work, the data sourcing, email finding, validation, deduplication, Clay table creation, Supabase storage, is all automated. His team does not touch any of it manually anymore. Claude Code handles the orchestration. Scheduled jobs handle the timing. The system runs whether or not anyone is watching.

That frees up his human attention for the creative work: designing campaigns, writing examples, reviewing output quality, deciding when a campaign is ready to ship. The creative ideas campaign is the output of that freed-up creative attention. He can afford to hand-write examples and iterate on prompts because he is not burning hours on CSV uploads.

Most agencies and founders do this backwards. They spend hours on manual list building and data work, then rush the campaign design because they are out of time. The output quality reflects that. The emails are mediocre, response rates are low, and they blame the channel.

The real lesson from both videos is not about a specific tool or a specific prompt structure. It is about where you are allocating your creative energy. Automate everything that can be automated. Then spend your human attention on the things that actually determine whether someone replies.

If you want to see specific email scripts built on this kind of thinking, the top 5 cold email scripts page has frameworks you can adapt. And if you are trying to figure out how follow-up fits into a campaign like this, the cold email follow-up guide covers the sequencing logic in detail.

What to Actually Implement This Week

If you are running any kind of outbound campaign right now, here is the one thing worth doing before anything else: set up a version of the creative ideas campaign for your business.

Pick three specific things your service or product does. Not general categories. Specific capabilities. Write a tight constraint for each one so the AI knows exactly what it is working with. Then find ten companies in your target market and write the three bullets manually for each one. Do not use AI for this step. Write them yourself. Once you have those ten examples, you have your training data. Put them in a system prompt, add your business context, and run your next 500 contacts through it.

Compare the reply rate on that campaign against whatever you were running before. That is the test. Eric's data says this wins. In my experience with the campaigns I have seen perform, the specificity-first approach consistently outperforms generic value propositions. But test it on your list, with your offer, before you commit to scaling it.

The infrastructure Eric describes in video one is worth understanding at a conceptual level even if you never build it. The principles, caching validated emails, running processing during off-peak hours, forecasting list size based on validation rates rather than raw contact counts, apply at any scale. You can implement versions of all of them without writing a single line of code.

The tools are available. The frameworks are clear. The question is whether you are going to allocate your creative energy to the part of the process that actually determines your results.