The argument in one line.
Connecting an AI coding agent to a B2B data platform through natural-language goal prompts turns lead sourcing, contact enrichment, and personalized outreach copy into one workflow that costs cents per lead instead of hours of tool-switching.
Read if. Skip if.
- You're doing cold email outreach yourself and keep stalling on finding accurate contact info for the right businesses.
- You already use an AI coding agent for other tasks and want to see it orchestrate a paid data platform instead of writing code.
- You run or want to run an agency and need a repeatable way to generate a fresh batch of qualified leads without hiring an SDR.
- You're looking for a free lead-gen method — this workflow runs on paid Clay credits and an AI coding agent subscription.
- You want inbound or warm-network strategies — this is entirely a cold-outbound data-and-copy workflow.
The full version, fast.
Cold outreach usually breaks on three points: finding the right businesses, getting real contact info, and writing copy that doesn't read as mass AI spam. This video pairs Claude Code, used purely as an orchestrator, with Clay, a B2B data platform whose 'waterfall' checks multiple providers per lead to lift email-match rates from around 30% to 80-90%. A single natural-language goal prompt has Claude Code spin up parallel sub-agents, source and enrich fifty HVAC leads across six cities, and write personalized subject lines and bodies for roughly $12-24 in credits. The resulting CSV imports straight into Clay to launch a warmed-domain email campaign, with no new dashboard to learn — just plain English.
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.

01 · What We're Building
Cold open shows the finished result, then frames the episode's premise: Claude Code as orchestrator, Clay as the data source, run entirely in natural language.

02 · The Data Problem vs The Tool Problem
Cold outreach breaks on three points — finding the right business, getting real contact info, and writing non-spammy copy. The fix splits into a data problem (Clay) and a tool problem (Claude Code).

03 · Why Clay and the Waterfall
Clay blends its own dataset with negotiated access to other B2B data providers, paid for with credits instead of separate subscriptions. Its waterfall lifts email-match rates from ~30% to 80-90%.

04 · Setting Up Clay and the Plugin
Sign up for Clay, grab an API key, then install the Clay marketplace plugin — which must be done from the Claude Code terminal, not the desktop app or IDE extension.

05 · Connecting Your Account
Reload the plugin, ask Claude Code to authenticate with Clay, follow the generated authorization link, and confirm the Clay MCP server is connected.

06 · The Goal Prompt
A single detailed prompt sets the target (50 enriched HVAC/home-service decision-maker leads with emails, pain points, personalized subject lines and bodies) and an explicit end condition, using Claude Code's /goal workflow.

07 · How It Runs and What It Cost
The goal prompt spins up six parallel sub-agents (one per city), each sourcing and enriching 25 leads, then merges, dedupes, and verifies the results. The full run cost about 172 Clay credits, roughly $12.

08 · The Enriched Data and Email Copy
Walkthrough of the 50-row output: verified emails, phone numbers, business pain points, personalization hooks, and a written subject line and body for every lead.

09 · Launching the Campaign in Clay
The CSV imports into Clay as a table, gets attached to a new campaign via merge variables for subject and body, and new sending domains get purchased and warmed up for outreach.

10 · Final Thoughts
Wrap-up noting Clay's own agent connector can't yet manage campaigns end-to-end, plus a plug for the creator's free community where the full project write-up is shared.
Lines worth screenshotting.
- Cold outreach fails for three separable reasons: finding the right business, getting real contact info, and writing copy that doesn't read as AI-generated spam.
- Splitting lead generation into a data problem and a tool problem lets one system own sourcing and enrichment while an AI agent owns orchestration and copywriting.
- A single data vendor typically nets a working email on around 30% of leads; a 'waterfall' that falls through several providers in sequence pushes that to 80-90%.
- One natural-language goal prompt can spin up six parallel sub-agents, each sourcing and enriching 25 leads from a different city, then dedupe and cross-check the merged list.
- A thorough goal-prompt run with multiple verification passes and rewrites took about an hour for 50 leads; a simpler one-shot version of the same request took five minutes.
- Fifty fully enriched B2B leads with verified emails, phone numbers, and personalized subject lines and bodies cost roughly $12 in data-platform credits.
- At even a 1% reply-to-client conversion rate, spending about $24 in credits per 100 leads is cheap next to the value of a single closed client.
- An AI agent can only write outreach that sounds specific to a business if it's first given that business's actual context: offer, case studies, FAQs, proof, and website copy.
- A completed lead-and-copy CSV can be re-imported into a data platform as a table and mapped straight into an email campaign with merge variables, with no re-typing.
- New sending domains need to be purchased and warmed up with a capped daily send volume (roughly 30/day) before they're safe to use for cold outreach at scale.
- As of this recording, the data platform's own agent connector could source and enrich leads but could not yet manage campaign creation or sending — that step still required its own UI.
Split cold outreach into a data problem and a tool problem
Pointing an AI coding agent at a data-enrichment platform turns lead sourcing, contact verification, and personalized outreach copy into one natural-language workflow instead of five disconnected tools.
- Cold outreach has three separate failure points: finding the right businesses, getting real contact info, and writing copy that doesn't read as mass-produced AI spam.
- Splitting the workflow into a 'data problem' handled by a lead-data platform and a 'tool problem' handled by an AI orchestrator means you're not learning a new dashboard for every piece.
- A single data vendor typically nets a working email on roughly 30% of leads; stacking several providers in a fallback sequence pushes that to 80-90%.
- Paying for enrichment by the credit instead of by the subscription means cost scales with actual usage, not with how many tools you're renting.
- Wiring an AI agent into an external data platform's connector removes the need to learn that platform's UI — you describe the outcome you want instead.
- This kind of integration usually has to be installed through the terminal version of the agent, not the desktop app or IDE extension.
- Before an agent can act on your behalf inside a paid platform, you authenticate it explicitly — it doesn't get standing access by default.
- A single upfront 'goal' prompt with target criteria, required fields, and an explicit end condition is enough for an agent to plan its own multi-step approach.
- The agent can split work across several parallel sub-agents (one per segment), then merge, deduplicate, and cross-check the combined results before calling the goal met.
- A run with multiple verification passes and rewrites took about an hour for 50 leads; a simpler one-shot version of the same request took five minutes — the extra time buys accuracy checking, not more leads.
- Fifty fully enriched B2B leads with verified contact data and personalized copy cost roughly $12-24 in credits, which is cheap next to the value of even one converted client.
- An agent can only write outreach that sounds specific to a business if it's first given that business's real context — offer, case studies, proof, FAQs; without it, the copy defaults to generic AI phrasing.
- Personalization that references something concrete and public about the prospect reads as researched, not templated.
- The output of an agent-run enrichment workflow can be re-imported into the sending platform and mapped to email fields with merge variables, so nothing gets re-typed.
- New sending domains need to be purchased and warmed up with a capped daily send volume before they're safe to use for cold outreach at scale.
Terms worth knowing.
- Clay
- A B2B data and workflow platform that sources, enriches, and can send outreach for lead lists, combining its own dataset with negotiated access to many other data providers.
- Claude Code
- Anthropic's coding agent, used here purely as an orchestrator that calls an external tool's actions through natural language and writes the resulting outreach copy.
- MCP server
- The connection layer that lets an AI agent call a specific external tool's actions directly, instead of requiring a custom-built integration for each tool.
- Waterfall enrichment
- A lookup sequence that tries one data provider for a match, then falls through to the next provider on a miss, raising the overall chance of finding a valid contact.
- Goal prompt
- A command that gives an AI agent an end condition to work toward, letting it plan its own multi-step approach and keep running until that condition is verified as met.
- ICP
- Ideal Customer Profile — the specific description of the business or decision-maker a lead search should target.
- Domain warming
- Gradually increasing a new email domain's send volume over time so mailbox providers don't flag its messages as spam.
- Clay credits
- Clay's usage-based currency, spent per lookup or enrichment action rather than paid as a flat subscription.
Things they pointed at.
Lines you could clip.
“The workflow I'm showing you guys today, Clay is going to fix the data problem and Claude Code is gonna fix the tool problem.”
“If you're looking at just one vendor, then you might just get, you know, 30% of your leads, you'll find the right email. But with the waterfall, that number is going to increase somewhere up to more like 80 to 90%.”
“This costed us a 172 clay credits, which is only about $12.”
“Clay's MCP server cannot yet manage all of this. I'm assuming maybe by the time you're watching this, that's already out there.”
Word for word.
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
The bait, then the rug-pull.
In one prompt, Nate Herk has Claude Code find, verify, and personalize cold outreach for fifty leads, turning what's usually a multi-tool grind into a single natural-language workflow.
Named ideas worth stealing.
The three blockers to cold outreach
- Finding the right business
- Getting real contact info
- Writing outreach that isn't AI-generated and spammy
Cold email fails on three distinct points, not one — a useful checklist for diagnosing why an outreach effort isn't working.
Data problem / tool problem split
- Data problem: sourcing and enriching accurate leads
- Tool problem: switching between too many disconnected platforms
The video's core structure: a data platform (Clay) fixes the data problem, an AI agent (Claude Code) fixes the tool problem by replacing UI-hopping with natural language.
The waterfall enrichment model
Instead of relying on one data vendor, the platform checks a top provider for a match and falls through to the next provider on a miss, repeating down the list until it gets a hit — raising match rate from ~30% to 80-90%.
The goal prompt
A single upfront prompt states the target criteria, required output fields, and an explicit end condition ('don't stop until every lead has an email, subject line, and body'), letting the agent plan sub-agents and verification passes on its own.
How they asked for the click.
“Join my free school community. The link for that is down in the description.”
Soft close after the demo, not a hard pitch — frames the free community as where the full project write-up (context files, prompts, outputs) will be shared.






































































