You're making this outbound mistake

I built an entire GTM outbound strategy in Claude Code. Then I deleted it and started again from zero.

The build was fine. What I had built it on was wrong.

Apartment in Poblado, Medellín. Blinds down, ten hours at the desk going properly deep on it, and I had pointed every hour of it at a market that was not actually buying from my client.

12 prospects into the relaunch he had a live lead with a VP.

Here is how it happened, because you are probably doing the same thing right now.

Where it went wrong

He handed me his ICP document in week one and I built off it. That is what almost everyone does.

Here is the problem with the ICP doc. It is usually fiction. Not a lie. Fiction. Written a year ago, often by someone who has since left, describing the market the founder wants rather than the market that actually pays him.

Or worse, you build off one good case study.

That is the real trap, and it is an easy one to fall into. A client hands you a story they are proud of. Big logo, clean result, quotable. So you point the campaign at more companies that look like that logo.

But one case study is one deal. It tells you nothing about who is actually easiest to sell to, where the product genuinely fits best, or which market closes without a fight. That is a different question, and there is only one file that answers it.

The file I should have asked for on day one

The closed-won and closed-lost export from the CRM. Every deal. Won and lost. Deal source, contract value, every contact on the deal, every title, and the meeting count per contact.

Almost nobody asks for it. I asked for it late, and it cost me the build.

He sent it over as a CSV. I loaded it straight into Claude Code and started querying it.

Win rate by industry. Win rate by deal source. Contract value against outcome. Then meeting counts per contact per deal, which is the query almost nobody runs and the one that changed everything.

This is what AI is genuinely good for. It will count anything you point it at, across 40 deals and every contact on them, in seconds. The questions are still yours. I knew which ones to ask because I have done this before, and that is the only reason the file gave up anything useful.

Around 40 deals. Roughly 19 won, 20 lost.

Then I threw the targeting out and started again.

What they had actually closed and lost was not the ICP they had written down.

Here is the part I did not expect. When I walked him through it, he was not surprised or defensive. He half knew already. He knew that was the market they were really selling into, he just had never seen it laid out as a file. What he said was that this is exactly how they would have gone about it internally.

That is usually how it goes. The knowledge is in the business. It is sitting in individual people's heads as a hunch, and it has never once been written down in a form a campaign can be built on.

Three things came out of the file.

1. The vertical was wrong

Going in, the hypothesis was DTC health. That is where he believed the market was, and the ICP doc agreed.

The export did not.

His best closed business sat with funded AI companies. B2B SaaS second. The vertical he thought of as home turf was where his losses clustered.

Founders remember the deals they enjoyed. The spreadsheet remembers the deals that paid.

2. The buyer was not the buyer

He told me they sell to the VP of Marketing. His whole team told me the same thing.

So I had it count meetings per contact across the won deals.

The VP appeared in 1 meeting.

On those same deals, a Senior Paid Social Manager sat in 44. A Senior Director in 40. A Performance Marketing Manager in 23. A Senior Data Scientist in 19.

Then I ran the identical count across the losses. On a lost deal, no contact anywhere cracked 6 meetings.

That is the route into the account.

The VP signs. A senior manager or a director carries the deal to him, internally, in rooms you will never sit in. Build your list off "we sell to VPs" and you have targeted about a tenth of the people who do the real work of closing you.

So the list got two title tracks per account. Buyer track: VP and above, the people who sign. Champion track: Director, Senior Manager, Performance, Growth, Paid, UA, plus the marketing data people.

Two tracks means two messages. That distinction matters more than it sounds, and I will come back to it, because it is exactly where the AI got it wrong.

3. Warm did not mean qualified

I split the wins and losses by where the deal came from.

Investor and advisor introductions: 3 won, 8 lost. The worst source in the file.

Partner and operator introductions: 4 won, 1 lost. The best.

Organic LinkedIn inbound: 4 won, 2 lost. Content quietly beating the investor network.

Everyone treats a warm intro as free pipeline. This file says a warm intro from the wrong referrer is worse than a cold list, because you spend six weeks being polite to a deal that was never closing. Warm tells you how to reach someone. It tells you nothing about whether they will buy.

There was a fourth finding I will not spend long on. The losses clustered above a certain contract value and the wins clustered in a band well below it. Which tells you the pricing conversation is a targeting problem long before it is a sales problem.

Then Claude wrote the messaging, and it was wrong

Claude Code drives this campaign. It builds the lists, it drafts, it does the volume work. But AI is terrible at knowing what good looks like, and it always marks its own homework.

First round of messaging came back and it read fine. Clean, confident, no obvious errors.

It was not how these people talk about their problems.

Both tracks were being sold the same benefit, and those two people do not want the same thing at all.

The champion cares about ROAS. He cares about finding budget that is currently being wasted, and about not defending numbers he already knows are wrong.

The VP cares about visibility and confidence in the data. Because he has to sit in front of a board or a CFO and be believed.

Same product. Completely different reason to reply.

I only caught that because I had done the work to understand the offer and how the problem lands on each of them. Skip that and you approve the draft, because nothing in it looks broken.

So I went back and queried it. These are not the right examples for the champions, here is what they actually care about, here is what the VP is actually afraid of. Rewrite against that.

How I get inside the buyer's head

Before any of the messaging, I have it play the ICP and then I interrogate it:

→ How does this person think about the problem

→ How does it show up in their business day to day

→ What does it cost them personally when it goes wrong

→ What are they seeing right now

→ What do they care about, and what do they genuinely not care about

→ The lexicon: the actual terms in this niche, last-click attribution, MMM, incrementality

→ Old way versus new way, and where the market currently sits between them

→ The exact words they use for the problem out loud

Then I read every document it produces, check every file I feed it, and check them against each other for consistency.

People ask me what my tell is for spotting when AI is confidently wrong.

There is no tell.

You have to know the space. That is the whole answer. Working with AI now requires you to have enough knowledge to push back, otherwise you are just being dictated to, and it will happily hand you the medium of how to approach a company along with everything else. A "this is great, Luke" from a model that has never sold anything is a mirror.

The rest of the build

→ The list. Roughly 90% of outbound performance. Industry, then persona, then a set of criteria for what a good pull looks like. For this one: Series A funded, around $50k a month in ad spend, already talking about attribution, and active on LinkedIn in the last 30 days. That last one matters more than people think, because otherwise you are messaging dead accounts. Save the criteria and you can pull more like that on demand, forever.

→ Messaging. Across the last 100 campaigns I have seen almost no difference between one-line personalisation and a clean template that reaches the right person with a clear offer. One-line personalisation is the new slop. Everyone is doing it, so nobody is reading it.

→ The handoff. The whole engine got saved as a Claude skill so his team runs it without me. He keeps me for the thinking. His team presses the buttons.

The rebuild broke most of the rules the GTM crowd on LinkedIn will sell you. The stack was three tools and none of them was an AI SDR. I handwrote every message myself, and the strategy presentation with them.

Why most agencies never do this

Because it is slow and it does not look like work.

They take what Claude gives them and ship it. They never ask the client for the actual data. They never spend the time to properly understand the offer. And they never come up with their own idea for a way into these accounts, because the model will offer them one for free in four seconds.

That is the whole gap, and no tool closes it.

Ten hours of build, deleted, because I started in a campaign builder instead of a spreadsheet.

Reply with "STRATEGY" and I will send you how I build an outbound strategy from zero, and the order to do it in.

-Luke