Building Your ICP

Hello Predictable Revenue community,

Book update: I printed 3,000 copies of The Terrifying Art of Finding Customers. So far, I’m tracking ~379 retail print sales (250 in the US, 129 in Canada, and that tracking is partial). On the distribution side, 2,016 copies have shipped out so far: 600 into Canada, 1,344 into the US, and 72 direct. These don’t include international sales, which are reported differently for some confusing reason.

Work update: you probably already know I help coach startups with finding their first customers and building their first sales teams. Well, I’m just about full of clients and need to hire another ex-founder. If you know a founder with a background in revenue that’s in between startups, drop me a line. I’d like to keep PR as a place for founders run by founders.

Newsletter thought - this is a longer one that gets a little sales geeky, hit reply and let me know if this style is interesting or not.

Onto the newsletter…

2025 is the year we started communicating with code. When I'm asking my co-founder to build a new feature, instead of just sending a product spec over the fence, I'm now using Lovable to code something up alongside the spec so he can see exactly what I had in mind. Learning tools like Lovable, Windsurf, and n8n allow me to communicate with much higher fidelity, we move faster, and it seems the bar has been raised across the board.

But when it comes to founders communicating our ideal customer profile to our go-to-market teams, I haven't seen that bar being raised.

I've worked with a lot of founders and companies where the slide deck version of the ICP is something the founder creates, then that gets passed to a marketing person who doesn't understand 100% of it. Even if you explain it well to the next person down the chain, there's probably 20% information loss at each step. Which means by the time the person building the list of accounts gets their hands on it, they're only dealing with part of the picture and need to make their own assumptions.

I've been on both sides of this equation. I've been the one delegating something important without sufficient context and getting back something that was not even close to my original intention. I've also been the guy on the other end asking a marketing leader for clarity on what the founder meant, getting a glazed over head shake, and having to make some pretty big assumptions because they needed the list asap.

I get that startups are hectic and we're all trying to move fast. But there are a few moments when it makes sense to slow down and make sure everyone's on the same page. This is one of them. If you're a founder, investing in GTM, and you're not deeply involved in the list building process, you're doing it wrong.

Fortunately, AI has made this process much easier to communicate in higher fidelity. In the past we had to accept the limitations of your data provider's drop downs. Remember trying to target B2B SaaS companies and being limited by "IT and Technology" as a filter? If you've been at it long enough, you might even recall a time when all we had were SIC or NAICS codes to build our lists.

Now? All you need to do is explain what you're looking for in a target customer to ChatGPT, have it write a Classifier Prompt, and then have a tool like Clay loop that prompt through each account on your list. The end result is you have the ability to filter these lists of companies in any way you can imagine.

Your only limitation is your own creativity.

This post isn't sponsored, I just like the tool and think that more founders should know how it works so they can improve how they communicate with their GTM teams. I do have an affiliate link, but you don't have to click it.

That being said, today's post is about a few of my favorite things in Clay and why founders need to be more involved in the list building process. These are the things that are easy to miss as a new user but make it an extremely powerful tool.

A word of warning. Clay is the gateway drug to the AI automation space. I've been using Clay for a few years now and have gone deep into N8N, windsurf, lovable, and a dozen other tools I didn't know existed a year ago. Clay is like the weed of the AI space. You start here, soon you'll be into harder stuff. Fortunately, it's way more productive than heroin.

Before we get to Clay

Please record your calls. All of them, from early customer interviews to later stage sales calls. And make sure you save the recordings somewhere accessible.

One of the biggest challenges when communicating your ICP is remembering every last detail, especially if what you're doing is evolving week to week. Here's my rough high level process for extracting that information:

1. Record all my calls

2. Take the transcripts and have an LLM summarize the persona, company firmographic details, jobs to be done mentioned, and any indicators that they were feeling the pain we solve.

3. Feed the summaries to an LLM and ask it for a list of the indicators and a way of finding that information using publicly available information only. I find Perplexity is the best at this step.

4. Take Perplexity's summary and use my 4.1 Prompt Optimizer GPT with the instruction to turn it into a research prompt you can use with Claygent. Make sure to tell it to provide explicit instructions for the prompt to source each signal.

5. The first prompt won't be perfect, you'll need to run it, look at the results, and adjust the classifier. I typically will loop through this 3-10 times depending on how good the first pass was.

Recording your calls takes zero effort and you're probably already doing it. Working with the transcripts might feel like a big task but it's really just copy, pasting, and reading. When I ran my SDR agency, I used to pay humans to do this work manually. It used to cost thousands, if not tens of thousands of dollars. Now it'll take you an hour. To quote the youtuber 2 Minute Papers, "what a time to be alive."

I also use a similar process when I'm reviewing our startups roadmap or documenting a new feature request.

What Clay actually is

Most people use Clay as an enrichment pipeline. Data goes in, contact info comes out. Which is fine, but it's the least interesting use case.

Clay is where your ideal customer profile comes to life. It's a spreadsheet that connects to any data provider you could dream of. And they've added Claygent, which are agents that can browse the internet and answer questions about companies that no enrichment provider has ever been able to answer.

It's the best tool I've found for communicating who your best customers are in high fidelity.

My favorite things about Clay

1. Claygent is the secret sauce

Think of Claygent as a web research agent. The input for Claygent is the prompt we put together using my process above. Now we can run it across 25k accounts and filter the output to just the good ones.

This is the thing nobody else has replicated as nicely and neatly as Clay. You can have Claygent scrape a company's website for specific signals, check if they mention certain keywords, find whether they have specific keywords in their job postings, or answer questions that no data provider has in their database.

Your first prompt won't be perfect. You'll need to run it, look at the results, and adjust. I typically loop through this 3-10 times depending on how good the first pass was. Each time you iterate, you're communicating your ICP with higher fidelity than any slide deck ever could.

2. Write to other table + look up data in other table

This is what turns Clay from a toy into a tool that lets you build sophisticated ICP scoring, routing and sourcing.

Here's the power. You can run a research prompt that finds a list of all the companies mentioned on a website, then push that entire list to another sheet. Or extract all their job postings and push those to a separate table. Or pull their tech stack and route it somewhere else.

Then you can look up data across those tables. Cross-reference. Build scoring systems that weight different signals differently.

Your ICP isn't just one thing. It's a combination of signals. Do they have the right size? Do they hire for the right roles? Do they use the right technology? Have they mentioned the problem you solve in their content?

You can have one table that's accounts, another that's people, another that's job postings, another that's technologies used. Push data between them. Look up information across them. Update your classifier in one table and it flows through the rest of your system.

3. Webhooks and HTTP post

Webhooks, for bringing in data from external services and keeping your ICP current.

Hubspot creates a new lead. It sends a webhook to Clay* with that lead's company name and website. Clay runs your ICP classifier, scores the account, finds the right people to target, enriches their contact info, then sends a webhook back to Hubspot with a qualified, scored, and researched account.

You just automated what used to take an SDR 20 minutes per account. And more importantly, you ensured that your ICP criteria is being applied consistently to every single account. Your ICP becomes executable code that runs automatically, not a slide deck that gets interpreted differently by everyone who touches it.

*While you can do this natively in Clay / Hubspot I use n8n to monitor Hubspot for these updates and push them into Clay via webhook because I’m thrifty and you need to be on an expensive plan to enable CRM integrations. 

4. Version control your prompts (manually)

There's no version control in Clay. No undo. No save as.

I have lost so many good ICP classifier prompts because I copy-pasted over them. Here's what I do now: I almost never copy-paste a new prompt into an old enrichment. I always duplicate.

When I'm testing and iterating on an ICP classifier, I'll duplicate the enrichment. Make my changes in the new version. Run both versions on the same data. Compare the outputs side by side.

I'll do the first 10 rows, compare results. Then 20. Then 100. Maybe 1,000 depending on the scale. But I always keep the old version until I'm absolutely sure the new one is better.

This is especially important for ICP work because your understanding of what makes a good fit will change as you talk to more customers.

5. Use AI to create a formula

I don't know regex or SQL and this feature lets me set columns to run only if a certain condition is met. It's a simple but powerful ability. This is the lynchpin that lets you filter out bad fits early.

You define the logic of what runs and what doesn't. Then you can run a large table with multiple enrichments and have it waterfall from a big unfocused list to only running expensive enrichments on your best fit accounts. Start broad with 25k accounts. Run your ICP classifier. Filter down to 5k good fits. Then run more detailed enrichment only on those 5k.

Without this, I’d be copying and pasting into ChatGPT, then trying to reformat the output into whatever syntax Clay expects. Instead, you just describe what you want, click generate, and it writes it.

Start here

Start practicing with pushing, pulling, looking up, and then running Claygents. Those are the big unlocks for building ICP classifiers that actually work.

And if you're a founder who hasn't been deeply involved in list building, start now. Build your first ICP classifier yourself. Even if you hand it off to someone else later, you'll understand what's possible. You'll communicate your ICP with higher fidelity.

Your only limitation is your own creativity.

Collin

PS - I hope you all had a great break and actually took some time for yourselves.

PPS - one last reminder, let me know if you know an ex-founder that would be a good fit to help me help other founders.