How to Build Custom AI Agents in Clay

Hey, GTM Engineers!

What’s the secret to build great AI prompts for Claygent (or in general)?

Let me take you behind the scenes of how I do it for customers at Clay. This is a process I’ve developed that’s been consistently produces great outputs.

But first, did you catch wind of the GTM Engineer content competition?

The short of it is my team ‘Screaming Seagulls’ will be dropping template, prompts, and our best Clay tips. Here’s who to follow on Screaming Seagulls so you don’t miss the action:

• Veronica Steiner 

• Joseph Fiorie 

• Marilin Berg 

• Rahul Mehta 

• Sabrina Glaser 

• Tong-Tong Li 

• Travers Nammack

Back to this week’s post. These are these steps and general process I take when building custom Claygents for customers:

1. Objective

2. Possibilities

3. What are the steps to get there?

4. Considerations

5. Build prompt

6. Test prompt & spot check

7. Iterate with your AI side kick

8. Repeat step 6

9. Organize

10. Add to Prompt Library + Create Template

11. Other Tips for Fine Tuning

This may seem like a lot, but it can be pretty fun and you’ll almost always learn things along the way.

1. Objective

What do I ultimately want to achieve with the prompt I’m developing?

In the context of PLG companies - I want to figure out all ways I can derive or infer insights from an email address to use for sales and marketing purposes.

2. Possibilities

Today I validate emails in Clay to figure out if they can be used, if they are personal, or work emails, and then use the email in a waterfall enrichment to find a LinkedIn profile. But, am I missing anything?

Turns out I was missing a lot. This is what I asked ChatGPT (I also use Perplexity for similar questions):

This is what I got back:

Useful Outcomes

Most people just classify personal emails vs. work emails.

But now that we have additional insights, we can build out a more powerful workflow/prompt.

This is how I’ll use the insights:

• Classify the email into one of the following

  • Personal Email

  • Work Email

  • Service Email

  • Cannot Determine

• Identify service accounts to:

  • Reduces noise for sales reps

  • Tells us what department is using the service

  • Now that we have the department, we can source people in that department

  • The use of a service email tells us this could be more significant usage compared to an individual user

• Extract the name + company (if it’s a work email)

  • This gives more additional inputs to improve coverage and accuracy of finding LinkedIn URLs with Claygent

    • The use of this approach in conjunction with waterfall for finding the LinkedIN URL increases coverage

3. Steps

Now lets figure out what steps the agent needs to take to get us to the above.

There’s 2 ways I do this:

• Use Clay’s Meta Prompter to get a baseline prompt

• Ask Perplexity or ChatGPT what the steps would be so we can include in the prompt

4. Considerations

This prompt likely doesn’t need any. However, sometimes I ask AI separately what constraints, guidelines, or simply “what else am I missing?”. This adds another layer of learning and works great for building more complex chain-of-thought prompts like the previous one I shared to find GTM Ideal Customer Signals.

5. Build Prompt

I do one of 2 things:

1. Use Clay’s Meta Prompter (see below)

2. Prompt ChatGPT or Claude

Clay’s Meta Prompter 👇️