How Temp Impacts AI Email Copy

In partnership with

Hey, GTM Engineers!

Want to know the secrets to getting better AI generated email copy?

There’s a few suggestions that I give everyone:

• Giving AI a persona — “your an ad copy writer with 10 years of experience”

• Adding constraints — “the email should be no more than 4 sentences.”

• Providing examples — “this is an example outbound email [insert email]”

• Giving AI a system prompt — added context on the subject matter

• Explaining style & tone — “the tone should be friendly, and style succinct”

• Providing a list of what NOT to include — “don’t start the email with ‘I hope all is well’”

These are all straightforward to understand.

But, what about temperature?

You may have seen this as an option in the LLM or Claygent that you’re using.

This is a variable you can experiment as it impacts the output in several ways.

This post will help you better understand how you can play around with temperature to dial in your outputs.

P.s. Connect with me on LinkedIn.

🎯 Objective:

Understand how AI temperature setting impacts the output of AI generated copy.

To test this, we’ll take a newsletter post and then create a 4 sentence summary that is written in ad style copy.

We’ll generate a variations with the following different temperature settings:

• Very Low

• Low

• Medium

• High

• Very High

⁉️ WTF is AI Temperature?

In short, temperature is the degree of creative freedom and accuracy that you give AI.

⚙️ Workflow

1. Copy link to newsletter into Clay

   • Alternatively you could send a blog post (when it’s published) into the table automatically via webhook monitoring

2. Claygent will find the latest blog post & outputs

   • Title

   • Subtitle

   • Link to Post

   • Author

3. DeepSeek reads blog post and

   • Creates copy based on a prompt

   • Provide variations by repeating prompt with different temperature inputs

We’ll be using my last post on claymation for the example:

📹 Table Walkthrough

🔉 The Outputs & Table Template (you can copy):