AI PRDs: Everything You Need to Know

How to write PRDs in the AI Era (for AI and non-AI features)

The Head of Product for Google Gemini posted that Google is moving from PRDs to prototypes:

This post stirred quite a response. So I wanted to get the truth of the matter: Is there really a role for the PRD in this era? And if so, what should it look like?

That’s the subject of today’s newsletter.

---

I have teamed up with someone at the epicenter of AI: Miqdad Jaffer at OpenAI.

Miqdad runs the AI PM Certification, which I took myself in 2024 and enjoyed. I recommend it, so I secured $550 off his next cohort for my subscribers:

---

Today’s Post

We’ve put together everything you need to know about AI PRDs:

1. Why a spec is still a great document

2. The history of the PRD and how it lost its way

3. The value of a PRD and a modern AI PRD template

4. How PRDs interact with vibe coding/ AI prototyping

5. Steps to write a PRD in 2025

6. Real-World Tear Downs

---

1. Why a Spec is Still a Great Document

Here’s just a few of the things a vibe coded prototype doesn’t specify:

1. What is the hypothesis of the change?

2. How does this fit within the overall strategy?

3. Is it going to be rolled out to everyone or A/B tested?

4. What constitutes passing metrics to graduate the experiment?

5. What are non-goals and potential side effects that are worth accepting?

What do all of these things have in common? These are the ingredients of what actually made a good PRD in the past, and they still do now.

And even beyond the factual information contained in the spec, a spec is a great artifact that allows you to align humans. It’s what you discuss, debate, refer to, and sync on.

That’s why OpenAI actively still regards specs as one of the most important documents. As Sean Grove shared in his viral talk, “The New Code:”

---

2. How the PRD Lost Its Way

In the era before ChatGPT, a long written document was a sign of a human having put in a lot of work. Many humans used that opportunity to show all their hard work by writing very long and thorough documents.

Unfortunately, this is how the PRD lost its way.

Even before ChatGPT, most PMs could relate to this meme:

It only got worse post-ChatGPT. PMs started using LLMs to create overly long PRDs that said nothing (the sad tragedy of most LLM writing).

As a result, how much PRDs were read further dropped off a cliff.

This created the modern situation, where Heads of Product in AI are calling for an end to writing cultures towards build-first cultures.

---

3. The Value of a PRD and a Modern AI PRD Template

But a PRD, when done right, is very different from a prototype.

Let’s work through the first principles of what is useful in a great PRD for the AI era.