Blog Article
Does ChatGPT 6.1 Terra Need a PRD? I Tested It Twice
I reran my complex-build bench on ChatGPT 6.1 Terra with a PRD, then again with a brand PRD. The PRD fixed the skipped work, but the builds were still underwhelming.
Short answer: yes. If you're building with ChatGPT 6.1 Terra, give it a PRD, or it will cut corners.
Longer answer: even with a PRD, I wouldn't pick Terra for a web app with a lot of moving pieces. I ran the test twice, and this is what I found.
Why Terra is my exception to "skip the PRD"
A PRD, or product requirements document, is basically a detailed spec that tells the model exactly what to build.
For most of this year, I've been steering away from PRDs unless it's a client build. Models have become more and more capable of building what I want without one.
Terra is the exception. It's so cheap, cost-efficient, and fast that it skips over important details of a base prompt.
I saw that in my last video, where I tested Terra on every effort level with a basic prompt. Six out of six runs finished in under 5 minutes, and it missed the beat. It built a landing page with a few pages, and that's it:
- No backend
- No go-to-market strategy
- No motion graphics
So I wanted to know: does a PRD fix that?
The setup
I use a bench that runs the same prompt on every effort level side by side: low, medium, high, extra high, max, and ultra.
I prepared a PRD in ChatGPT on extra high and ran two tests:
- PRD run: the same complex build, with a PRD in place of the basic prompt.
- Brand PRD run: the PRD plus our brand, persona, positioning, and color palette, using the Clearmud website as a reference.
One thing to keep in mind: the bench doesn't give the model any skills or tools beyond fetching things from websites and doing its own research.
Test 1: the PRD run
I expected these runs to take a little longer, since there are more instructions to follow. The fastest was medium, at 2 minutes and 28 seconds. Ultra finished very quickly too.
The surprise was low. Of the runs that had finished at that point (Max was still going), low took the longest. I did not see that coming. [02:37]
### It finally built a backend
On low effort, with a PRD, Terra built a backend. A decent looking one, with room for improvement. The basic prompt never gave me that on any effort level.
### AI gibberish text is back
Then I opened medium and high, and there it was: gibberish text with random characters on screen. I have not seen that in ages with an AI model. One of them didn't even tell me how to sign in. [04:09]
### I thought it skipped the work. It hid it.
At first I couldn't find the go-to-market strategy or the creative anywhere. I was expecting another route on the website.
Then on Ultra I found them. The launch plan and the creative library were sitting inside the admin login. I went back through the other effort levels, and they were there too. I had gone through them so quickly that I overlooked them. So they didn't miss the beat. They just put it all in the backend.
The motion graphics were very basic. That makes sense, though. We didn't give it any reference skills, and it had no access to Remotion or HyperFrames. Those would make a drastic difference.
### The mid-test verdict
The PRD wasn't perfect. I could have spent more time on it, 110%. But comparing a basic prompt to a PRD for a complex build, 10 times out of 10, we're going PRD.
Test 2: the brand PRD run
The first PRD run didn't pull our color palette, and that was on me. The PRD had no specific brand details. So I prepped a second PRD with all of that and ran the bench again.
This run was rougher:
- Low: It saved my info in the admin login, which I liked. But the text was too small, the go-to-market strategy wasn't as in-depth as I wanted (a fault in the PRD), and the motion graphics didn't animate at all.
- Medium: Gibberish text again, and it wouldn't let me sign in. So it's broken.
- High: Missed the beat altogether.
- Extra high: A very basic launch plan that doesn't do much.
- Ultra: A nice little mockup up front, but the launch plan and creative library did not load.
It was slower, too. Extra high took 9 minutes on this run, versus 5 minutes on the first one. Low took longer than medium again, though it was a bit faster than high.
Max keeps winning
Across all of this, one effort level stood out.
On the first PRD run, Max was the winner. A great looking dashboard, a more in-depth launch plan, and better motion graphics. It was the first output that showed the creativity and depth I want to see in models. [08:25]
It's not perfect. It still defaults to that shade of green GPT models love, the text needs to be larger, and I wouldn't ship it as is. But it was far better than everything else on the bench.
I went back to the basic prompt run to compare. Max was my favorite there as well. It looked good above the fold and even pulled in the Clearmud logo. The problem? No backend. It built the front end and ignored all the other work I asked for.
On the brand PRD run, Max was my favorite again: great dashboard, great looking leads view. But the launch plan was thinner than I wanted, and the motion graphics wouldn't load, even after a refresh. Not a one shot.
So we're seeing a trend. If you're going to build with Terra, Max is the go-to.
So, should you build with Terra?
I'm going to be honest with you. Even with a PRD, it's a bit underwhelming.
The PRD made Terra follow instructions far better. We got the backend, the launch plan, and the creative library that the basic prompt skipped. But we also got gibberish text, broken sign-ins, and pages that didn't load.
My takeaways:
- Building with Terra? Use a PRD. Otherwise it will cut corners.
- Use Max. It was the best result in every run.
- Building a web app with a lot of moving pieces and depth? Don't choose Terra.
- Use Terra where it's strong. As an agent's brain: quick actions and repeatable tasks that follow a well-defined skill file. It also does a great job with video editing.
For websites and web apps with backends, go-to-market strategies, and motion graphics, I'm simply not going to use Terra.
Now, I'm not an AI expert. I'm building in public and sharing what works. If you want a test like this for your own use case, drop a comment on the video with who you are, what you do, who your target audience is, and what your question is. I'll add a video to my queue tailored to you.
Watch the full test: https://www.youtube.com/watch?v=xBU1rJpI21c
Watch the full walkthrough on YouTube.
Watch on YouTube