Cooking while GPT-5 cooks: testing a one shot, unsupervised prompt

Like many of you who live deep in the world of software engineering, I’ve been following the buzz around GPT-5—especially its ability to stay on task.

That’s what I care about. If I can trust a model to get on with a job while I’m in a standup or grabbing lunch, we’re getting closer to real delegation.

So I put it to the test. I gave GPT-5 a minimal one-shot prompt: upgrade an old Google Scholar scraper, build the project out, and deploy it. Then I walked away.

What stood out

• Focused follow-through: It parsed the blog post, understood the APIs, and built the full tool—README, tests, Docker, Cloud Build and all.
• Self-correcting: When it hit errors, it often fixed them—though not always without a nudge.
• Context retention: It held the task in mind better than Claude Sonnet 4, even across multiple attempts.
• Still needs supervision: It got stuck in an infinite loop and missed that a key API was deprecated.

Bottom line? GPT-5 is the first model I’ve tested that feels like it’s approaching useful autonomy for real engineering work—but it’s not fully hands-off yet.

Want to see what happened?

I walk through the full run in the video—errors, fixes, decisions, and all. If you’re wondering what GPT-5 is really capable of in a dev workflow, it’s worth a watch.

Watch the demo:

Thoughts? Let’s talk: Tommy.Hinrichs@equalexperts.com

Disclaimer: This demo is for educational and exploratory purposes only. Equal Experts is not affiliated with OpenAI or any AI tool provider mentioned. All views expressed are those of Tommy Hinrichs.