ChatGPT-5 Test Series - Part 1: Conceptual Work in Action
🕓 Read Time: ~5 minutes
When OpenAI rolled out GPT-5, I asked you what I should test first. Thanks for all your input — this is Part 1 of the series, where I dig into how it handles conceptual work.
By “conceptual work,” I mean the kind of big-picture, structural thinking that many of us coaches, consultants, and solopreneurs do every week. Think:
-
Redesigning a curriculum or program
-
Structuring a new offer
-
Reworking core business workflows
I put GPT-5 to the test on two real projects:
-
Updating one of my AI-focused offerings
-
Refreshing one of my core workflows
In this edition, I won't talk about tool specs (I'm too lazy to actually read them in much detail). I always test tools the same way I expect my clients to: in real workflows, with real stakes.
So what changed between GPT-4o and GPT-5?
Many things have changed. The focus here is on my experience with conceptual work. At first, I tried GPT-5 in auto mode, where it decides whether to respond fast, “thinking,” or “long thinking.” To be honest, it didn’t feel like a big leap forward. The outputs looked similar to GPT-4o.
But then I switched to manual model selection. Instead of letting it choose, I selected the middle “thinking” mode.
That’s when things changed.
Suddenly, it felt like working with a colleague at the whiteboard. We bounced ideas back and forth. It structured concepts with me, tightened the flow, and worked at a speed that felt better than 4o. When I say the speed felt "better", that doesn't mean ChatGPT responded faster. It actually responded more slowly. But the quality of the responses was way better, saving me time on iterations.
For me, the co-creation experience was noticeably sharper.
The big surprise: GPT-5 is unforgiving with bad prompts
Here’s the part that really stood out: GPT-5 will not rescue you from a bad prompt.
Where GPT-4o often managed to turn a vague or fuzzy ask into something usable, GPT-5 is less forgiving. If your input is unclear, your output will be too. Good old "garbage in, garbage out".
If I'm perfectly honest, I actually think that’s a good thing:
Why?
GPT-4o would often “fill in the blanks” when a prompt was vague. Sounds helpful but in practice, it often invented details, produced generic content, and gave results that felt flat.GPT-5 forces you to do your part: to be explicit, structured, and intentional. The trade-off? If you put in the work, you get output that’s sharper, more relevant, and actually usable.
And it’s not just me noticing this. In fact, OpenAI has released an official GPT-5 Prompting Guide — which in itself is a clear signal that things are changing. Prompting isn’t an afterthought anymore... It’s a key skill to have.
This is where GPT-5 aligns beautifully with my favorite prompting formula:
-
Frame the role: “You are an instructional designer…”
-
Set the constraints: length, style, tone, audience.
-
Specify the output format: outline, table, list, etc.
-
Iterate interactively: ask it to critique, refine, or re-organize its own answers.
-
Check its work: prompt it to suggest improvements.
With this approach, GPT-5 doesn’t just respond, it collaborates.
It sounds different, more "normal"
One more thing I noticed: the tone shift.
GPT-4o can sometimes feel like a hyperactive co-worker: over-eager, full of emojis, always saying “Yes, absolutely!”.
GPT-5 doesn’t do that. It’s calmer, more neutral, even slightly robotic at times. Of course, I can always instruct it to be more upbeat, supportive or whatever. But that's up to me.
For conceptual work, I actually liked this. It made the collaboration feel more grounded.
Will that tone still work for more emotional or creative writing? That’s a test that deserves its own edition.
And one more thing: Your CustomGPTs may need a review
As many of you know, most of my daily work runs through my virtual team members — the custom GPTs (aka "Assistants") I have built for specific workflows.
With GPT-5, I found that most of them still work just fine but not all. The more structured ones survived the update well. The less detailed ones? They needed tweaking.
So I’ve been reviewing them one by one, and here’s what I’d suggest if you rely on custom GPTs:
-
Re-check the instructions: Make them explicit and step-by-step.
-
Set the default model: Decide whether each GPT should run on “fast” or “thinking.”
-
Test for clarity: Run them on your core tasks and adjust where they stumble.
One surprising discovery: you can actually ask GPT-5 to rewrite your custom GPT instructions for GPT-5 compatibility. And it does a pretty good job.
Key Takeaway
GPT-5 is powerful for big-picture thinking... but only if your prompts are structured, specific, and intentional. Unlike earlier models, it won’t cover for vagueness. That’s a good thing: it pushes you to get clear, and rewards you with sharper, more original results.
Til next time, keep working on those prompting skills!
Elena
P.S. Want to explore OpenAI's ChatGPT 5 Prompting Guide? Here it is.
P.P.S. On a another note. Next week (24 Sep @ 3:15 p.m. CEST), I’ll be joining Václav Sulista, entrepreneur career consultant, live on LinkedIn to talk about how to make AI your career advantage (instead of your fear). If this has been on your mind, you can sign up here and bring your questions.
