I have no idea how to prompt GPT-5

Thank god OpenAI gave us some guidance

It’s been two weeks since the chaotic launch of ChatGPT-5 and people are still confused.

Is 5-Thinking the new o3? Should 5-Auto be your new daily driver?

One thing is clear. If OpenAI’s goal was simplification, they just created a new model soup by bringing back all their legacy models (including o3).

It turns out I’m not alone here. My AI gurus over at

Confluence: AI, Leadership, and Communication

recently wrote:

And after a week, we’re not quite sure how good it really is.

Their experience ranged from “I do not like GPT-5 at all” to “GPT-5 Pro is very smart.” They struggled to gauge the value-add of the new auto-router.

This GPT-5 confusion persisted across my various AI communities and consulting clients.

But OpenAI quietly dropped a little utility that immediately made things a lot better.

---

Stop leaving 10 hours of productivity on the table every week. ChatGPT Mastery for Professionals transforms knowledge workers, like you, into AI power users in under 3 months. While your colleagues are still treating ChatGPT like Google, you'll be automating reports, building custom tools and actually getting home on time.

---

Meet the GPT-5 Prompt Optimizer

Buried deep in OpenAI’s resource library you’ll find a prompt optimization tool.

The promise is simple: drop in an old prompt and the tool will make it better.

I started with a quick and dirty prompt about early retirement:

Build me a simple web calculator that determines if I can retire early. It should have < 10 variables such as:
- Withdrawal rate
- investment return
- spending
-age
- inflation

The goal is simple testing, not detailed analysis

Here’s the optimization result (with the recommended changes in blue):

I encourage you to click on the little text bubbles on the right. They provide clues into how GPT-5 needs to be prompted.

Here are a few of the recommendations:

• Adding headers: Introduced clear and structured sections (role/objective, instructions, etc.) to improve prompt clarity and support better parsing and task comprehension by language models.

• Adding the checklist: Introduced a pre-coding checklist requirement (as per GPT-5 best practices), compelling thorough planning and improved workflow transparency for complex, multi-step tasks.

The result was a pretty nifty calculator that can now be run within the browser:

Is GPT-5 better at research?

Next, I wanted to see how the prompt could be modified for a quick research task. (Not Deep Research, I wanted a “quick explainer.”)

As an ex-hedge fund investor, I’ve had this nagging question about the growth of the private credit market — a phenomenon that seems to have taken off after my semi-retirement.

Here’s the old prompt:

I want to understand the growth of the private credit market. Particularly the explosive growth over the past decade. Who were the key players? What are the key sectors? What caused the growth? How has the product evolved? What does the product consist of? Provide a 2000 word bulleted report that uses high-quality primary sources and is targeted at someone who understands the alternative investment industry very well. 

And here are some ideas (in pink) called out by the prompt optimizer.

Laying out the structure of the report:

A more detailed context section that shows both in scope and out of scope:

I was keen to better understand how I could guide the reasoning process, but this looked more like citation verification:

The results were good, here’s one section of the output:

While I didn’t do an apples-to-apples comparison, I did take the lazy route and asked an independent model (Claude Opus 4.1 Extended thinking) to judge the two. Here’s what it said:

Both documents cover private credit's $2 trillion market evolution, but while the first offers a solid executive-level primer, the second provides institutional-grade research with 50+ clickable sources and granular data points . Think Wikipedia entry versus investment committee deck—both useful, but for different audiences and purposes.

(I recognize that this is extremely lazy, but gets to the heart of the issue — does prompt optimization matter?)

How to prompt going forward?

Like most things in AI, when it comes to GPT-5 the ground is still wobbly.

Which means that research and experimentation should be rewarded.

Unfortunately, I don’t have any definitive conclusions (other than use the prompt optimizer) but here’s what I want to explore:

What are the key sections of the prompt?

So far I see:

• Role and Objective

• Instructions (with an emphasis on 3-7 steps)

• Context

• Reasoning

• Structure

• Output Format

• Planning and Post-Action Validation

What goes under reasoning and planning?

These are the most unfamiliar headers generated by the optimizer. Since these advanced models are all reasoning models it makes sense to better understand the underlying prompting principles.

Asking me to lay out the plan in the prompt makes sense — yet often I don’t know enough about the question to be able to propose a plan.

More to come here.

This looks a lot like pseudo-code

One hypothesis is that GPT-5 wants you to be less conversational and more “code-like” in your problem solving approach. If that’s indeed the case, the transition is definitely going to be more complicated.

Am I a broken record?

For the second time in two weeks, I’ll end with a familiar maxim.

If a model isn’t living up to your expectations, try taking responsibility to figure out if it’s a failure of your own imagination.

---

Stop leaving 10 hours of productivity on the table every week. ChatGPT Mastery for Professionals transforms knowledge workers, like you, into AI power users in under 3 months. While your colleagues are still treating ChatGPT like Google, you'll be automating reports, building custom tools and actually getting home on time.

---