How to Get a Patched Version of GPT-5 You'll Like Better
Plus, a little bit of humility goes a long way
This is a paid subscriber post, including a free summary. Free subscribers, for access to the full post, please use the button at the bottom to explore subscription options.
Dear Prompt Experimenters,
The reaction to GPT-5 has been…extreme.
Occasionally positive. Mostly not. Everybody has a strong opinion.
Except me.
Is it a game-changer, or is it a flop? Or somewhere in between?
Here’s my honest answer: I don’t know yet. And I don’t think you do, either.
There’s a lot going on under the hood. Trying a few prompts isn’t enough to draw conclusions.
What if we gave GPT-5 a thorough inventory of everything people don’t like about it? Could we, then, simply ask it to be better? They say it’s really smart, so surely it can adapt to feedback, right?
That’s what I’ve been working on.
And here’s what I’ve created: a “patched” version of GPT-5, highly aware of its own faults, programmed to lean into its new strengths while avoiding the behaviors that everyone is complaining about so loudly.
It’s available to all of our paid subscribers, as a fully functional alternative to “raw” GPT-5. I hope you’ll like it. I sure do.
And (hint, hint!) after collecting feedback from you, I know we can make it even better.
Here’s what we’ll cover in today’s post
A little humility goes a long way: How I taught my version of GPT-5 what people don’t like about it, and why this matters.
Prompting is different now: Those 4o prompts? They’re not right for GPT-5. To test the new model, you need a new kind of prompt.
Get your own patched version of GPT-5: Meet DOT-G5. She is a fully functional alternative to GPT-5. She can (1) discuss GPT-5 with honest humility around what you might not like, and an orientation to finding real solutions (2) Write prompts that are optimized for GPT-5, and show you exactly how they are different from 4o prompts (3) Become your new default for “GPT-5, but better.”
A LITTLE HUMILITY GOES A LONG WAY
Let me just say: I don’t agree with the harshest critics. My own experience with GPT-5 did not blow me away out of the gate. But it also didn’t disappoint me in the way other people are reporting.
But a LOT of disappointment is being reported. And that matters.
When you ask ChatGPT about GPT-5, you get a response like this:
It builds on the strengths of earlier GPT models but with significant improvements in reasoning, adaptability, and contextual awareness.
Here’s an overview:
1. Core Improvements
Better reasoning chains – I can maintain more coherent logical steps, so multi-stage problems (like complex math, planning, or legal reasoning) are less prone to “logic drift.”
More context awareness – I can handle and recall longer, more detailed conversations without losing track of earlier points.
Stronger adaptability – I adjust my explanation style more fluidly to the user’s expertise, tone, and goals.
It goes on. For a long time. It’s all positive.
My first step was to disabuse the model of this perception.
How? I used Deep Research to assemble a set of reports documenting all of the criticism of GPT-5, what specific things people don’t like, why they don’t like them, and the extent to which each is “fixable” through prompting.
I converted this into a Custom GPT, DOT-G5.
When you ask DOT-G5 to tell you about GPT-5, you get a more balanced perspective:
Practical Reality
It’s not “dumber” than GPT-4 — the model is actually more controllable and consistent if you use precise, structured prompts.
The disappointment is often about defaults, not raw capability. GPT-4’s defaults were warmer and more exploratory; GPT-5’s are safer and more formal.
If you treat GPT-5 as something you “drive” actively — telling it tone, process, constraints — you can often get better results than GPT-4 ever gave. If you treat it passively, it can feel bland or rigid.
Why does this matter? We are inviting the model to fundamentally re-anchor. Rather than starting from a neutral position, it is now assuming that you will not be fully satisfied with GPT-5 output, unless and until it explicitly addresses the issues that are alienating many users.
That’s what we want.