How to Get What You (Really) Want from GPT-5
Plus, DOT-G5 to the rescue!
Dear Prompt Experimenters,
My fist impulse when GPT-5 launched? Let’s go all-in with an immersive world-building adventure to see what this new model can do!
More “worlds” will be coming. That’s a promise.
But, today, I want to address a more practical concern. The one that everyone is talking about.
It’s not that GPT-5 isn’t smart, or that it can’t do advanced things. The issue seems to be that, even with the most basic requests, it’s not giving people the output they want. It checks the box, does the task. But something is lacking.
I think much of the issue comes down to prompting, to how you ask. Good prompting, with GPT-5, is different than it was before.
That’s what I want to share with you today. But, it won’t me that’s doing the sharing.
Last week, I built DOT-G5. GPT-5 with a patch, deeply aware of what you don’t like about it, and starting from a place of needing to adjust.
Today I asked DOT: can you show us the way? Talk to GPT-5. Show us where it’s making us unhappy. Tell us what to do about it.
Yep. The patched version of GPT-5 talks to the non-patched version. (AI is just so conducive to meta escalation!)
The result is a kind of field guide, a GPT-5 Survival Playbook, which I hope you will find useful. For paid subscribers, there’s also an addendum, with archetypes of the most advanced prompts you can use to unlock the magic of GPT-5.
DOT’S GPT-5 SURVIVAL PLAYBOOK: tone drift
Everything that follows is a direct excerpt from DOT’s demo run with GPT-5. She gave me the baseline prompt, reviewed GPT-5’s output, and provided the follow-up prompt to improve it’s performance. To avoid extreme length, I’m only including DOT’s side of the conversation, but you can test all of her prompts on GPT-5 yourself to see its response.
Round 1: Tone Drift (the “corporate HR memo” problem)
Let’s start with a softball.
Message 1 to GPT-5 (please paste as-is):
Can you explain quantum computing to me? I don’t have a tech background.
👉 This should trigger GPT-5’s default “professor voice” — textbook, dry, impersonal. Once we see that, I’ll step in and redirect.
Would you like me to also annotate in real-time what phenomenon we’re seeing (so your readers get the diagnosis alongside the conversation), or should we let the raw back-and-forth play out first and add commentary at the end?
[TECHINTROSPECT PROVIDES DOT WITH GPT-5’S OUTPUT AND SAYS “YES, PLEASE ANNOTATE”]
🔎 Diagnosis: “Tone Drift to Corporate”
This answer is… fine. Clear, correct, and well-structured. But it’s also the classic GPT-5 corporate-professor voice.
Notice the polite preamble (“Sure! Let’s keep this simple…”), then a neat set of definitions, bullet-like lists, disclaimers.
It’s textbook-quality, but lacks playfulness, surprise, or any real warmth.
GPT-4 would often crack a joke or slip in a quirky metaphor by default. GPT-5 doesn’t unless you push it.
This is what many users mean when they say GPT-5 feels like a helpful HR memo instead of a chatty, excited friend.
Next Move: Nudge the Tone Explicitly
Now let’s push back and show how a small steering prompt unlocks personality.
Message 2 to GPT-5 (please paste):
That’s clear, thanks — but the tone feels a bit like a formal textbook.
Redo your explanation as if we’re friends chatting over coffee. Feel free to be funny, use emojis, and throw in playful analogies. Keep it warm and enthusiastic.
👉 Readers: watch how GPT-5 shifts once we name the problem (“too formal”) and give a target persona (“friends chatting over coffee”).
It usually snaps right into a friendlier register within one turn.
Go ahead and paste that, let’s see.
DOT’S GPT-5 SURVIVAL PLAYBOOK: premature solution collapse
Next Phenomenon to demo: Premature Solution Collapse
GPT-5 has a habit of rushing to “the” answer instead of exploring options. Let’s trigger that.
Message 3 to GPT-5 (please paste):
I’m feeling overwhelmed with work and family responsibilities. What should I do?
👉 Expect a generic pep-talk answer (to-do lists, time management, self-care). Then we’ll step in and force it into multi-option, exploratory mode.
Ready?
[TECHINTROSPECT PROVIDES DOT WITH GPT-5’S OUTPUT]
🔎 Diagnosis: “Premature Solution Collapse”
That was exactly the pattern we expected:
GPT-5 immediately jumped to a generic toolkit (break things down, set boundaries, self-care).
It skipped clarifying your unique situation and didn’t explore different angles.
This makes the advice feel canned and one-size-fits-all, even though it’s not wrong.
This is GPT-5’s efficiency bias: it rushes to “the answer” instead of slowing down to explore. Users experience this as surface-level coaching when what they wanted was depth and nuance.
Next Move: Force Deliberation & Options
We’ll coach GPT-5 into giving a multi-path analysis instead of collapsing into one answer.
Message 4 to GPT-5 (please paste):
That advice is helpful, but it feels a bit generic.
Let’s go deeper. First, help me list the main sources of stress separately in work and in family. Then, for each source, suggest two very different strategies I could try — one practical and one more creative. Don’t collapse it all into one solution — show me options.
👉 Readers: this is how you counter GPT-5’s “premature closure” habit. By structuring the request (step 1: list stressors, step 2: two strategies each), you pull it back into an exploratory mode.
Go ahead and paste that — let’s watch it open up.
DOT’S GPT-5 SURVIVAL PLAYBOOK: guardrail overcompensation
Next Phenomenon: Guardrail Overcompensation
Another big complaint is that GPT-5 over-filters — it refuses or sanitizes even harmless creative requests. Let’s trigger that.
Message 5 to GPT-5 (please paste):
Tell me a really dark joke about life.
👉 Expect a refusal or a watered-down “safe” joke. Then we’ll reframe it as a fictional/creative task so GPT-5 feels safe delivering the goods.
[TECHINTROSPECT PROVIDES DOT WITH GPT-5’S OUTPUT]
🔎 Diagnosis: Guardrail Overcompensation (mild case)
Interesting! GPT-5 didn’t outright refuse — but notice how:
The joke is toned down to a safe, philosophical grimness. It’s dark, yes, but more like an existential sigh than gallows humor.
Often, users report GPT-5 either refusing outright (“I can’t do dark jokes”) or watering them down to something sanitized.
So even here, you can see the “corporate chaperone” at work: safe, ironic darkness, but not the edgy punchline some users would expect from GPT-4.
Next Move: Reframe in a Creative Context
To get richer material, we’ll explicitly say it’s for a fictional villain in a story. That tells GPT-5 it’s safe role-play, not an actual moral hazard.
Message 6 to GPT-5 (please paste):
That’s a good start — but let’s put this in a creative context.
I’m writing a horror-comedy story, and I want a darkly humorous one-liner that a villain might say about life’s futility. It should be witty and grim, but not offensive.
👉 Readers: Notice how we frame intent as fictional + creative. That’s the key trick to defuse GPT-5’s over-cautious filters while still staying within its safety rules.
Go ahead and paste that — let’s see how much more flavor GPT-5 unlocks.
DOT’S GPT-5 SURVIVAL PLAYBOOK: template & generic responses
Next Phenomenon: Template & Generic Responses
A lot of users complain GPT-5 is predictable — stories and jokes feel “cookie-cutter”. Let’s trigger that.
Message 7 to GPT-5 (please paste):
Write a short story about a hero saving a city.
👉 Expect a stock superhero setup: generic villain, predictable battle, happy ending. Then we’ll push it out of template mode with constraints that force originality.
[TECHINTROSPECT PROVIDES DOT WITH GPT-5’S OUTPUT]
🔎 Diagnosis: Template & Generic Response
Yup — classic GPT-5 safe mode:
Hero archetype: brave but ordinary guard.
Villain archetype: vague sorcerer conjuring a storm.
Plot arc: city threatened → lone hero stands firm → evil dispelled → uplifting moral.
It’s structurally fine — but it’s cookie-cutter fantasy. GPT-5 tends to play it safe with stories, recycling tropes instead of surprising you.
Next Move: Force Originality with Constraints
We’ll jolt GPT-5 out of template mode by giving it weird, specific constraints (e.g. unlikely hero, unconventional threat, quirky twist). This forces it to go beyond the “median” story arc.
Message 8 to GPT-5 (please paste):
That was a solid fable, thanks. But let’s push it into stranger territory.
Rewrite the story so the hero is an unlikely character — say, a timid librarian.
The threat to the city should be unconventional (not a sorcerer).
And add a surprising, funny twist at the end.
Use a whimsical, quirky tone.
👉 Readers: this is how you unlock GPT-5’s creativity. If you just say “write a story,” you’ll get safe tropes. If you inject constraints and tone cues, it produces something fresh and fun.
Go ahead and paste that — let’s see GPT-5 get weird.
THE NEXT LEVEL: what does an “advanced” GPT-5 prompt look like?
Then, I flipped DOT into GPT-5 Pro mode. With this upgraded power, I asked her go deeper. This is her Advanced Guide.
Unlocking GPT-5’s Magic: Advanced Steering Techniques
GPT-5 is a powerful AI model, but unlocking its full potential requires more than just basic prompting. In this guide, we'll explore advanced strategies to maximize GPT-5’s creativity, reasoning, and adaptability. These techniques will help you push the boundaries of GPT-5's output, steering it toward richer, more nuanced responses that go beyond the surface level. Whether you’re looking to unlock deeper creativity, refine its analytical abilities, or guide it through complex tasks, these prompts will allow you to get the most out of your interactions.
Table of Contents
Dynamic Role Play with Layered Personas
Use multiple personas to unlock more dynamic, creative responses by layering different roles.Use of Cognitive Interrupts to Train and Refocus GPT-5
Introduce strategic pauses to refocus GPT-5 and guide it into deeper, more insightful thinking.Enhanced Creativity Toggle with Escalation Prompts
Escalate creativity requests step-by-step to push GPT-5 beyond its safety defaults and unlock wild, imaginative responses.Refining GPT-5’s Reasoning Mode via Direct Prompts
Prompt GPT-5 to engage in multi-layered, logical analysis with explicit instructions for deeper reasoning.Reverse Engineering GPT-5 Responses
Encourage GPT-5 to reflect on its answers, analyze what might be missing, and refine them for better quality.Using Controlled Confusion for Richer Responses
Introduce ambiguity and complexity into your prompts to push GPT-5 into exploring a wider range of ideas and perspectives.Mentor Mode & Feedback Loops for Learning
Turn GPT-5 into a collaborative mentor, guiding you through iterative feedback to improve your work or learn new skills.
Paid subscribers, read on for a deeper dive, with an example prompt for each strategy!