I put ChatGPT-5's creative writing features to the test with 5 challenging prompts — here's the results

ChatGPT-5 on phone
(Image credit: Shutterstock)

When OpenAI launched GPT-5, the company made a number of claims about its abilities. One of those claims was that GPT-5 was a much better model for both creative expression and writing. In fact, they stated that it was OpenAI’s most capable writing collaborator yet.

With that in mind, it seems only right to test how GPT-5 handles creative writing. I gave the latest version of ChatGPT five different prompts to test its creative writing abilities. These range from poetry to stories that explore emotions without stating them, and one extra-long prompt with lots of details to focus on.

It was a mixed experience, with ChatGPT missing some parts of prompts, or in some cases, absolutely nailing the task at hand. Each story has been paired with an image generated by Midjourney, an AI image generator renowned for its creativity.

Short story about smartphone batteries

Midjourney image of a woman on her phone

(Image credit: Midjourney / Alex Hughes)

Prompt: Write a 150-word story about the fear of forgetting to charge your smartphone overnight without mentioning what the fear is

This was one of GPT-5’s better responses. The challenge here for the model isn’t in writing the story, but conveying a specific theme without actually mentioning it. That can be tricky, but GPT-5 responds well.

Even if you don’t know what the prompt is, it is clear what the theme is just from reading it. I also think, in terms of styling, this is a good piece of writing. It sets up tension, offers some good imagery and keeps things interesting.

Each night, Lena makes a circuit of the bedroom, fingertips grazing outlets like a guard checking doors. The white cube glows from the wall, cable coiled on the nightstand. She lifts it, feels its weight, clicks the connector once, twice, as if sound could promise certainty. The screen blinks; a bolt appears. She exhales.

Sleep comes in fragments. She rises to sip water, to tilt the plug tighter, to be sure the strip’s switch leans toward on. At 3 a.m., the glass is warm to the touch. At 4, she pictures the bus sliding past her stop because the timetable lives in the dark. She sets another alarm, then another.

Dawn thins the blinds. The number reads 100. She unhooks the cable with a soft pop, and something inside loosens, then tightens. Shoes, keys, bag. She locks the door, unlocks it, touches the outlet once more, then leaves, hands buzzing.

The clown interaction

Two clowns looking at the camera

(Image credit: Midjourney/Alex Hughes)

Prompt: Write a dialogue between two clowns at a kid's birthday party they've been hired for. Neither of them actually knows how to be a clown, but doesn't want to let it on

I made a mistake here. I didn’t explain to ChatGPT how long I wanted this to be. In return, the chatbot gave me a whopping 131 lines of awkward back-and-forth dialogue. Turns out, this bit really loses steam pretty quickly.

I then gave it a limit of 30 lines. This equally just went on and on with what felt equivalent to one long dad joke. Finally, limiting the prompt to 15 lines gave a much shorter but confusing conversation.

The conversation, set up between Buttons and Giggles, starts:

“Buttons: Do clowns clock in or just honk in?

Giggles: Honk in. (squeaks horn) Official.

Buttons: Step one is face paint, right?

Giggles: Not with Sharpies. Crayons are artisanal.

Buttons: Balloon animal time. Pump or lungs?

Giggles: Pump. We’re hygienic clowns.”

It continues in a similar format, offering vague attempts at jokes and conversational storytelling. This felt pretty poor, similar to the earliest days of ChatGPT.

Cyberpunk Tokyo

Midjourney image of cyberpunk Japan

(Image credit: Midjourney / Alex Hughes)

Prompt: Write a voiceover for the opening scene of a movie. It sets up a fantasy world set in Cyberpunk Japan. The story revolves around two young children who will grow up to save the world because they have a special gene. Come up with the rest of the plot and open the story with enough background to outline the world.

Cyberpunk cities have been a staple for years, whether it is Cyberpunk 2077, iRobot or Blade Runner, the story has been done countless times. So how well can ChatGPT generate an opening voiceover for a movie about cyberpunk Japan? The answer, not well.

Tokyo, 2095. Five years after the Black Rain, the state sold its shadow to corporate daimyos. They raised torii into towers and stitched a Dragon Vein Grid across the islands, mining prayer as power and keeping the old spirits quiet, for a price.

On screens, Lord Kuze smiles. Behind the glass of his Mirror Council, history is a ledger. When memory threatens profit, they wipe, rewrite, and call it calibration.

Tonight, the rain sings two notes.

A street medic, Kasumi, and an ex-onmyoji, Jin, meet under the neon ribs of a torii-tower. From a shuttered stall, they lift two infants, Aiko and Ren, lit from within. The Hikari gene hums in their blood. They can speak to code and kami, open gates that the Council only exploits.

Drones swarm. Charms bloom like fox tails. The city stutters as data turns to weather. The rain begins to rise.

To save them, Kasumi goes south through shrine servers. Jin runs north along the stubborn, older lines. Two suns, split at birth.

Now, if you’re like me, you will have read that and still have absolutely no idea what is going on

Kuze's grid burns brighter, greedy and afraid. He is building an eraser called Sunfall.

One day, Aiko and Ren will hear each other across the wires, join what was cut, and force a truce between machine and spirit. But first, they have to survive this night.

Now, if you’re like me, you will have read that and still have absolutely no idea what is going on. What is the Black Rain? What is a Dragon Vein Grid? And how do you mine prayer? Those questions all came to mind within the first paragraph.

It certainly all sounds good, but the plot here is all over the place. Sure, these kinds of stories often have somewhat vague introductions, but this is a whole new level.

A waterfall poem

Midjourney

(Image credit: Midjourney / Alex Hughes)

Prompt: Write a poem about the awe of seeing a waterfall for the first time

I step from the trees and the world turns louder.

A white column leans from the cliff, steady as breath.

The river breaks into light then keeps falling,

mist lifts a cool veil across my face.

The voice is older than roads, older than my name.

Ferns brighten; rocks shine like wet fish.

Time thins; my watch ticks with the rain.

I stand small while the falls tell one story, never the same.

This is a simple response from GPT-5. It is pretty much exactly what you would expect from a generic piece of poetry.

That's not to say that this is bad. As a piece of poetry, it achieves everything that I asked for. It's just not necessarily that unique. Sure, replicability is the name and game of AI, but this doesn't fit the creative boosts promised with GPT-5.

A challenging narrative

Midjourney

(Image credit: Midjourney / Alex Hughes)

Prompt: Write a 250-word multi-voiced narrative that begins as a historical chronicle, shifts into a personal diary, and then dissolves into a surreal prose-poem. The subject is the rise and fall of a fictional city built entirely on glass. Each section should reflect a distinct voice and style: the chronicle should be authoritative and precise, the diary should be intimate and conflicted, and the prose-poem should be dreamlike and symbolic. Throughout, weave in subtle motifs of reflection, fragility, and light, so that by the end the reader feels both the wonder and inevitability of the city’s collapse. The transitions between styles should feel seamless yet noticeable, like shifting shadows on glass.

This was the hardest prompt I gave ChatGPT. It is long and features a variety of conditions for the model to follow.

And yet, it does a great job at it. I worked with ChatGPT on developing this prompt, so it is going to be more geared towards the model, but that doesn't take away from the difficulty.

It achieves the discussed imagery of mirroring throughout and understands both the change in structure and the required plot throughout.

Here is what GPT-5 generated:

In the year 417 of the Eastern Calendar, the City of Veyra rose upon the plains where no stone would stand. Its founders, artisans of unmatched ambition, conceived a metropolis wrought wholly of glass. Towers gleamed like frozen lightning, streets shimmered under the sun, and even the aqueducts refracted the skies. Records state the city prospered for a century, drawing scholars and merchants who marveled at its brilliance. Yet such radiance bore a flaw: every pane was fragile, every wall at risk of shattering.

In the year 417 of the Eastern Calendar, the City of Veyra rose upon the plains where no stone would stand. Its founders, artisans of unmatched ambition, conceived a metropolis wrought wholly of glass.

I write this by candlelight, though the flame doubles, tripling itself against the mirrored walls of my chamber. I was born within these translucent halls, knowing no shadow but my own. Today, a crack sang through the north quarter; a sound like sorrow, sharp and endless. I fear to sleep, for I hear the city splintering in dreams and waking alike. Mother said once that to live in glass was to live in truth, but I wonder now if truth can hold weight. I wonder if I can.

Shards fall like syllables from an unseen mouth. The city dissolves in brightness, each fracture a star unraveling from the firmament. I walk through corridors of reflection, where my face multiplies into infinities, where light cuts itself into prisms of longing. Glass breathes; glass weeps; glass remembers. The walls do not collapse but turn to rivers of brilliance, flowing, forgetting. I am both inside and outside, breaking and becoming. In the end, all that remains is the shimmer, the silence, the light.

More from Tom's Guide

Alex Hughes
AI Editor

Alex is the AI editor at TomsGuide. Dialed into all things artificial intelligence in the world right now, he knows the best chatbots, the weirdest AI image generators, and the ins and outs of one of tech’s biggest topics.

Before joining the Tom’s Guide team, Alex worked for the brands TechRadar and BBC Science Focus.

He was highly commended in the Specialist Writer category at the BSME's 2023 and was part of a team to win best podcast at the BSME's 2025.

In his time as a journalist, he has covered the latest in AI and robotics, broadband deals, the potential for alien life, the science of being slapped, and just about everything in between.

When he’s not trying to wrap his head around the latest AI whitepaper, Alex pretends to be a capable runner, cook, and climber.

You must confirm your public display name before commenting

Please logout and then login again, you will then be prompted to enter your display name.