When I first started playing with AI for creative writing, I didn’t want it to write for me. I just wanted it to push me out of writer’s block. So I needed something that would toss me weird opening lines, metaphors I wouldn’t come up with, and jolt my brain awake. What I didn’t expect was how wildly different the output is depending on how you prompt it — even if you just change the order of words.
Thank you for reading this post, don't forget to subscribe!
For example, writing “Write a sad poem about an abandoned train station” gives fairly standard AI output — like a high school assignment. But twist it a little: “You’re a rusty coin wedged in a bench at an abandoned station. Speak your loneliness” — and suddenly it’s deliciously bizarre. The AI gives something haunted, specific, and way harder to predict. That one-start detail — the rusty coin — reframes its style and tone immediately.
In short: You need to decide if you want the AI to entertain, to break your habits, or to ghostwrite. The vibe changes how you prompt.
Break Prompts Into Hot Specifics
Too often the default prompt people use is something like “Write a story about a dragon and a village.” This lands you in generic land. The trick is to break the idea apart into high-impact chunks: sensory challenges, conflicting goals, and charged objects.
Sensory framing: Ask it to describe smells before worldbuilding. Like “Write the scent of a dragon’s lair before we see the dragon.”
Conflicting goals: Instead of “Make a knight fight a dragon,” try “Write a scene where the knight has fallen in love with the dragon, but her village expects her to kill it.” This gives AI emotion to grind against.
Charged object: I gave GPT-4 the prompt: “Describe a poem found carved into a dragon’s tooth” — that alone built a whole culture in the reply without me asking.
This works well with both fiction and poetry. The same principle applies — shape constraints around the emotions or symbols you want explored. The AI will fill in gaps if you don’t, and usually not in a way you’ll like.
In short: Don’t ask AI for the story. Ask it for the moment the whole story turns.
Use AI as a Weird Friend in the Room
I’m not saying the AI’s stuff is always usable. But sometimes it gives me the reaction I want even when it’s bad — because it makes me want to fix it. Here’s the weird little workflow I go through:
I prompt it with a situation like “A chef makes meals for ghosts who can only taste memories.”
The AI gives me a too-clean story. But one line — “The scent of burnt rosemary summoned her father’s silence” — jumps out.
I then prompt: “Write 4 different opening paragraphs that lead to that sentence naturally.” Now the thing builds itself backward and becomes mine.
This method makes AI feel less like a writer, and more like a strange insider trading partner: it leaks pieces I wasn’t supposed to see yet, and I pick up the thread.
In short: Let AI be valuable even when it’s wrong — the point isn’t output, it’s the emotional breadcrumb.
Tweak Tone with Form Pointers That Work
One of the few consistent things I’ve seen across GPT, Claude, and Gemini is this: Form-based prompting (as in “Write in the form of”) controls tone better than theme-based prompting.
Theme Prompt
Form Prompt
“Write a poem about heartbreak”
“Write a heartbreak poem in the form of a Yelp review”
“Write a story about revenge”
“Write a monologue of someone leaving angry voice memos after a betrayal”
“Write a hopeful poem”
“Write a hopeful poem as a message-in-a-bottle found decades later”
Why does this work better? Because you’re not just telling it what to write — you’re telling it how and who is speaking. That tone overlay changes the language more radically than asking it to simply sound inspiring or sad.
In short: If you want something that feels original, pick a strange voice or container. Form shapes feeling more than topic.
Add Creative Friction with Prompt Chains
One of the best ways to get surprising fiction or poetry isn’t with one clever prompt — it’s with a chain of mismatched ones. Basically, you pile emotions or imagery into a sequence the model has to juggle. Here’s the chain I used when stuck writing a poetic sci-fi vignette:
First prompt: “Describe a siren song written by machines.”
Then: “This song has been corrupted. What are three lines in a damaged version?”
Then: “What kind of person would sing this broken song, and why?”
Finally: “Write a paragraph of their internal monologue while singing it to someone they love who can no longer hear them.”
That last output — I couldn’t have written it myself from scratch. But because I walked it through that friction chain, the AI took on a weird poetry logic. This is efficient if you’ve already written a scene and need inspiration to give it more texture.
In short: Build prompts in layers. Isolated requests give you flat writing. Chains create collisions.
Try Prompt Mining Instead of Prompt Writing
What worked even better for me than brainstorming creative prompts? Mining published literature voice-by-voice. I paste a given author’s paragraph into the prompt box and ask:
“Give me a writing prompt that would result in something like this tone but in a totally different situation.”
When I submitted a short slice from Cormac McCarthy, GPT gave me something like:
“Describe a man walking across a flattened amusement park after a storm, carrying a set of brass keys he can’t name.”
I immediately loved that.
Using author work as inspiration for prompt generation does two things:
It mocks the tone instead of the structure (so you don’t clone plots).
It forces the AI to stretch ideas laterally instead of linearly.
In short: Stop trying to invent the prompt — just reverse engineer the vibe you want using your favorite writers’ ghosts.
Inject AI Only Into the Dry Spots
If you’re writing something and hit a part where the scene drags — that’s usually NOT where you want AI to fix it wholesale. Instead, highlight that line or section and prompt:
“Reword this sentence to sound emotionally distant, like it’s being recalled years later during a police interview.”
Or:
“Replace this paragraph with one where the character hides the actual memory and talks around it indirectly.”
This lets AI generate slant, not content. You keep control of the truth underneath, but it adds framing guilt, or coldness, or instability. I even sometimes throw text into Claude with a very short prompt like:
“Make this section feel like it was recorded on decaying cassette tape.”
It’s subtle but it reframes rhythm and vocabulary without changing facts.
In short: Use creative prompts like filters, not writers. AI isn’t better at writing, it’s better at changing temperature.
Don’t Edit the Output — Prompt Again Smarter
Last important thing I learned: Don’t spend an hour cleaning up weak responses. If you have to reword or delete half the poem, your prompt failed. I used to waste time trying to fix bland results from vague ideas. Now I just reprompt smarter.
For example, this input:
“Write a poem about insomnia”
…gave me basic metaphors: clocks ticking, empty streets. Replacing it with:
“Write a poem from the view of a pillow that never gets slept on”
…totally changed it. Suddenly the writing had longing, confusion, even jealousy. And I barely edited a word.
So if your AI writing feels boring — the answer isn’t to fix it. Throw it out and prompt better. Spend ten more seconds writing the input to save half an hour wading through junk.
In short: If the result is flat, your prompt isn’t sharp enough yet. Loop until it sings back to you.
