AI Prompts for Poetry Writing Without the Greeting-Card Mush

You are helping me write poetry. Follow these rules for everything you produce:
- NO abstractions (love, time, soul, journey, heart). Use concrete images only.
- NO forced end-rhyme. Rhyme only if it lands naturally.
- Prefer the specific over the beautiful. A cracked coffee mug beats "shattered dreams."
Now: give me 5 concrete images for the feeling of missing someone who's still alive.

The first poem I ever asked an AI to write was about grief. What came back rhymed "heart" with "apart" and contained the phrase "the tapestry of time." It was, technically, a poem. It was also the exact thing you'd find printed under a stock photo of a sunset. A rhyming fridge magnet.

That failure taught me the core problem with ai prompts for poetry writing: left alone, models sprint toward the most average version of a poem in existence, because that's what "poem" statistically looks like. Abstraction, easy rhyme, tidy sentiment. Getting real poetry out requires fighting that default at every step. Here's how.

Quick-Start (Copy This Right Now)

Paste this before you ask for anything:

You are helping me write poetry. Follow these rules for everything you produce: - NO abstractions (love, time, soul, journey, heart). Use concrete images only. - NO forced end-rhyme. Rhyme only if it lands naturally. - Prefer the specific over the beautiful. A cracked coffee mug beats "shattered dreams." Now: give me 5 concrete images for the feeling of missing someone who's still alive.

What this does: It bans the model's three worst habits up front — abstraction, forced rhyme, and prettiness — then asks for raw material (images) rather than a finished poem. Images are where poems start.

⚡ Pro tip: Ask for images, not poems. A model handed "write a poem about X" produces a fridge magnet. A model handed "give me 8 concrete, surprising images for X" gives you ingredients you can actually cook with. Assemble the poem yourself.

Understanding the Variables

Concreteness. This is the master dial. The more you push toward physical, specific, nameable things — the smell of a specific room, a particular object — the less mush you get. Abstraction is where poems go to die, and it's exactly where models drift.

Form. "Write a poem" gives you free-verse sludge. "Write a 14-line sonnet with a volta at line 9" gives the model rails to run on. Constraint sharpens output — a fixed form forces choices, and choices are where craft shows.

Register. Tell it whose mouth this is in. "Plain, flat, unsentimental, like a person too tired to perform their grief" produces something honest. "Poetic" produces the fridge magnet.

⚡ Pro tip: If output feels generic, add a "forbidden words" list. Ban "soul," "heart," "dance," "whisper," "eternal." Watching the model route around its own crutches often produces its best lines — necessity breeds invention, even for a machine.

Step-by-Step: Writing a Poem With AI as a Sparring Partner

Step one — gather images (use the Quick-Start prompt above).

Step two — find the tension:

From these 5 images, which two create the most friction when placed next to each other? Explain the friction in one sentence. Don't write a poem yet.

What this does: Poems run on tension between images. This finds the electric pair before you commit to a structure — the juxtaposition is often the whole poem in miniature.

Step three — draft yourself, then get line-level pushback:

Here's my draft. For each line, mark it KEEP, CUT, or SHARPEN. For SHARPEN lines, tell me what's vague — don't rewrite them. [PASTE POEM]

What this does: It gives you a surgeon's map of your own draft without letting the model paint over your lines. You do every rewrite; it just points at the wounds.

⚡ Pro tip: Never let the model rewrite a line of your poetry. Diagnosis, yes. Prescription, no. The moment it rewrites, your poem picks up its accent — and its accent is the average of all poems, which is the one thing you're trying to escape.

Pro-Level Variations

For a songwriter testing lyrics: "Read this like a listener hearing it once, live. Which image survives a single listen and which needs the page?" Lyrics and page-poems fail differently.

For a poet working in form: "I'm writing a villanelle. My two refrains are [X] and [Y]. Do they have enough meaning-range to survive six repetitions each, or will they go stale?" Form has technical failure modes AI can flag.

For a teacher or a hobbyist learning craft: "Take my free-verse poem and show me where line breaks are doing work and where they're arbitrary." Line-break logic is one of the genuinely teachable things a model can explain well.

Troubleshooting Common Issues

"It only writes in rhyme." Explicitly ban end-rhyme, or specify free verse. Models default to rhyme because it's the most recognizable "poem" signal.

"Everything is abstract and vague." Escalate the concreteness rule and add forbidden words. Demand objects you could photograph.

"The feedback rewrites my voice." Restrict it to diagnosis: KEEP / CUT / SHARPEN, never rewrite.

⚠️ Common mistake: Judging a poem by whether it "sounds poetic." That instinct is exactly what leads to the fridge magnet — "poetic" sound is the cliché signal. Real poems often sound plain and strange, not pretty. When you ask AI to make something "more poetic," you're asking it to add mush. Ask instead for "more specific" or "more surprising," and you'll climb the opposite direction — toward the real thing.

Prompts for Revising a Poem You've Outgrown

Sometimes you have an old poem that almost works. It's easy to over-tinker and kill it. These prompts help you revise with restraint.

Find the load-bearing line:

In this poem, which single line is doing the most emotional work — the one that, if I cut it, would collapse the whole thing? Explain in one sentence.

What this does: It identifies your poem's keystone so you protect it while cutting elsewhere. Writers often revise away the one line that mattered.

Test the ending's landing:

Read only the last line. What does it want me to feel? Now read the whole poem. Does the last line earn that feeling, or does it reach for one the poem didn't set up?

What this does: Poems fail at the exit more than anywhere else. This checks whether your final line is paying off a promise or grabbing for unearned weight.

⚡ Pro tip: When revising, cut before you add. Ask the model, "Which 20% of this poem could disappear with no loss?" Most weak poems are strong poems with too many words around them. The fridge-magnet effect often comes from over-explaining an image that was already clear.

Working With Sound, Not Just Sense

Poems are heard even when read silently, and AI can help you audit the music — carefully. Ask for observation, never correction:

Read this poem aloud in your analysis. Mark any place where the SOUND stumbles — awkward consonant clusters, a rhythm that trips. Describe the stumble; do not fix it.

What this does: It surfaces the moments where sound fights meaning, so you can decide whether the friction is a flaw or the point. You keep every choice.

⚡ Pro tip: Read your poem aloud yourself before and after any AI note. Your ear catches things no model will, and it protects you from "smoothing" a deliberate roughness that was doing real work.

The Uncomfortable Truth About AI and Poetry

Poetry is the form where AI helps least with the actual writing and most with everything around it — reading, diagnosing, pushing you toward specificity. That's not a limitation to apologize for; it's the correct division of labor. A poem is one person saying a true thing in a way only they could say it. A machine trained on the average of all poems is structurally incapable of that. Use it to clear away the mush so your true thing has room to stand — and write the true thing yourself.

Your Turn

Take a feeling you can't quite name and run the Quick-Start image prompt. Pick the two images with the most friction. Then write eight lines yourself, and only then ask for KEEP/CUT/SHARPEN. That loop — machine gathers and diagnoses, you write and decide — is the whole method.

The forbidden-words lists, the concreteness rules, the diagnosis-only framing: these are the prompts you'll want every time you sit down to write, and rebuilding them from memory is how good habits die. Save them in PromptABCD as a poetry set, and your anti-mush toolkit is one click away — so the next poem starts as a poem, not a fridge magnet.