How to Write a Haiku With AI (Beyond the 5-7-5 Trap)
May 31, 2026
Ask any AI for a haiku and you'll get something that counts to 5-7-5 and says nothing. It will be technically correct and completely dead. That's because the syllable rule is the least important part of a haiku — and it's the only part most people (and most AI tools) know.
This guide teaches the parts that actually matter: the seasonal image (kigo), the cut (kireji), and the turn. Then it shows you exactly how to prompt for a real one — with a full before/after I wrote out so you can see the difference, not just read about it.
What a Haiku Actually Is (Beyond 5-7-5)
A haiku is not "a three-line poem with 5, 7, and 5 syllables." That definition is a school-handout simplification, and it's the reason most haiku — human or AI — are bad. A real haiku is a tiny machine for producing a moment of recognition: it places two concrete images next to each other and lets the reader feel the spark jump the gap.
Three traditional elements do the real work.
Kigo — the seasonal reference
A kigo is a word that anchors the poem to a season. "Cherry blossom," "first frost," and "long shadow" all carry a season inside them. This isn't decoration. The kigo grounds an abstract feeling in the physical, shared world — it gives the reader a temperature, a quality of light, a mood, without the poet naming the emotion. "Cicada" already means late summer, heat, and a sound that won't stop. You get all of that for one word.
Kireji — the cutting word
In Japanese, a kireji (cutting word) is a particle that creates a sharp pause inside the poem. English has no exact equivalent, so we use punctuation — usually a dash, sometimes a colon, a period, or a hard line break. The cut splits the haiku into two parts. It's the hinge. Everything hangs on where you place it.
The turn — juxtaposition
The cut sets up a turn: two images placed side by side with a pause between them, so the reader's mind leaps across the gap and makes the connection itself. This is the entire art. The poet does not explain the relationship between the two images — the poet trusts the reader to feel it. A haiku that explains its own meaning has already failed.
So the working definition to carry into your prompt is: two concrete images, a seasonal anchor, and a cut between them that the reader completes. Syllables are the container, not the content.
Why AI Defaults to Bad Haiku
AI defaults to bad haiku because it optimizes for the rule it can see (5-7-5) and ignores the rules it can't (kigo, cut, turn). Left to its own devices, a model produces a poem that's the syllabic equivalent of a greeting card: it counts correctly, names its theme three ways, and states its emotion out loud.
Here's what a model gives you for "write a haiku about autumn" with no further guidance:
Leaves fall gently down,
Autumn whispers soft and cold,
Nature says goodbye.
Count the syllables — they're right. Now count the problems. "Whispers soft," "says goodbye," and "gently" all tell you how to feel instead of showing you something and trusting you. There's no cut, no juxtaposition, no second image to leap to. "Nature says goodbye" is a thesis statement, not an image. It's three lines that all point the same direction and explain themselves on the way. There's nothing for the reader to do.
The model isn't bad at language — it's doing exactly what a vague prompt asks. "Write a haiku about autumn" requests a summary of autumn in 17 syllables, and that's what it delivers. To get a real haiku, prompt for the machinery, not the topic. (This is the same problem behind most flat AI writing — see using AI for creative writing without losing your voice for the general version.)
Prompting for a Real Haiku: Constraint + Seasonal Image + a Turn
The fix is to specify the three real elements in the prompt and explicitly forbid the failure modes. Don't ask for a haiku about a topic. Ask for two specific images, a season, and a cut — and ban abstraction and emotion-naming. Here's a copy-paste prompt that works across ChatGPT, Claude, Gemini, or SentX:
Write a haiku in English. Follow these rules exactly:
1. Use two concrete, physical images — things I could photograph.
Do NOT name any emotion (no "lonely," "peaceful," "joy," "sad").
2. Include a seasonal reference (kigo): a word that tells me the time
of year through a physical detail, not by naming the season.
3. Place the two images on either side of a cut — use a dash or a hard
line break as the turning point. The reader should feel the
connection without me explaining it.
4. Do not state a conclusion or a lesson. End on the image.
5. Keep it short and close to 5-7-5, but never pad a line to hit the
count. A real image beats a correct syllable.
Topic seed: [a winter morning after an argument]
Give me three versions, each using a DIFFERENT pair of images.
The key moves: you've replaced "topic" with constraints, you've banned the named emotion (the single biggest tell of AI haiku), and you've asked for the cut explicitly. The "three versions, different images" line is doing quiet work too — it forces the model off its first, most generic association and into specificity.
If the output still feels explained, add one follow-up:
Version 2 is closest. The last line still tells me what to feel.
Replace it with a single physical detail from the same scene and let
the cut do the work. Don't explain the connection.
For a deeper library of constraint-based prompts beyond haiku, see creative writing prompts for AI.
A Before/After I Wrote Myself
Here's the same poem at two stages so you can see what "the turn" actually changes. The topic seed was a winter morning after an argument.
Before (the default the model wants to give you)
Cold and quiet house,
We said hurtful words last night,
Now my heart feels numb.
This is the 5-7-5 trap in its purest form. Every line names the feeling: "cold," "hurtful," "numb." Line two narrates the backstory instead of showing an image. There's no seasonal anchor doing real work, no cut, and the poem hands you its conclusion ("numb") so there's nothing left to discover. It's a status update with line breaks.
After (with kigo, a cut, and a turn)
Frost on the window —
two coffee cups, one still full
going cold alone
Look at what changed. "Frost on the window" is the kigo — winter, cold, and a literal barrier of ice, all in one image. The dash after it is the cut. Then the turn: two coffee cups, one untouched. Nobody says "argument." Nobody says "lonely." But the full, cooling cup is the argument — someone left the room, or never sat down, or stopped speaking. The reader assembles the whole story from a cup of coffee and a sheet of frost. The emotion lands harder because it was never named.
That's the difference between counting syllables and writing a haiku. The "after" isn't more elaborate — it's actually plainer. It just trusts the reader.
You can run this exact loop yourself in a couple of minutes: paste the prompt above into a chat with SentX, feed it your own scene, and use the follow-up line to push it off any line that explains itself. The craft isn't in the generation — it's in knowing what to ask for and what to cut.
Frequently Asked Questions
Does a haiku have to be exactly 5-7-5 in English?
No. The 5-7-5 count comes from Japanese on (sound units), which aren't the same as English syllables — a faithful English haiku is often shorter, around 10–14 syllables. Most contemporary English-language haiku poets treat 5-7-5 as a loose guideline and prioritize the image, the cut, and brevity over hitting an exact count. Never pad a line with a filler word just to reach 17 syllables; a real image always beats a correct count.
What's the single most important thing to fix in AI haiku?
Delete the named emotion. The fastest upgrade to any AI-generated haiku is to find the line that says "lonely," "peaceful," "joy," "sad," "calm," or any feeling word, and replace it with a physical detail that causes that feeling. Show the cold coffee, not the loneliness. This one move fixes most bad haiku on its own.
Can AI actually write a good haiku, or is it always going to be generic?
AI can write a genuinely good haiku, but only if you supply the craft constraints it doesn't apply on its own. The model has read enough real haiku to produce the form when you ask for two concrete images, a seasonal anchor, and an unexplained cut. The generic output isn't a limit of the model — it's a limit of the vague prompt. Specify the machinery and the quality jumps immediately.
What is a kigo and do I have to use one?
A kigo is a word that signals the season through a physical detail — "first frost," "cicada," "falling leaves," "long evening light." It's a traditional element, not a hard rule, but it's worth keeping because it does so much work for free: it grounds the poem in the shared physical world and sets a mood without naming it. If you drop the kigo, make sure another concrete image is anchoring the poem in something real.
How do I stop the AI from explaining the poem in the last line?
Tell it to end on an image, not a conclusion, and ban "lesson" lines explicitly. The most common AI failure is a third line that summarizes the meaning ("now I understand," "life goes on," "nature says goodbye"). In your prompt, add: "Do not state a conclusion or a lesson. End on a single physical detail." If it still explains, ask it to replace the closing line with one concrete object from the same scene and let the cut carry the connection.
Is the "turn" the same as the cut?
They work together but they aren't identical. The cut (kireji) is the pause — the dash, colon, or hard line break that splits the poem in two. The turn is what that pause produces: the leap the reader makes between the two juxtaposed images. The cut is the mechanism; the turn is the effect. A well-placed cut creates the turn; a poem with no cut has nothing to turn between, which is why so many AI haiku feel flat even when the images are fine.