What the AI Poem Writer does
The AI Poem Writer generates a poem on a topic you choose, in a poetic form you pick from a menu. Type "coffee" as the topic, select "haiku" as the form, click Generate, and the tool returns a 17-syllable poem in three lines. Pick "sonnet" instead and you get 14 lines of iambic pentameter with a rhyme scheme. Pick "limerick" and you get five lines of AABBA with the bouncy rhythm that makes a limerick a limerick.
The point of the form menu is that poetry isn't one thing — a haiku and a sonnet are two different kinds of object with two different sets of constraints. A tool that ignores form produces vaguely poetic prose. A tool that respects form produces actual poems, with the structural decisions a poet would make. That's the difference we're trying to draw.
Worked example. Input: topic = "coffee", form = "haiku". The tool returns:
Steam rises slowly
Bitter dark embrace warming
Monday mornings tamed
Five syllables, seven syllables, five syllables. A seasonal-ish anchor ("Monday mornings"). The juxtaposition of "bitter" and "embrace" doing the work a haiku is supposed to do — placing two images next to each other and letting the reader feel the connection. Click Generate again with the same inputs and you get a different haiku. The model writes from scratch each time.
The poetic forms, and the rules each one obeys
The form menu isn't a stylistic flourish — each form has formal rules that change what the model produces. Pick by the shape you want, not the topic.
| Form | Structural rules | What it's good for |
|---|---|---|
| Haiku | 3 lines. 5 / 7 / 5 syllables. Traditionally includes a seasonal reference (a "kigo") and a juxtaposition between two images. | A small moment, a single observation, a feeling captured in one image. Good for greeting cards, captions, meditative writing. |
| Sonnet (Shakespearean) | 14 lines. Iambic pentameter (10 syllables per line, alternating unstressed-stressed). Rhyme scheme ABAB CDCD EFEF GG. Final couplet usually carries a turn or twist. | A sustained argument, a love letter, a meditation on a theme. The form has gravitas — sonnets feel weighty even when the topic is light. |
| Limerick | 5 lines. AABBA rhyme scheme. Anapestic meter (da-da-DUM, da-da-DUM). Lines 1, 2, 5 are longer (7-10 syllables); lines 3, 4 are shorter (5-7 syllables). Always comic. | Birthday cards, toasts, light verse, anything that wants to be funny in a structured way. Usually the punchline is in line 5. |
| Free verse | No fixed meter, no rhyme requirement. Line breaks are deliberate and carry meaning. Imagery, rhythm, and condensed language do the work formal rules would otherwise do. | Modern poetry, emotional writing, when the formal forms feel artificial. Default choice if you don't know what you want. |
| Acrostic | First letter of each line spells out a word (usually the topic). No strict meter or rhyme — the constraint is the spelled word. | Personalized poems (someone's name), classroom exercises, gifts where the recipient's name is the spine of the poem. |
| Ode | Praise poem to the topic. Loose form, but typically several stanzas of 4-6 lines, elevated language, direct address of the subject ("O coffee, you..."). | Mock-serious praise (an ode to a houseplant), genuine appreciation, ceremonial writing. The form leans grand. |
The form decision is doing more work than the topic decision. "Coffee, haiku" and "coffee, sonnet" produce two completely different poems with the same subject — one is a small observation, the other is a 14-line argument about coffee's place in the morning. Pick the form first, then plug in the topic.
When you'll use it
Most people don't sit down to write poetry on a normal Tuesday. The Poem Writer earns its place in the moments where a small poem would land but you don't have time to write one yourself:
A birthday card or anniversary card. A haiku for someone's birthday hits harder than "Happy Birthday!" because it's specific. Plug in the recipient's name or hobby as the topic; the form does the rest.
A wedding toast or eulogy. Opening a toast with a four-line poem about the couple's first meeting is the kind of thing people remember. The Acrostic form is especially good here — the spelled-out word can be the couple's joint surname or a shared word like "TOGETHER."
Social media captions. A limerick about your weekend, a haiku about your morning view, a free-verse line about how the dog is sleeping. The poem is the post.
Teaching poetic forms to students. Generate a sonnet about a topic, then compare it to a Shakespeare sonnet on the same topic — the structural rules are visible in both. Students can see what "iambic pentameter" actually looks like in practice.
A gift inside a gift. An ode to a houseplant attached to a houseplant. A limerick taped to a coffee mug. The poem is a small piece of personalization that costs nothing and makes the gift feel considered.
Writer's block in your own poetry. Generate three sonnets about your topic and read them. They won't be your sonnet, but reading three near-misses often unlocks the line you'd been trying to write.
How the model handles form constraints
Sonnets and haikus are easier to grade than jokes — there's a right answer to "is this 5-7-5?" Either it is or it isn't. The model handles strict syllable counts most of the time, but it does miss occasionally, especially on words with ambiguous pronunciation (is "fire" one syllable or two? is "every" two or three?). If your generated haiku reads 5-8-5, regenerate. The model usually nails it on the second try.
For sonnets, the harder constraint is iambic pentameter (the 10-syllable, alternating-stress rhythm). The model gets the syllable count right almost always, but the stress pattern is rougher — about three lines out of fourteen will scan awkwardly when read aloud. Read the sonnet aloud before using it. If a line feels lumpy, regenerate or rewrite that line yourself.
For limericks, the AABBA rhyme is reliably correct. The anapestic meter is hit-or-miss — if reading it aloud, your voice can't find the bounce, the meter is off. Free verse and odes have no strict rules, so the model "passes" by default; the only test is whether the result feels like a poem rather than chopped-up prose.
How the AI writes a poem
Under the hood, the AI Poem Writer sends a prompt to a language model (Claude, in our case) that contains the topic, the form, and an instruction to follow the form's rules. The prompt looks roughly like:
System: You are a poet. Write a [form] on the topic of [topic]. Strictly follow the structural rules of a [form]: [the rules for that form]. The poem should feel like a real poem, not a description of one.
The model then writes one word at a time, predicting the most likely next word given everything that came before. Because the prompt includes the structural rules, the model "anchors" on them as it generates — counting syllables for a haiku, tracking rhymes for a sonnet, listening for meter on a limerick. It's not always perfect, but it's closer to following a form than a model writing free-flowing text.
The interesting thing about LLM poetry is that it's much better at poems with strict rules than at poems with no rules. Constraints help the model — they give it a target to aim at. A haiku is a small, well-defined target. A free-verse poem is an open field, and the model wanders. If you're getting weak output, try a stricter form. The constraint usually improves the result.
Why this costs credits (and the membership math)
Each poem generation uses a small amount of compute on the AI provider's side — generally less than a third of a cent. Microapp passes that cost through at the actual rate; we don't mark up AI compute. Members have a monthly credit allowance that covers normal use; non-members get a daily cap and see ads. That's the whole pricing model.
A sonnet costs slightly more compute than a haiku, but the difference is fractions of a cent. The cost is real but small enough you won't notice unless you're generating hundreds of poems per session. If you are — seeding a classroom with examples of each form, say — the membership math is in your favor.
The reason we charge anything is that ad-funded AI tools have a conflict of interest baked in: the operator wants you to click a lot and not think about quality. A member-funded one wants you to get the poem you needed and close the tab. We picked the second model on purpose.
Tips for getting better poems
Be specific about the topic. "Love" produces clichés. "The exact moment you realize you've forgotten what someone's voice sounded like" produces something that feels written. Specific topics force the model to grip details rather than generalities.
Match form to occasion. A sonnet for a wedding is fine. A sonnet for a Tuesday email about a deadline is overkill. Haiku and limerick are good defaults for daily use; sonnet and ode reserve for moments that earn the weight.
Generate three and pick one. The model is uneven. The first generation might be flat; the second might be the one. Most poems benefit from the comparison.
Edit the last line. Endings are hard for the model. About a third of generated poems land 90% well and then end on a weak line. Rewriting the last line yourself is usually a five-second job that lifts the whole poem.
Read it aloud. Especially for sonnets and limericks, the rhythm is what makes the poem work. If reading it aloud feels lumpy, the meter is off — regenerate or fix the lumpy line.
The acrostic trick for personalized gifts. If you're writing for a specific person, use the Acrostic form with their first name. The poem then literally spells out the recipient — every line starts with one of their letters. It takes ten seconds and reads as far more personal than a generic poem with their name dropped in.
Poetry vs. AI-written prose
AI-written prose has a famous problem: it sounds like AI-written prose. The model averages out the corpus and produces text that's grammatically clean but voiceless. Poetry partially escapes this trap because poetic forms are constraints that push the model away from default register. A haiku can't be 200 words of filler. It has to be 17 syllables in three lines — so tight the model can't fall back on its usual habits. That's why generated sonnets and haikus often read better than generated essays. If a poem feels generic, switch to a stricter form.
Related writing and generation tools
The AI Poem Writer is part of a small toolkit of generation tools on Microapp. Tools that pair naturally with it:
- The AI Joke Generator writes original jokes in five styles. A limerick is a joke with a meter; a knock-knock is a joke with a script. The two tools overlap.
- The AI Bio Generator drafts professional or personal bios. A short poem at the end of a bio is a memorable signature.
- The Caption Generator writes social-media captions. Pair a poem with a caption for a complete post.
- The Random Word Generator gives you single words on demand — useful when you want to write a poem about something unexpected but don't know what.
- The Paraphrasing Tool rewords passages while keeping meaning. Useful when a generated poem's meaning is right but a line's phrasing is off.
Frequently asked questions
Can I use these poems for a wedding, eulogy, or commercial project?
Yes. Anything the AI Poem Writer produces is yours to use however you want — read it at the wedding, print it on the card, use it in your marketing copy. Microapp doesn't claim copyright on generated output. Standard caveat: AI-generated text isn't currently eligible for copyright protection in most jurisdictions, so anyone else can also use the same output if they happen to regenerate it. For most uses (a personal card, a toast), that doesn't matter.
Why does my haiku sometimes have the wrong syllable count?
The model is good at syllable counts but not perfect. Words with ambiguous pronunciation ("fire" — one syllable or two? "every" — two or three?) trip it up about one time in five. If your haiku reads 5-8-5 or 4-7-5, regenerate. The second attempt is usually correct. If you want to verify the count yourself, syllable-counting is a quick mental check: clap once per syllable as you read aloud.
Are the poems original?
Yes — the model writes each poem from scratch. It's drawing on a deep training corpus of existing poetry, so the style might feel familiar (a sonnet by the model will read recognizably as a sonnet, not as something new and weird). But the specific lines are generated, not retrieved. If you're worried about an accidental match with an existing famous poem, search a distinctive line on Google after generating. Anything truly common will surface.
Can I write in languages other than English?
The model handles many languages, but English is where the form rules are tightest. A haiku in Spanish works fine because the form (5-7-5 syllables) is language-agnostic. A sonnet in Spanish works but the meter changes (Spanish sonnets traditionally use 11-syllable lines, not 10). The model knows these conventions but is more reliable on English-language poetry. For non-English output, generate and then verify the form with someone who knows the local convention.
Why are my free-verse poems weaker than my haikus?
Counterintuitively, the model handles strict forms better than open ones. A haiku has a tight target (17 syllables, three lines, one image); the model has something concrete to aim at. Free verse has no target, so the model defaults to a slightly-prosey register. If you want a "proper" poem from the generator, pick a stricter form. The constraint usually helps.
Can the generator write in the style of a specific poet?
To a degree. The model has read enough Whitman, Dickinson, Frost, and Rumi that asking for "a free-verse poem in the style of Walt Whitman about coffee" produces something with Whitman-ish features (long lines, repetition, expansive register). It won't be a Whitman poem, but it'll feel closer to one than the default. Works well for canonical poets; less reliably for obscure or very recent ones.
What's the difference between this and just asking ChatGPT for a poem?
Mechanically, the same thing happens: a language model writes a poem from your prompt. Practically, the Microapp tool removes the prompt-writing step — pick the form from a menu and the rules are pre-loaded. No signup, no chat interface, no waiting. Drop in, get the haiku, close the tab.
How long should I expect a sonnet to be?
14 lines. Each line is about 10 syllables, so the total length lands around 100 words. If your generated sonnet is dramatically shorter or longer, regenerate. The form fixes the length.