Four-line poems from chemical elements are the hardest creative format in this app — and the most honest thing to say is that they work, but barely

Can you write poems with chemical element symbols? Yes. The app already ships 150 of them and they are real, verified, four-line poems where every word on every line resolves to a chain of element symbols.

Are they good poems? That depends on your definition. They are structurally valid and often evocative — words like “agony,” “shore,” and “bliss” recur because they survive the element filter. But the vocabulary is so constrained that the writing tends toward the incantatory rather than the conversational. If you value precision under constraint, some of these are genuinely impressive. If you want freedom of expression, poems are the wrong format to start with.

How does this compare to sentences? Sentences use 138 unique words in English and reuse each one 3.5x on average. Poems use only 67 unique words and reuse them just 2.1x. The creative palette shrinks, but so does the repetition — poems are trying harder to be different each time, which is the expected behavior of a creative format.

The poem corpus is not human-authored in the traditional sense. Every poem is generated by Google Gemini as a candidate, then run through an elementizability filter that checks whether each word on each line can be decomposed into a valid chain of periodic-table symbols. Only poems where every single word passes survive. Finally, the build step enforces a hard structural rule: only poems with exactly four lines are kept.

This pipeline means the creative process is fundamentally generative-then-filtered. The LLM proposes, and the periodic table disposes. A poem that sounds beautiful but contains one word that cannot be elementized is discarded entirely — there is no partial credit. That strictness is what makes the surviving poems interesting: they represent the intersection of two very different constraint systems (linguistic coherence and chemical symbol decomposability).

The honesty here matters. These are not poems written by a chemist with a dictionary. They are machine-generated candidates that passed an unusually strict filter. The quality of the output reflects both what the LLM can imagine and what the periodic table will accept.

All of the numbers in this article — word counts, vocabulary sizes, line-position analyses — are derived from the 150 poems that survived this pipeline. They are corpus statistics, not projections.

The pattern is subtle but real. In English, line 1 uses 22 unique words and line 4 uses 23. Lines 2 and 3 — the middle of the poem — are where vocabulary peaks, at 28 and 27 unique words respectively. Opening lines and closing lines are more constrained, which aligns with how quatrains tend to work even in unconstrained poetry: the first and last lines carry structural weight (they set up and resolve), while the middle lines have more room to wander.

The word-count story is different. Line 2 averages 4.2 words per line — the highest of any position — while line 1 averages only 3.6. This suggests that opening lines lean shorter: a few decisive words rather than a clause. Closing lines sit in between at 3.9 words.

The word “is” dominates every line position, which is not surprising given that it is one of the most elementizable words in English (I + S, Iodine + Sulfur). It appears 24 times across all line positions. That single word tells you a lot about what element poetry sounds like: declarative, present-tense, and built around copula structures (“X is Y”) because the verb “is” is one of the few connective words that survives the filter.

The vocabulary reuse number deserves its own discussion because it is counterintuitive. In most natural language contexts, a lower reuse ratio means the text is more diverse — each word appears fewer times, so the writing covers more ground. But in an element-constrained system, low reuse can also mean the system is struggling to find valid words and has to keep reaching for new ones.

English poems land at 2.1x, which is low. For comparison, French poems — the largest poem subcorpus — sit at 1.4x. That gap tells you something about the relative depth of element-friendly vocabulary in each language. French has more elementizable words to draw from, so each poem can afford to introduce fresh vocabulary. English poems, working with a much smaller elementizable set, still manage to avoid heavy repetition, but the effort is visible in the constrained diction.

Dutch sits at the other extreme with 3.2x reuse across 18 poems — the highest reuse in the poem corpus. This means Dutch element poetry leans heavily on a small core vocabulary, recycling proven words across poems rather than exploring. The most frequent Dutch poem word is “is” at 64 occurrences, which acts as the same kind of structural crutch that “is” provides in English.

The most frequent words in English poems are revealing: “is” appears 24 times, “here” appears 8 times, and “i” appears 7 times. These are not poetic showpieces — they are the structural glue that makes element poetry grammatically viable. The actual creative content comes from lower-frequency words like “agony,” “shore,” “arcane,” and “consciousness.”

The French dominance is worth explaining. French has an unusually large set of short, elementizable words — many one- and two-syllable nouns and verbs that happen to decompose cleanly into element symbols. Words like “neige,” “bruine,” “fracas,” and “espoir” all survive the filter. That gives the LLM a much deeper vocabulary to draw from, which translates to more poems clearing the pipeline.

Spanish is second with 30 poems, partly because short function words like “un” (which maps to Uranium + Nitrogen) appear constantly and provide grammatical scaffolding. The most common Spanish word “un” appears 83 times — an anchor word that holds the corpus together. Italian benefits from similar Romance-language patterns, contributing 27 poems.

English trails at 9 poems because its phonology and spelling produce fewer element-compatible words — the vowel-consonant clusters in English are often one character off from matching a symbol pair.

German is the most striking outlier: only 4 poems. German compound words are typically long and multi-syllabic, which makes them harder to decompose. The element filter rewards compact, symbol-dense words, and German vocabulary skews long. With a vocabulary reuse of just 1.2x and only 41 unique words, the German poem subcorpus feels like an experiment that barely survived the pipeline.

Iodine (I) and Sulfur (S) dominate both poems and sentences, which makes sense — they map to the most common English letters. But Tungsten (W) is the standout poem-specific element: it covers 5.1% of poem symbol occurrences versus roughly 0% in sentences. That is because poems use more words starting with “w” — “white,” “worn,” “was” — which reflects the particular emotional register that element poetry favors.

Rhenium (Re) also gains prominence in poems at 4% share. This element powers words with “re” prefixes or internal clusters — “here,” “caress,” “ruin.” Again, the periodic table is shaping the diction: elements that map to common English letter pairs become load-bearing infrastructure for the poems.

Helium (He) appears more frequently in poems than sentences as well, powering words like “here” and “her.” The creative implication is that poems lean toward immediacy and presence — words about being here — partly because the periodic table makes those words easy to construct.

With 50 poems, French is the only language with enough data to draw confident line-by-line conclusions. The French subcorpus uses 185 unique words — far more than English’s 67 — and has a vocabulary reuse of 1.4x.

That low reuse is remarkable for a corpus of 50 poems. It means French element poetry is exploring vocabulary broadly rather than leaning on a few reliable words. The most common word is “personne” at only 5 occurrences — compare that to English where “is” appears 24 times. French poems do not need a structural crutch word because the language has enough element-friendly vocabulary to support diverse constructions.

The French line-by-line pattern is also distinct. French poem lines average just 1.3 words compared to English’s 3.9. French poems are terser — fewer words per line, but longer and more complex individual words. The result is a different texture: where English element poetry sounds declarative (“X is Y”), French element poetry sounds imagistic — single nouns or noun phrases that evoke rather than state.

If the constraint-heavy nature of element poetry appeals to you, start on the home page and enter a four-line poem idea. The app will evaluate each line independently. When all four lines pass, move to the designer for a print-ready layout.

If the poem idea collapses — which it will, more often than not — the fallback path is productive rather than frustrating. Drop individual lines into the sentence flow via the sentences article, or pull individual words from the words article and rebuild from there. The element constraint is fractal: the same rules that govern a poem also govern a word, so everything you learn from failed poems makes you better at finding successful words.

For reference, the French poem below is one of the best showcases in the corpus. Try entering it to see how the system renders multiline content: