A poetry humanizer with stated limits

One request to our AI server, carrying your poem and a rewriting instruction. That instruction is to rewrite the supplied text to accomplish the transformation implied by the tool name, while preserving the original meaning and factual claims, and to return only the rewritten text with no preamble, commentary, quotation marks or fencing.

Everything the tool does follows from that sentence, and so does everything it fails to do. Nothing runs locally on this route, which is why an allowance applies: five runs per rolling 24 hours, counted from first use rather than reset at a clock boundary, with no account and no paid tier in the product at all.

Read the instruction again for what is absent. It protects meaning and factual claims. It does not mention line breaks, stanza breaks, metre, rhyme, syllable count or the visual shape of the poem on the page. It is a general rewriting instruction with a tool name interpolated into it, not a prosody-aware one.

The consequence is precise rather than vague: lineation is unprotected by default. For prose that is correct behaviour, since a paragraph's line breaks are an artefact of rendering. For a poem it inverts the priority, because the break is frequently the unit carrying the meaning, and the wording is the part more safely changed.

This is worth stating rather than hiding, because it is also the thing you can do something about. The instruction field is where you close the gap yourself.

The sequence is deliberate. Step two exists because step one gives the model no constraint it will infer on its own.

Step 1

Paste the poem. The meter confirms the character count against the 12,000 ceiling.

Step 2

State the invariant in the instruction field, naming what must not change rather than what should improve.

Step 3

Run once, then diff the result against the original by line count before reading it for quality.

There is no style selector, no tone preset, no language dropdown and no strength slider. There is one optional free-text field of up to 300 characters, and whatever it contains is appended to the system instruction verbatim, introduced as an additional user instruction.

Free text is a strictly more expressive control than a menu for a form like this, because the constraint that matters in a given poem is usually idiosyncratic. A dropdown can offer free verse or spoken word. It cannot offer keep the caesura in line four, or keep the same number of syllables per line, or do not touch the final couplet.

The corollary is that a vague instruction buys nothing. Specificity is the entire mechanism: name the invariant, not the aspiration.

The input ceiling is 12,000 characters, enforced in the text box and again at the server, which rejects anything outside 1 to 12,000 characters before a request is made.

The output ceiling is 1,400 completion tokens, and it is unrelated to the first. The two are measured in different units and applied at different ends of the round trip, which is why a poem can pass the input check and return truncated. A result that stops mid-stanza has hit the second ceiling, not produced a considered ending. For anything long, sectioning the poem before the first run is the only remedy, since no setting raises either limit.

Sampling temperature on this route is 0.55, a deliberate non-zero value shared by every rewriting tool here. Identical input therefore yields different output on every run, and no option exposes a deterministic mode.

Against a five-run allowance that has a real cost. Rerolling is the obvious response to a result you dislike, and it is the expensive one: four rerolls exhaust the day on a single poem, and each roll is an independent sample rather than a refinement of the last. Constraining the first run is the only form of iteration this design actually rewards.

No claim is made, no detector is run here, and no third-party verdict is predicted. Competing pages in this category advertise otherwise: a 98 percent bypass rate against a named list including GPTZero, Turnitin and Copyleaks on one, and 100 percent bypass on another's paid tier while its own free tier is documented as unable to. A percentage against software you do not control, measured by nobody independent, is not a specification. There is no affiliation with any detection service.

The separate and more interesting question is whether the output is good, and the honest answer is that the tool is not evaluating that. It resamples wording. Whether an image is load-bearing, whether the ending earns itself, whether a line should be cut entirely, are judgements outside what a rewriting request performs.

It fits best on the smallest unit. A single line pasted alone cannot hit the output ceiling, raises no lineation question, and returns several plausible alternatives across runs, which is exactly what someone stuck on one line wants. It fits well on verse written to a brief, lyrics and occasional poems, where the line shape is subordinate to the words.

It is weakest precisely where the poem is most itself: short lines, deliberate white space, enjambment doing the work. There the ratio of what a wording rewrite can improve to what it can disturb is unfavourable, and the instruction field is a mitigation rather than a fix. Knowing which of those two cases you are in is the whole decision.