A Grok sentence rewriter without modes

The operative instruction is short enough to quote in full effect: our server asks for a rewrite of the supplied text that preserves the original meaning and factual claims, returning only the rewritten text with no preamble or commentary. Everything this route can and cannot do follows from that constraint.

It operates on the surface of a passage. Word order, clause length, register, the position of the main assertion, the density of framing. Those are all properties of expression, and a constrained rewrite can change them without touching what is being asserted.

It does not operate on the information content. Preserving claims means the claim set at the output equals the claim set at the input, so a passage that asserts nothing specific cannot be improved into one that does. The route is remote, metered at 5 runs per rolling 24 hours, and bounded at 1 to 12,000 characters per run.

The only Grok-specific source we fetched is an editorial guide on rewriting Grok text. Its account is qualitative throughout and it publishes no measurements, no sentence-length distributions and no frequency counts. We report it at that level of confidence and assert nothing numeric about xAI's model.

Audience-neutral voice

The guide describes a generic voice that could apply to executives, students or marketers. Correct, fluent and addressed to nobody in particular.

Assertion buried under framing

Text that sounds confident while the main point sits beneath extra framing. This is the one fault on the list that reordering alone resolves.

Default rhythm

Phrasing the guide calls too synthetic, too stiff, or too close to the default AI rhythm. Uniformity across sentences rather than any individual word choice.

Unjoined blocks

Paragraphs described as separate blocks of information placed beside each other, with the transitions between them absent.

Silent term substitution

The guide notes drafts swapping similar terms in a way that seems natural. This is a property of automated rewriting generally, ours included, and it is the risk the rest of this page returns to.

The method exists because two constraints bind at once: a five-run allowance and the absence of any verification surface in the product.

Step 1

Select a passage small enough to verify. The ceiling is 12,000 characters, but verification cost scales with length and there is no diff view, so the practical unit is a paragraph.

Step 2

Specify the transformation in the instruction field. Three hundred characters is enough for a testable constraint: lead with the conclusion, cap sentences at twenty words, leave every quoted term unaltered.

Step 3

Perform a side-by-side comparison before adopting the output. Check nouns first, since substitution is the likeliest defect, then confirm each assertion in the source survives unchanged in the result.

The competing pages in this result set are organised around mode lists. One prominent rewriter publishes two free modes, Standard and Fluency, and places Formal, Humanize, Academic, Simple, Creative, Expand, Shorten and Custom behind a premium tier. Another offers Clearer, Shorter, More formal and Friendlier with a 5,000 character cap.

Read structurally, a mode is a stored instruction with a label on it. Tiering them is a commercial decision rather than a technical boundary, which is worth knowing before concluding that a free tool is a less capable one.

This route exposes the instruction directly: one free-text field, 300 characters, no tier. The cost is that a vague request produces a vague result, because an adjective is not a specification. The benefit is that a compound constraint, shorter and less formal but with the terminology untouched, is expressible here and is not selectable from any menu.

Meaning preservation is an instruction, not a guarantee, and the failure it fails toward is specific. Rewriting substitutes near-synonyms, and near-synonyms are near only outside a domain. Deprecated and removed are interchangeable in ordinary prose and are not interchangeable in a release note.

The cost is asymmetric. A rewrite that leaves a passage clumsy costs a run. A rewrite that converts estimated into measured, or recommended into required, produces a fluent sentence carrying a claim the source did not make, and fluency is what stops you noticing.

Two controls are available and both are manual. Enumerate protected terms in the instruction field before the run, and read the output's nouns against the source afterwards. Nothing in the interface performs this check, so the page states it rather than implying the model handles it.

We make no prediction about any detector's output, run no tests against one, and hold no affiliation with any detection service. A competitor shipping a mode named for this outcome does not make the claim more defensible.

The deeper problem is that a detector verdict is not a property of the text in isolation. It is the output of one classifier, with its own threshold and its own error profile, on one day. Optimising against it means optimising against an instrument rather than a reader.

The properties a rewrite can be held to are the ones you can inspect: length, ordering, register, and whether every claim in the source survived. Those are verifiable by reading, which is what makes them worth targeting.

No mode selector, tone preset or style tier. No similarity metric, no meaning-kept score and no confidence value, so no number on this page will vouch for a rewrite. That is a deliberate omission rather than an oversight, because a score we could not justify would be worse than none.

No change visualisation. No diff, no highlighting, no revision history. Output is plain text in a read-only pane with a Copy button, and the comparison is performed by the user.

No document intake in any format, no download, no batch queue and no programmatic access. One run at a time against a five-run rolling allowance, with formatting retained only in whatever document you copied the text out of.

The productive case is a passage with a diagnosed fault. The writer already knows the point is buried, or the register is wrong for the audience, and wants an alternative arrangement of material they have already assembled.

Technical writers apply it to procedures that drifted formal enough to obscure the step. Analysts apply it to an opening paragraph that returned as framing rather than finding. Second-language writers apply it as a check, to see whether a sentence they suspect is stiff has a plainer form they had not located.

The unproductive case is submitting a full draft to discover what changes. Under a five-run allowance with no diff view, an unexamined bulk rewrite is an unverified one, and an unverified rewrite is exactly where a substituted term survives into publication.