ChatGPT grammar checker that reports

This route runs a classifier, not a rewriter. The server instruction for that family is to analyse the supplied text and return clear findings, which means the output is a list of identified problems with reasons attached, not a corrected document. The distinction is the design decision this page is built around.

A published side-by-side test of ChatGPT against Grammarly articulated the cost of the alternative precisely. The reviewer pasted a deliberately error-filled paragraph into the chat window, received corrected paragraphs within seconds, and recorded that there was no way to know what the mistakes had been. Grammarly, by contrast, highlighted them. A silent correction transfers the error out of your document and out of your knowledge at the same time, which is a bad trade on anything you will write again.

The searchable literature on this topic consists largely of prompt lists to paste into a chat interface. That wrapper is fixed here, which removes the main source of run-to-run variance a user controls.

Step 1

Submit between 1 and 12,000 characters. Submit contiguous prose, since agreement and reference errors are only decidable with their antecedents present.

Step 2

Supply an instruction of up to 300 characters narrowing scope: a single error class, protected terminology, or a request for explanations.

Step 3

Adjudicate each finding yourself. Nothing is applied automatically, so a wrong flag costs you a moment rather than a silent edit.

A deterministic grammar checker parses and applies rules. Its behaviour is reproducible: the same input yields the same output on every run, it catches exactly the violations its rules encode, and it is silent on everything else. Its characteristic failure is a sentence that satisfies every rule and communicates nothing.

A model-based check has the inverse profile. It can flag prose that is grammatically well formed and genuinely unreadable, which no rule set covers, and it can register a construction as odd when it is merely uncommon in the register it expects. It is also not reproducible: two submissions of identical text can return different lists, because generation is sampled rather than derived. The practical implications are specific. Run anything consequential twice and take the union. Treat a single flag on an unusual but deliberate construction as weak evidence. And do not read silence as clearance, because absence of a finding is not the same as a rule having been checked and passed.

The competing genre is a list of prompts to type into a chat window. A 300-character instruction supplied alongside a fixed analytical wrapper is the same expressive power with less to get wrong. The useful constraint classes are these.

Error-class narrowing

Restricting to punctuation, or to agreement, or to tense. Reduces the flag volume that dilutes attention on a long document.

Protected vocabulary

Naming terminology, product names or defined terms as fixed. A check that does not recognise a term will otherwise treat it as a mistake.

Explanatory mode

Requesting a reason with each finding. Converts a correction list into something with transfer value beyond this document.

Register declaration

Naming the document type. Register violations are not grammatical errors and go unreported unless asked for.

Learner targeting

Declaring a first language or a known weak area. Article and preposition use in particular benefits from being named rather than discovered.

Model-drafted text is rarely ungrammatical at the surface. What remains is the set of errors that read smoothly, which is precisely the set a human proofreader skims past.

Long-distance agreement

Subject and verb separated by intervening clauses, with the verb attracted to the nearest noun rather than the actual subject.

Tense inconsistency

Aspect shifting across a sentence boundary inside a single narrative span. Invisible on a sentence-by-sentence read.

Misattached modifiers

A participial opener whose grammatical subject is not the one the writer intended, producing a parseable and absurd reading.

Comma splices

Independent clauses coordinated by a comma alone. Fluent, frequent, and still an error in edited prose.

Unresolved anaphora

A demonstrative pronoun with multiple available antecedents and no disambiguation. Grammatical, and unreadable.

Stacked hedging

Two modal qualifiers doing one job. A stylistic defect rather than a rule violation, and reported as such.

There is no browser extension, no editor plugin and no in-document integration of any kind. The comparison literature is consistent that this is the real dividing line: a resident checker runs in real time wherever you are typing and offers instant suggestions, while a pasted check requires you to leave the document and come back.

The consequences follow directly. Nothing is underlined as you write, nothing is corrected on acceptance, and no document is modified by this tool under any circumstance. There is also no file input, so a document has to be copied out as text first. The compensating property is that a single pass over a finished draft presents the complete finding set at once, which surfaces distribution-level problems, a tense that drifts across three paragraphs, a hedge that recurs eleven times, that a sequential stream of underlines does not make visible.

Submissions are accepted between 1 and 12,000 characters, approximately 1,800 to 2,000 words, with 5 runs per rolling 24 hours measured from first use rather than reset on a calendar boundary.

Splitting a longer document has a methodological cost worth planning around. Agreement and reference errors are only decidable when the antecedent is inside the submitted span, so a split that cuts mid-paragraph strands the evidence the check needs and produces both misses and spurious flags. Split at section boundaries. Given the 5-run allowance, a document requiring three splits consumes most of a day's capacity, which is a genuine constraint on using this as a production proofreading step.

Writers producing documents with a single delivery: submissions, applications, proposals, anything published rather than iterated. The complete finding set on a finished draft is worth more here than incremental correction during composition, because the errors that matter at this stage are distributional.

Writers working in a second language who want the reason rather than the fix, since a silently corrected article teaches nothing about when the article was required. And editors processing model-drafted copy, where surface fluency is high and the residual errors are concentrated exactly in the long-distance and anaphoric classes above.