Every writing tool makes an implicit claim about what counts as wrong. Most make it badly, and the result is software that sands the character off prose while calling it correction.

The distinction that matters is between a mistake and a choice.

What is genuinely a mistake

Facts that are wrong. Names misspelled. Subject-verb disagreement that was not intentional. A pronoun with no clear referent. A sentence that means something other than what the writer intended. Inconsistency with a defined house standard. A claim contradicting one made three paragraphs earlier.

These are errors regardless of style. Flagging them is unambiguously useful.

What is a choice

Sentence fragments. Starting with "And." The passive voice, which is frequently correct and has been maligned by a generation of tools that cannot tell when. Long sentences. Very short ones. Repetition for rhythm. Uncommon word choices. Paragraphs of one line. Semicolons; some people like them.

Every one of these gets flagged by mainstream writing tools. Every one is a legitimate decision a writer might make deliberately.

Why the conflation is harmful

A tool that flags choices as errors trains writers toward the median. Accept every suggestion and you produce prose that is grammatically unimpeachable and completely characterless — which is precisely the register people complain about in AI-assisted writing.

The damage compounds. Writers who accept enough suggestions stop making the choices at all, because the friction of defending them exceeds the effort of conforming. The tool does not just correct the document; it changes the writer.

The harder cases

Some things sit genuinely in between, and the honest answer is to present rather than correct.

Clarity. A sentence that took two reads is a problem — unless the difficulty is doing work, which sometimes it is.

Length. A 60-word sentence may be a mess or may be building deliberately.

Jargon. Wrong for a general audience, correct and efficient for a specialist one. The tool does not know who is reading.

Repetition. Accidental repetition is sloppy. Deliberate repetition is one of the oldest devices in the language.

For these, the right behavior is to surface the observation with its reasoning — "this sentence is 62 words and has three subordinate clauses" — and let the writer decide. Information, not instruction.

How this shapes Writewell

Writewell separates the two categories structurally rather than by tone. Mistakes are flagged as mistakes. Choices are surfaced as observations, if at all, and never with the visual language of an error.

The rule we hold to: the tool may tell you what it noticed; it may not tell you who to be.

The consequence is that Writewell flags fewer things than competing tools, which looks like weaker capability in a feature comparison. It is the opposite. Restraint is the harder engineering problem — knowing when not to speak requires understanding what the writer is doing, and flagging every deviation from the average requires understanding nothing at all.