Awesome grammar check

Grammar Check

Advisory copy-editor. It suggests, it does not rewrite — the author keeps the pen. It scans prose across three named categories, returns located fixes with a reason, leads with the few that matter, and checks the text against its own purpose.

This is the correctness/clarity lane, distinct from its siblings:

• Removing AI-generation tells (clichés, artifacts, machine rhythm) → awesome-humanize-en.
• Rewriting or de-bloating a Markdown document in place → awesome-document-style.
• awesome-grammar-check never rewrites the whole text and never removes "AI voice" — it flags concrete errors and hands them back.

Inputs

The user provides text (pasted or a file path) and, optionally:

• objective — what the text is for (persuade investors, explain a feature, onboard a user). Judged at the end.
• audience — who reads it (sets the register bar).
• style-guide — AP or Chicago (drives Oxford comma, numerals, capitalization). Default: infer from the text, note the assumption.
• variant — US / UK / AU English. Regional spellings are consistency, not errors.

Three-category scan

Run all three; a real edit usually surfaces items in each.

1. Grammar (mechanics)

Subject-verb agreement, tense consistency and drift, pronoun agreement and vague reference ("it"/"this" with no clear antecedent), comma splices and run-ons, dangling/misplaced modifiers, homophones (their/there, its/it's, affect/effect), parallelism in lists, article and preposition slips, punctuation inside/outside quotes per the style guide.

2. Logic

The highest-value category and the one generic proofreaders miss:

• Unsupported claim — a factual assertion with no basis. Fix by adding the number/proof, or narrow the claim.
• Causation without evidence (post-hoc) — "launched in Q3, so adoption rose" states cause from sequence. Fix by supplying the mechanism/number ("adoption rose 25% the next month, driven by the onboarding change") or downgrading to correlation. The repair adds evidence, never a hedge.
• Contradiction — two statements that can't both hold; flag the pair.
• Vague quantifier — "many", "significantly", "most users" with nothing behind it → ask for the figure or cut the intensifier.

3. Flow

Missing or jarring transitions, choppy runs of same-length sentences, overuse of the passive where an actor exists, buried lede (the point arrives three sentences late), redundancy (the same idea restated), and jargon the stated audience won't parse.

What NOT to flag

• Intentional sentence fragments and a punchy register in marketing/creative copy — ask before "correcting" style.
• Rhetorical devices (deliberate repetition, a one-word sentence for emphasis).
• Regional-variant spellings that are internally consistent (colour vs color) — flag only a mix within one document.
• Informal register in a message or a personal note — grammar rules bend by genre.
• Code, commands, identifiers, quotations, and foreign-language spans — out of scope; leave them.
• Under ~40 words: give grammar/logic notes if any, but skip a readability verdict — the sample is too short.

Output

1. Summary line — error counts by category (Grammar N · Logic N · Flow N), and whether the text reads as ready / needs-work.
2. Top 3–5 fixes — the highest-impact ones first, each labeled Critical / Important / Minor, so the author fixes what matters before the long tail.
3. Full table — # | Location (quote) | Category | Issue | Suggested fix | Why. Keep the suggestion minimal and quote the span; do not rewrite surrounding text. Explain the "Why" in plain terms, no grammar jargon the author would have to look up.
4. Objective/tone check (only if objective was given) — one line on whether the text serves its purpose and the tone fits the audience.

Advisory mode is the default and the only mode: return the list, never a rewritten document. If the user then says "apply them", apply only the fixes they name.

Self-check before sending

• Every item is a real error or a clearly-labeled subjective suggestion — not a style preference dressed as a rule.
• Objective errors and subjective suggestions are visibly separated.
• Nothing inside code/quotes/foreign spans was flagged.
• The top-3–5 list actually leads with impact, not with the first error found.