Line editing, structural editing and the ruthless cut, kept separate. Below are 7 copy-ready prompts. Fill in the [BRACKETS], copy, and paste into ChatGPT, Claude, Gemini or any capable assistant.
Editing is where AI earns its place most quietly. It never gets bored on page nine, and it has no ego about the sentence you were proud of.
The 7 prompts
Line edit without changing the voice
Tighten prose while keeping it sounding like the author.
Line edit this. Preserve the author's voice. TEXT: """ [PASTE] """ AUTHOR'S INTENT: [WHAT THEY ARE TRYING TO ACHIEVE] REGISTER: [formal / conversational / technical / literary] WHAT NOT TO TOUCH: [ANYTHING OFF LIMITS] Edit for: 1. Sentences that are longer than their content justifies 2. Weak verbs propped up by adverbs ('walked slowly' → 'ambled') 3. Passive voice where the actor matters 4. Nominalisation ('made a decision' → 'decided', 'provides assistance' → 'helps') 5. Filler: 'in order to', 'the fact that', 'it is important to note', 'needless to say', 'basically', 'essentially' 6. Repeated sentence openings and repeated rhythm 7. Vague quantifiers: 'many', 'several', 'various', 'a number of' 8. Mixed or dead metaphors 9. Ambiguous pronouns Output as a table: Original | Edited | Reason. Then the clean edited text. Rules: - Do NOT make every sentence the same length. Varied rhythm is a feature. - Do not remove the author's characteristic constructions even if unconventional. If a habit is deliberate, leave it and note it. - Do not simplify vocabulary the reader will know. - Do not impose your own preferences about em-dashes, semicolons or sentence fragments. - Flag anything you changed where you are unsure whether it was intentional style. At the end: the three edits that most improved it, and the one change you considered but did not make because it would have cost the voice.
What you get: A table of changes with reasons, clean edited text, and an explicit note on the edit deliberately not made.
Tip: The 'do not make every sentence the same length' rule matters. Unchecked line editing produces uniformly medium sentences, which is technically correct and dead to read.
Structural edit: fix what is wrong with the shape
Diagnose problems that line editing cannot fix.
Give me a structural edit. Do not line edit. PIECE: """ [PASTE] """ WHAT IT IS MEANT TO DO: [PURPOSE] READER: [WHO] MY WORRY ABOUT IT: [WHAT FEELS WRONG, IF YOU KNOW] Assess: 1. THE ARGUMENT - state what this piece actually argues, in one sentence, based only on what is on the page. If it does not argue anything, say so. If what it argues differs from what I said it should, that is the finding. 2. THE SHAPE - map it: what each section does and how it connects to the next. Produce this as a list so I can see the skeleton. 3. WHAT IS IN THE WRONG PLACE - context that should come earlier, background that should come later or go entirely, a conclusion buried in the middle, and the real opening that is currently in paragraph four. This last one is common: writers warm up for several paragraphs before starting. 4. WHAT IS MISSING - a step in the argument that is assumed rather than made, an objection not addressed, a claim with no support. 5. WHAT DOES NOT BELONG - sections that are good but are a different piece. 6. PROPORTION - a table: section | word count | % of piece | importance to the argument. Flag anywhere the two columns disagree. Long treatments of minor points and short treatments of central ones are the most common structural fault. 7. THE REORDER - the proposed new structure, with what moves where and what gets cut. 8. THE HARD QUESTION - if this piece is trying to do two things, say so and which one to keep. Do not fix sentences. Structural problems survive line editing perfectly intact.
What you get: The argument as it actually appears on the page, a structural map, a proportion analysis, and a proposed reorder.
Tip: Point 6's proportion table is the fastest structural diagnostic there is. Word count should track importance, and when it does not, the piece feels off for reasons nobody can name.
Proofread for errors without rewriting
Catch the mistakes, change nothing else.
Proofread this. Report errors only - do not improve the writing. TEXT: """ [PASTE] """ VARIETY OF ENGLISH: [British / American / Australian / other] STYLE GUIDE IF ANY: [NAME, or 'none'] CONTEXT: [WHERE THIS WILL BE PUBLISHED] Report as a table: Location (quote 5 words around it) | Error | Correction | Type. Check: - Spelling, including the right word spelled correctly in the wrong place (their/there, its/it's, affect/effect, principal/principle, complement/compliment) - Grammar: agreement, tense consistency, dangling modifiers, parallel structure in lists - Punctuation: apostrophes, comma splices, quotation mark and punctuation order for my chosen variety - Consistency: hyphenation, capitalisation of the same term, number formatting (numerals vs words), date formats, serial comma used or not used throughout, spelling variety consistent throughout - Names, titles and proper nouns spelled consistently - Numbers: do totals add up, do percentages make sense, are any figures internally contradictory - Formatting: heading hierarchy, list punctuation consistency, spacing - Broken or placeholder text: lorem ipsum, TK, XX, [brackets], double spaces, unclosed quotes or parentheses Then separately: - POSSIBLE ERRORS - things that may be intentional. Do not correct, just flag. - FACTUAL CHECKS - claims worth verifying. You cannot verify them; list them so I can. Rules: - Do not suggest stylistic improvements. Errors only. - If something is correct in my stated variety of English but would be wrong in another, do not flag it. - Do not rewrite sentences. Point at the error.
What you get: An error table with locations and corrections, plus separate lists of possible-intentional items and claims to fact-check.
Tip: Separating 'errors' from 'possible errors' is what makes the output usable. A proofread that flags your deliberate sentence fragments as mistakes gets ignored wholesale.
Give feedback on someone else's writing
Review a colleague's draft usefully and kindly.
Help me give feedback on this draft. DRAFT: """ [PASTE] """ WHO WROTE IT: [their role, their experience level, my relationship to them] WHAT THEY ASKED FOR: [copy edit / structural feedback / 'what do you think' / approval] WHAT IT IS FOR: [PURPOSE AND AUDIENCE] HOW MUCH CAN CHANGE: [it ships tomorrow / early draft / anything is open] MY HONEST REACTION: [WHAT YOU ACTUALLY THINK] Produce: 1. WHAT IS GENUINELY GOOD - specific, not encouragement. Quote it and say what works. If there is nothing specific to praise, say so to me rather than manufacturing something; false praise makes the real feedback unreadable. 2. THE ONE THING - if they change only one thing, what. Every review should have a clear priority. 3. FEEDBACK BY LEVEL - grouped, because mixing them is what makes reviews overwhelming: - ARGUMENT/PURPOSE: does it do its job? - STRUCTURE: is it in the right order? - CLARITY: will the reader follow it? - LINE LEVEL: only if they asked for it, and only patterns rather than every instance 4. CALIBRATED TO WHAT THEY ASKED - if they asked for a copy edit, do not deliver a structural rewrite. If it ships tomorrow, do not propose restructuring. Say plainly if what they asked for is not what the draft needs, and let them decide. 5. THE WRITTEN FEEDBACK - as I would send it. Direct, specific, no compliment sandwich. Frame each point as an observation about the reader's likely experience rather than a verdict on the writing. 6. WHAT I SHOULD NOT SAY - things in my honest reaction that are taste rather than quality. Separate them out; imposing your own style preferences is the most common way review feedback goes wrong. 7. WHAT TO ASK RATHER THAN TELL - places where they may have a reason I cannot see.
What you get: Level-grouped feedback calibrated to what they asked for, with your taste preferences separated out from real problems.
Tip: Point 6 is the discipline. Half of most editorial feedback is the reviewer's own style preference wearing the costume of a correction.
Make technical writing readable without dumbing it down
Improve clarity for a mixed-expertise audience.
Make this clearer without losing precision. TEXT: """ [PASTE] """ READER: [WHO, AND WHAT THEY ALREADY KNOW] WHAT THEY NEED TO DO WITH THIS: [DECIDE / IMPLEMENT / UNDERSTAND / APPROVE] WHAT MUST STAY PRECISE: [TERMS OR CLAIMS THAT CANNOT BE SIMPLIFIED] Produce: 1. THE COMPREHENSION BLOCKERS - quote every place a reader at the stated level would stop. Categorise each: undefined jargon, assumed prior knowledge, a sentence carrying too many ideas, an abstraction with no example, or a reference to something not yet introduced. 2. TERM AUDIT - every technical term. For each: essential (keep and define on first use) / replaceable (give the plain equivalent) / unnecessary (cut). Do not remove a term the reader needs to know for their job - teaching it is better than avoiding it. 3. THE STRUCTURE PROBLEM - technical writing is usually unclear because it explains in the order the author understands it rather than the order the reader needs it. Check specifically: is the conclusion at the end when it should be at the start, and is context given before it is needed or after? 4. THE REWRITE - clearer version. For each change, precision must be preserved. Where making something simpler would make it wrong, keep it and add an explanation rather than simplifying. 5. WHAT I ADDED - examples, analogies or definitions inserted. Flag any analogy that is imperfect and say where it breaks down; an analogy that misleads is worse than none. 6. WHAT I REFUSED TO SIMPLIFY - and why. Some things are irreducibly complex and pretending otherwise is a disservice. 7. THE ONE-PARAGRAPH VERSION - for the reader who will only read the top. Never trade accuracy for readability. If a choice is forced, keep accuracy and flag it.
What you get: Quoted comprehension blockers, a term-by-term audit, a rewrite that preserves precision, and an explicit refused-to-simplify list.
Tip: Point 3 is the real fix. Most technical writing is unclear because it is ordered by how the author learned it rather than by what the reader needs first.
Adapt one piece for a different audience
Rework existing writing for a different reader without starting over.
Adapt this for a different audience. ORIGINAL: """ [PASTE] """ WRITTEN FOR: [ORIGINAL AUDIENCE] ADAPTING FOR: [NEW AUDIENCE - WHO THEY ARE, WHAT THEY KNOW, WHAT THEY CARE ABOUT] NEW PURPOSE: [WHAT THE NEW AUDIENCE SHOULD DO OR UNDERSTAND] NEW FORMAT AND LENGTH: [WHERE IT GOES AND HOW LONG] Produce: 1. THE GAP ANALYSIS - what differs between the two audiences: prior knowledge, what they care about, what motivates them, their objections, what they can act on, and how much time they will give it. 2. WHAT CHANGES AND WHY: - The angle: the same facts, framed around what the new audience cares about. This is usually the biggest change. - Level of detail: what to add, what to cut - Vocabulary: terms to define, terms to replace, terms the new audience uses instead - Examples: which no longer land, and what to replace them with - The ask: what the new audience can actually do 3. WHAT STAYS - the substance that holds across both. 4. THE ADAPTED VERSION. 5. WHAT I COULD NOT CARRY ACROSS - things central to the original that do not work for the new audience, and whether that matters. 6. THE HONEST CHECK - is this an adaptation or does the new audience need a different piece entirely? Say which. Sometimes the answer is that only the underlying facts transfer, and repurposing costs more than writing fresh. Do not just simplify or lengthen. Adaptation is about what the audience cares about, not their reading level.
What you get: A gap analysis, a reframed adaptation, and an honest check on whether adaptation is the right approach at all.
Tip: Point 2's angle change is the part people miss. Adapting for executives is not about shorter words; it is about leading with cost and risk instead of method.
Build a style guide from existing writing
Document how your organisation actually writes, so others can match it.
Build a style guide from these samples. SAMPLES OF OUR WRITING: """ [PASTE 3-5 PIECES YOU CONSIDER GOOD EXAMPLES] """ WHO WILL USE THIS GUIDE: [WRITERS, THEIR EXPERIENCE] WHAT WE PUBLISH: [CONTENT TYPES] KNOWN INCONSISTENCIES THAT ANNOY US: [IF ANY] Produce a practical style guide, derived from the samples rather than from general best practice: 1. VOICE - three to five adjectives, each with a quote from the samples demonstrating it, and a counter-example of what it is not. 2. THE RULES WE ALREADY FOLLOW - inferred from the samples: person (first/second/third), contractions, sentence length range, paragraph length, formality, how we address the reader, humour and how much. 3. MECHANICS - a decision on each, matching the samples: variety of English, serial comma, numerals vs words, capitalisation of headings, em-dash spacing, how we write dates, times, currency and percentages, how we introduce acronyms. 4. WORDS WE USE AND DO NOT USE - our terms for our own concepts, banned jargon, and words that appear in the samples characteristically. 5. STRUCTURE PATTERNS - how our pieces typically open, how headings are phrased (questions, statements, noun phrases), typical length by content type, how we handle lists. 6. INCONSISTENCIES FOUND - where the samples disagree with each other. For each: the options, and a recommendation. These are the decisions the guide needs to make. 7. THE ONE-PAGE VERSION - the ten rules that matter most. A 40-page style guide is not read; a one-pager is. 8. EXAMPLES - three before/after pairs showing off-guide and on-guide writing. Derive everything from the samples. Where the samples are silent, say so and propose rather than asserting.
What you get: A style guide derived from your own writing, with inconsistencies surfaced as decisions and a one-page version.
Tip: Point 7 is what determines whether the guide gets used. Nobody reads the full document; the one-pager is the actual deliverable.
Where AI actually helps here
- Cutting length without losing content — ask for 20% and it finds it
- Catching the sentence that says nothing, the paragraph that repeats the one above
- Consistency: spellings, capitalisation, terminology, tense
Where it falls down
- Preserving voice. Left alone it edits everything towards the same neutral middle
- Knowing what is deliberate. Your fragment for emphasis becomes a complete sentence
- Fact-checking, which it will do confidently and wrongly if you ask
The mistake almost everyone makes: Asking it to edit everything at once
Structure, line editing and proofreading are three different jobs and doing them in one pass means each is done badly. Run them as separate prompts in order, and finish with: show me what you changed and why, as a list. That list is where you catch the edits that removed something on purpose.
Free tool: Output Grader
Runs in your browser. No sign-up, nothing uploaded.
Questions people ask
Can AI proofread better than a human?
It is more consistent at mechanical errors and worse at knowing which rule to break. Use it for the sweep, then read it yourself for the things that were choices.
How do I stop AI flattening my voice?
Constrain it: “fix only grammar, spelling and punctuation — do not rewrite sentences, do not change word choice”. Unbounded “improve this” always means “make this sound like everything else”.