TL;DR
Paste your full draft into Claude or ChatGPT with a critique prompt that bans rewriting, demands line-quoted issues ranked by impact, and names your audience and goal. Claude Opus 4.7 and Sonnet 4.6 keep your voice best on long pieces; GPT-5.5 with Canvas is the best place to act on edits inline. Accept roughly 60% of the suggestions, reject the rest, and never let the model rewrite the piece for you, that is where your voice goes to die.
The task
You have written a draft (essay, blog post, cover letter, application) and you want sharp feedback before you publish or submit. You do not want a friend to be polite. You do not want AI to rewrite the whole thing in its own voice. You want what a senior editor would say in a 15-minute review: what is working, what is broken, what to do about it.
When AI is the right tool
- You need feedback faster than a human reviewer can give.
- You have already done two self-edits and stopped seeing your own mistakes.
- You want a second perspective on logic, structure, and pacing.
- You are practicing writing and want repeated, low-cost reps.
When not to rely on AI alone
- For factual claims that are specialist-level (legal, medical, technical research). AI will not catch domain errors a real expert will.
- For high-stakes pieces (book proposals, college essays you only submit once). Use AI for one pass, then a human editor.
- For voice. AI is bad at telling whether a piece sounds like you. A human who knows you matters more here.
Which model to use (June 2026)
The big three frontier models all give usable critique now. The differences are about voice preservation, how much you can paste, and where you edit. Figures below are as of June 2026.
| Tool (tier) | Best for feedback | Context you can paste | Price |
|---|---|---|---|
| Claude (Pro $20/mo) | Voice-preserving line edits, long drafts | 1M tokens on Sonnet 4.6 / Opus 4.7 | Free tier limited; Pro is $17/mo billed annually |
| ChatGPT (Plus $20/mo) | Acting on edits inline via Canvas | ~320 pages in-app on Plus (full 1M only on $200 Pro) | Free tier has tight limits and ads (US, since Feb 2026) |
| Gemini (Google AI Pro $19.99/mo) | Very long documents plus Google Docs | 1M tokens on Gemini 3.1 Pro | Free tier available; Pro renamed from “Gemini Advanced” early 2026 |
Practical rule: draft and critique in Claude (it stays in an editor persona and won’t flatten your phrasing), then if you want to revise sentence by sentence, paste into ChatGPT Canvas for inline comments and version history. For a 90,000-word manuscript or a long thesis, paste it whole into Gemini 3.1 Pro or Claude, both hold 1M tokens.
What to feed the AI
- The full draft (not a summary).
- Audience: who reads this, and what do they already know.
- Goal: what should the reader do or feel by the end.
- Genre: essay, op-ed, blog post, application, etc.
- Constraints: word limit, tone, format requirements.
- The rubric, if one exists. If a professor or editor grades against criteria, paste them verbatim and ask the model to score each one. This is the single biggest jump in feedback quality.
Copy-ready prompt
Replace each bracketed placeholder with your own text.
Critique my draft as a senior editor doing a 15-minute review.
Audience: [who reads this and what they already know]
Goal: [what readers should do or feel by the end]
Genre: [essay / blog / cover letter / ...]
Constraints: [word limit, tone, format]
Draft:
"""
[paste full draft]
"""
Output:
1. 3 specific strengths, each quoting the line it refers to.
2. 5 specific issues, ranked by impact. For each: quote the line,
name the problem (logic gap, weak verb, buried lede, unclear
claim, etc.), and give one concrete fix.
3. 3 revision priorities for my next pass.
Rules: Do NOT rewrite the draft. Do NOT pad. Ground every note in
reader response, not abstract writing rules. Preserve my voice and
unusual word choices unless they obscure meaning.If you have a grading rubric, add this block:
Rubric (score each criterion 1-5 with the exact line that justifies
the score, then one fix to move up a level):
"""
[paste rubric criteria]
"""Recommended output structure
The reply should be skimmable in 60 seconds: a numbered strengths list, a numbered issues list with line quotes, and a short revision plan. Anything longer than 500 words is usually the model padding. Claude in particular will not pad unless asked, so if the response balloons, your prompt let it.
How to check the output
- Are the issues specific enough to act on? “Weak intro” is useless; “the first sentence is a definition, move the example up” is useful.
- Do the suggested fixes match your goal? Reject suggestions that drift the piece toward generic best-practice writing.
- Spot-check at least one quoted line against your draft. If the model misquotes or invents a line that is not there, treat the rest with suspicion, that is a hallucination signal.
Work in passes, not all at once
Editors do not fix everything in one read, and neither should you. Run the model once per layer:
- Structure pass: argument, order, and what to cut. Ask only about the skeleton.
- Paragraph pass: pacing, transitions, where attention drops.
- Sentence pass: weak verbs, overused words, tone slips.
- Proofread pass: grammar and typos last, after the shape is locked.
Asking for all four at once buries the structural notes (the ones that matter most) under a pile of comma fixes.
Common mistakes
- Asking AI to rewrite. You learn nothing and the piece loses your voice.
- Skipping audience and goal. Without them, feedback defaults to generic style notes.
- Taking every suggestion. A good edit pass might accept 60% and reject 40%.
- Mixing structural and proofreading feedback in one prompt, so the big fixes get lost.
Run a second round
After your revision, run the same prompt again with “second-round critique, assume I already fixed the obvious stuff.” Compare the two critiques. If the new issues are real and deeper than the first pass, you have found the next layer worth working through. If they look like padding the model invented to have something to say, you are done, stop revising and ship.
FAQ
Which AI gives the most honest writing feedback? Claude (Opus 4.7 or Sonnet 4.6) tends to give the sharpest editorial notes while preserving your voice, partly because it will not write at length or soften unless told to. GPT-5.5 is close and its Canvas mode is better for acting on edits inline.
Will the AI tell me if my piece is boring? Usually yes, if you ask explicitly. Add “flag any section where attention is likely to drop and say why.”
How do I stop the AI from flattening my voice? Two things: never ask it to rewrite, and add “preserve idiosyncratic phrasing and unusual word choices unless they obscure meaning.” Feedback teaches you to fix it; a rewrite replaces you.
Can I use this for fiction? Yes, but name the genre conventions you want feedback against (literary, thriller, romance). Otherwise the model defaults to generic prose advice that fits no genre well.
Is it cheating to get AI feedback on a school essay? Getting feedback is generally fine; having AI write or rewrite the essay is not. The safe line, echoed by university writing guides, is to use AI to identify problems and explain fixes, then make every change yourself. Check your institution’s policy.
Do I need a paid plan? No. Free Claude and free ChatGPT both critique drafts, with tighter usage limits. Paid plans ($20/mo as of June 2026) matter mainly if you run many long drafts or want ChatGPT Canvas and bigger context.
Related
- AI article rewrite
- AI blog post outline
- AI cleans up study notes
- Academic Debate Prompts for Cases and Refutation
External references: Northeastern: Using AI for Writing Feedback and Anna Mills’s PAIRR feedback prompts are solid, classroom-tested starting points.