Self-Editing: How to Edit Your Own and AI's Drafts

Self editing drafts is hard for a specific reason. You already know what you meant to say, so you read that instead of what's actually on the page. Your brain fills gaps, supplies missing logic, and hears the tone you intended even when the sentence doesn't carry it. This isn't a discipline problem. It's a perception problem, and it applies whether you wrote the draft yourself an hour ago or a model wrote it and you've now read it four times.

A serviceable first draft isn't the hard part anymore. Anyone can produce one, and a model can produce one in seconds. The work that separates a published piece from a pile of words is the edit, and most people are their own editor most of the time. Getting good at switching roles on purpose, from the person who wrote the thing to the person whose job is to find what's wrong with it, is the skill that actually matters here.

Why the page feels done too soon

A draft feels finished when you can read it without friction, and you can always read your own draft without friction because you're not really reading it. You're recalling it. The words on the page are a prompt for a memory, and the memory is smoother than the text underneath.

This is why a paragraph that repeats an argument made two paragraphs earlier can survive five read-throughs untouched. You're confirming that yes, this is the point you meant to make, and moving on. The same blind spot shows up with an AI draft, just on a different timeline. The first read, you're genuinely evaluating it. By the third pass, the phrasing has gone familiar, and familiarity gets mistaken for correctness. A sentence that sounds confident but is quietly wrong slides past, because confidence and accuracy read the same on a screen.

OpenAI's own research found that roughly two-thirds of ChatGPT writing requests are asks to modify existing text, not generate something new. source People aren't mostly asking a model to write from nothing. They're asking it to fix, tighten, or rephrase something that already exists, which means the editing pass is where most of the actual work already happens. Knowing you need to edit and actually seeing what needs fixing are two different skills, though. You can sit down intending to edit rigorously and still miss the same three problems you missed last time, because intention doesn't fix perception.

Switch roles before you start fixing

The fix is a stance change, not a technique. Stop reading as the writer and start reading as a skeptical editor who didn't write this and isn't obligated to like it.

The writer is generative and a little forgiving, willing to let a rough sentence stand so the next one can get written. The editor is detached, and asks one question of every sentence: does this serve the reader, or does it serve the person who wrote it? Padding, throat-clearing openers, the paragraph that exists because it was fun to write rather than because it's needed, almost all of it fails that test, and the editor's job is to cut it without negotiating for its survival.

Imagine the draft was written by someone whose work you respect but whose every claim you're paid to check. That framing gives you permission to be hard on the sentences without it feeling like an attack on yourself. It's also the right posture for an AI draft, where the prose often sounds more certain than the facts underneath it actually are.

Since you can't just decide to see your draft differently, change the conditions under which you read it. Let it sit for fifteen minutes at minimum, overnight if you can spare it. Change the format too, a different font, a printout, a phone screen instead of the editor you wrote it in, since your eye has memorised the current layout and a new one breaks the pattern. Read it aloud, probably the single highest-leverage reset, because your ear catches clumsy rhythm and repetition your eye slides straight past. For pure proofreading, read backward, last sentence to first, which strips out meaning and momentum so spelling and mechanics show up on their own.

For an AI draft, time and reformatting matter less, since you have no memory of writing it. Reading aloud and verifying the claims still carry the load, arguably more of it, because a fluent AI sentence is very good at sounding right whether or not it is. For a quick way to spot the patterns that give a machine-written passage away before you even start editing, the repeated rhythm and stock scaffolds are covered in what marks a passage as AI-written.

Self-editing sits next to, but isn't quite, copyediting, line editing, or developmental editing. Those are usually separate passes done by a second person against a fixed brief. Self-editing is the version you do alone, on your own or the model's output, before any formal QA ever sees the page.

Edit in passes

Editing goes wrong fastest when you try to fix everything in one read. You polish a sentence for elegance, then delete the whole paragraph it was sitting in, and the polish was wasted effort. Editing in passes, from biggest concern to smallest, fixes that. Ignore anything in a pass that isn't that pass's job.

Structure comes first. Does the piece answer the reader's actual question early, and does every section earn its place, or is it organized the way you happened to write it rather than the way someone needs to read it? Leading with the answer isn't just a courtesy to the reader. It's also what makes a piece easier for AI systems to extract and cite, which is the discipline covered in generative engine optimization.

Argument comes next, and this is where self-editing pulls in opposite directions depending on who wrote the draft. On your own work, you over-trust the logic, because you know what you meant and the connective tissue feels present even when it isn't on the page. On an AI draft, you under-scrutinize the claims, because the tone sounds authoritative and authoritative-sounding prose doesn't invite suspicion the way it should. Verify rather than assume, a discipline covered in more depth in fact-checking AI content.

Line editing comes third, once you know the paragraph is staying. Say a sentence reads: "In terms of the overall approach, it's important to note that the strategy we're proposing here will likely result in improvements to efficiency over time." A structural and line edit together turns that into: "The strategy should make the team faster." Same claim, a third of the words, and the hedging that was doing no work is gone.

Proofreading comes last, because there's no point fixing a comma in a paragraph that structural editing might cut entirely. This is where reading backward finally pays off, since at that stage you're checking mechanics, not meaning. Microsoft frames a version of this same order for reviewing Copilot output, working through goal, audience, constraints, evidence, and structure before touching the sentence level. source Tighten the brief before you tighten the sentence, whether the draft came from a person or a model.

Say a team drafts a landing page with a model, loves a clever line in paragraph three, and spends twenty minutes wordsmithing it before realizing the whole section argues a point the page doesn't need. That twenty minutes is gone. Passes exist so it doesn't happen twice.

The last review

The final pass isn't a vibe check. It's a specific set of questions asked in order, sitting on top of a short version control and sourcing habit.

Did the draft get time to rest, or at least a format change, before you started editing it? Have you actually switched into the skeptical-editor role, or are you still reading as the person who wrote it? Does the structure answer the core question up front, with every section pulling its weight? Is the padding gone, the hedging cut, and does the voice sound like something a specific person wrote rather than the average of everything the internet has said on the topic? Did you read the whole thing aloud at least once? Is it clean on spelling, links, and formatting as the very last step, not the first?

Keep the draft you started from, even a plain copy pasted into a separate file, so a structural cut is a decision you can reverse rather than a change you have to remember and rebuild from scratch. For any claim that isn't your own firsthand opinion, log where it came from as you edit, not after the fact. A quote pulled from memory during a fast proofreading pass is exactly the kind of thing that turns out to be misremembered later.

Your own draft usually has a voice already, so the risk is over-trusting the logic and under-cutting the sentimental bits you're attached to. An AI draft usually has no wrong facts you're aware of and no voice you'd recognize as yours, so verification and voice repair carry almost all the weight. OpenAI's own guidance on this is specific: treat the output as a draft to review, not a final authority, and revise it with concrete instructions rather than accepting or rejecting it wholesale. source Its prompt guidance pushes the same idea further upstream, recommending you review a response and adjust the prompt itself before you even touch the draft. source Microsoft Word's Copilot suggestions work on a related principle, surfacing changes to clarity, structure, and tone as suggestions you accept or reject individually rather than edits applied automatically, which keeps a human deciding what actually ships. source

A Harvard Kennedy School study found that people who practiced writing with AI assistance improved more on a writing test a day later than people who practiced without it, but they didn't outperform a group simply shown an AI-generated draft they weren't allowed to edit. source The gain came from exposure and feedback, not from handing the whole job to a model and walking away.

If you want a fuller checklist to run against every piece before it ships, an editing checklist for AI drafts covers the same ground in more granular form, and the broader case for why editing now carries most of the value in the process is laid out in editing AI content.