How to Use AI as an Editor (Not a Writer)

TL;DR: Ask AI to write and you get competent, generic prose with your thinking removed. Ask it to edit a draft you already wrote and it keeps your thinking while finding your blind spots. Below: nine separate editing passes, as copy-paste prompts, plus the one instruction that keeps every pass from rewriting your voice.

What actually goes wrong when you ask AI to write for you?

Hand a model a blank page and a topic, and it does the only thing it can: produce the statistically likely version of that topic, in its own default register. There is no version of your specific reasoning in there, no trace of the particular evidence you'd reach for, no sign of the thing you actually noticed that made you want to write this in the first place. It's competent. It's also nobody's writing in particular, which is most of what "generic" means. We cover the same failure mode from the reader's side separately: a broad prompt gets a broad answer, whether the output is an answer or a draft.

Hand the same model a paragraph you already wrote and ask it to edit that paragraph, and the job changes completely. It isn't generating your argument anymore. It's checking one you already made: is it sound, is it clear, is it missing something, would a skeptic buy it. That's a narrower, more specific task, and it happens to be one models are unusually good at, because spotting a hole in an existing structure takes less invention than building the structure in the first place.

The gain here isn't that editing is easier to prompt for. It's that editing is the version of this task where your thinking survives contact with the model.

What's the one instruction that makes every editing prompt actually work?

Every template below does one specific job. Before any of them, add one more line, because it's the difference between an editor and a ghostwriter with opinions: name what the model has to leave alone.

Left unconstrained, a model asked to "improve this" will rewrite it, and it will rewrite toward its own default: shorter sentences than you'd choose, tidier transitions than you'd write, the same handful of connective phrases it reaches for on every task. None of that is wrong on its own. It's just not yours, and multiplied across every sentence in a piece, the cumulative effect is a draft that reads correctly and sounds like nobody.

So say it directly: keep my sentence rhythm, don't replace my word choices, flag the problem instead of fixing it unless I ask. That last one is the single highest-value pattern in this whole piece. A flag-and-explain instruction leaves the decision with you and tells you why something is weak, instead of quietly substituting the model's version and hoping you don't notice the swap. You learn something from a flag. You learn nothing from a silent rewrite you didn't ask for.

This isn't the vague kind of prohibition that backfires. We've written separately about why "don't do X" often under-specifies: "don't be verbose" leaves the target open, because every length except one is still on the table. "Leave this sentence exactly as I wrote it" doesn't have that problem. The excluded set is one specific, checkable state, which is exactly the case that instruction names as the exception that holds up.

How do you prompt for a developmental edit?

A developmental edit asks whether the piece should exist in its current form at all: is the argument sound, what's missing, what's asserted without support. Run this pass before any sentence-level pass, because polishing a paragraph that shouldn't be there wastes the polish.

DE-1 · Claim-to-support map
Read the draft below. For each claim you find, note whether it
is directly supported in the text, only asserted, or contradicted
elsewhere in the piece. Do not rewrite anything. Output a list:
CLAIM / SUPPORTED, ASSERTED, or CONTRADICTED / WHERE.

[PASTE DRAFT]

DE-2 · The missing piece
A well-informed reader who does not already agree with me is
reading this. What would they need to see before accepting the
conclusion that the draft does not currently give them? List up
to five gaps. Do not suggest wording. Name the gap only.

[PASTE DRAFT]

DE-3 · Unsupported-assertion flag
Quote every sentence that states a fact, a number, or a causal
claim without a source or without evidence earlier in the piece.
Do not judge whether the claim is true. Just list what would
need a source if a skeptical editor asked for one.

[PASTE DRAFT]

How do you prompt for a structural edit?

Structure is order and inclusion: what comes first, what gets cut, what deserves more room than it currently has. This pass moves paragraphs and sections. It shouldn't touch a single sentence's wording.

SE-1 · Section value ranking
List every section heading in the order it currently appears.
Next to each, note in three words or fewer whether it is doing
real work, could be cut, or could be shorter. Do not reorder or
rewrite anything yet. Ranking only.

[PASTE DRAFT WITH HEADINGS]

SE-2 · Find the buried lede
What is the single strongest point in this draft? Quote the
sentence that makes it. Then tell me where in the draft that
point currently sits, and whether moving it earlier would help a
reader who might stop after the first three paragraphs.

[PASTE DRAFT]

SE-3 · Redundancy across sections
Identify any point that is made more than once across different
sections. For each repeat, name both locations and say which one
is the stronger version. Do not merge or cut anything. List only.

[PASTE DRAFT]

How do you get a line edit without losing your voice?

Line editing is rhythm and clarity at the sentence level, and it's the pass most likely to overwrite your voice if you don't constrain it, because "better" and "more like the model's house style" look identical from inside a rewrite. Ask for a map of problems instead of a replacement draft.

LE-1 · Long-sentence flag
Flag every sentence over 30 words. For each, note in one line
why it's hard to follow, if it is. Do not shorten or rewrite the
sentence yourself. I will decide whether and how to split it.

[PASTE DRAFT]

LE-2 · One unclear line per paragraph
For each paragraph, name the single least clear sentence, and
explain in one line what makes it unclear: an unclear referent,
a buried verb, a clause doing two jobs. Do not offer a rewrite.

[PASTE DRAFT]

LE-3 · Rhythm check, annotate only
Read this the way it would sound spoken aloud. Mark any place
where three or more consecutive sentences share the same length
and structure, which reads as monotonous. Note the location only.
Do not change the sentences. I will vary them myself.

[PASTE DRAFT]

What does a copy edit pass actually catch that the others don't?

Copy editing is mechanical: grammar, punctuation, consistency of terms and capitalization, the kind of error with one correct answer and no judgment call attached. This is the one pass where letting the model actually fix things, rather than only flag them, is reasonable, because there's nothing of your voice at stake in a comma splice.

CE-1 · Consistency-only list
List every place a term, name, or spelling is used
inconsistently across this draft (for example, both "email" and
"e-mail", or a product name capitalized two different ways).
Give the two variants and where each appears. Do not pick a
winner or fix anything. List only.

[PASTE DRAFT]

CE-2 · Grammar and punctuation, nothing else
Fix grammar, spelling, and punctuation errors only. Do not
change word choice, sentence order, sentence length, or tone.
If you are not sure whether something is an error or a
stylistic choice, leave it and list it separately instead of
changing it.

[PASTE DRAFT]

How do you prompt for a hostile read?

A hostile read asks the model to argue against you, which works with the grain of a specific, documented behavior rather than against it. OpenAI's Model Spec instructs its models that when someone asks for critique of their work, the assistant should "provide constructive feedback and behave more like a firm sounding board that users can bounce ideas off of — rather than a sponge that doles out praise" (model-spec.openai.com, version 2026-08-18, accessed September 2, 2026). Asking for a hostile read is asking the model to run that mode on purpose instead of hoping it defaults there on its own.

This is persona prompting aimed at your own draft instead of at a question, and it's a gentler cousin of a technique we've covered for testing an AI's answers rather than your own prose.

HR-1 · Steelman the skeptic
You disagree with this piece's main conclusion. Build the
strongest honest case against it, using only real objections a
knowledgeable reader could raise. If the case against is weak,
say so rather than manufacturing objections.

[PASTE DRAFT]

HR-2 · The weakest link
If a critic wanted to dismiss this entire piece by attacking one
sentence, which sentence would they pick, and what would they
say about it? Name one sentence, quote it, and give their
argument in two sentences.

[PASTE DRAFT]

HR-3 · Reviewer with nothing to lose
You are reviewing this for a publication that will be embarrassed
if it publishes something wrong. List every claim asserted
without support, every number without a source, and every
sentence that needs a caveat it doesn't have. Do not rewrite
anything. List only.

[PASTE DRAFT]

Can AI actually fact-check your writing?

Not in the sense of confirming that something is true. A model has no independent access to reality. It has patterns in training data and whatever it can infer from your text, and both can be wrong in ways that sound completely confident. Anthropic's own guidance on reducing its models' hallucinations is direct about the ceiling here: "while these techniques significantly reduce hallucinations, they don't eliminate them entirely. Always validate critical information, especially for high-stakes decisions" (platform.claude.com, accessed September 2, 2026). If that's the honest limit of Anthropic's own mitigation advice for its own model, it's the honest limit of a quick prompt run over your blog post too.

So don't ask for a fact check. Ask for a claim audit: which sentences assert something a reader could check, and where you'd go to check it. The model surfaces the list. You do the checking. That reframing is the entire difference between a useful pass and a false sense of security.

FA-1 · Claim ledger
List every sentence in this draft that states a number, a date,
a study result, or a named fact. For each, write NEEDS SOURCE if
none is given in the text, or ALREADY SOURCED if it is. Do not
tell me whether the claim is true. I will check each one myself.

[PASTE DRAFT]

FA-2 · Confidence tags
Tag every factual sentence as [MY OWN OBSERVATION], [WIDELY
KNOWN], or [SPECIFIC CLAIM NEEDING A CITATION]. Do not change
the wording. Tags only, inline or as a list referencing each
sentence.

[PASTE DRAFT]

FA-3 · Where would this be wrong
Name the three claims in this draft I would be least comfortable
defending if a reader pushed back, and say specifically what
evidence would settle whether each is correct. Do not attempt to
verify them yourself.

[PASTE DRAFT]

How do you find what your reader doesn't already know?

An audience check asks a different question than any pass so far: not whether the draft is correct or clear, but whether it's teaching the reader it's aimed at anything at all. A piece that only restates what its audience already believes has no reason to exist.

AC-1 · Assumed-knowledge audit
List every term, acronym, or concept this draft uses without
defining, for [DESCRIBE YOUR READER, e.g. "a marketing manager
who has used ChatGPT casually but never built a system prompt"].
For each, say whether that reader would already know it or need
one clause of explanation.

[PASTE DRAFT]

AC-2 · The nothing-new test
Read this as [DESCRIBE YOUR READER]. For each section, say
whether that reader would learn something they didn't already
believe, or whether the section only confirms what they'd
already assume. Name the sections in the second category.

[PASTE DRAFT]

How do you cut to a word count without gutting the argument?

Cutting for length is a different skill from cutting for quality, and confusing the two is how a tight edit turns into a thin one. Get a list of what could go before you delete anything. (If the model is already handing you clipped answers before you've asked it to cut anything, that's a separate, opposite problem.)

CL-1 · Removability ranking
This draft is [CURRENT WORD COUNT] words and needs to be
[TARGET WORD COUNT]. List every sentence or short passage that
could be removed with the least damage to the argument, ranked
from safest cut to riskiest. Do not delete anything. I'll decide
which cuts to make.

[PASTE DRAFT]

CL-2 · Per-section budget
Divide [TARGET WORD COUNT] evenly across this draft's [NUMBER]
sections. Tell me which sections currently run over that share
and by how much. Do not shorten anything.

[PASTE DRAFT WITH SECTION WORD COUNTS]

How do you fix a title and opening without flattening them?

Titles and openings are the one place where letting a model draft real candidates, rather than only flagging problems, earns its keep, because you're choosing between options instead of defending your own sentence. Do this pass last, after the piece has stopped changing shape.

TO-1 · Opening line audit
Read only the first two sentences of this draft. Do they tell a
reader what they're about to learn, specifically enough that
they could repeat it back? If not, say exactly what's missing.
Do not rewrite yet.

[PASTE FIRST TWO SENTENCES]

TO-2 · Five titles, one constraint held
Write five alternative titles for this piece. Each must keep
[YOUR TARGET KEYWORD OR CORE CLAIM] intact. Rank them by which
is clearest about what the reader gets, not which sounds most
exciting. Do not touch the body.

[PASTE CURRENT TITLE AND ONE-SENTENCE SUMMARY]

TO-3 · Opening rewrite, offered not applied
Draft one alternative opening paragraph that states the piece's
main point in the first two sentences. Label it clearly as a
suggestion. Do not replace my existing opening; give me both
side by side so I can compare.

[PASTE DRAFT]

If you'd rather run a first pass across all nine at once and dig into any one of them afterward, one message does it:

RUN-1 · All nine passes in one message, for a fast first look
Read the draft below and run all nine of these separately, each
under its own heading, without skipping any:
1. DEVELOPMENTAL: claim-to-support map, quoting sentences only.
2. STRUCTURAL: which sections do real work vs. could be cut.
3. LINE: flag long or unclear sentences. Do not rewrite.
4. COPY: grammar and punctuation fixes only.
5. HOSTILE READ: the strongest case against my conclusion.
6. CLAIM AUDIT: which sentences need a source. Do not verify them.
7. AUDIENCE: what will [READER] not already know here.
8. LENGTH: what could be cut to reach [TARGET WORD COUNT].
9. TITLE/OPENING: do the first two sentences deliver the point?

Do not rewrite my sentences at any step except step 4. Flag,
don't fix, everywhere else.

[PASTE DRAFT]

Is there a real risk to running all of this on the same draft?

Yes, and it's a different risk than the model getting a fact wrong: homogenization. Every one of the nine passes above is trained on a huge amount of writing, and every suggestion it makes is pulled from patterns common across that writing. Accept every suggestion, every time, on every draft, and your prose stops sounding like you and starts sounding like the statistical average the model keeps returning to, because that average is what every suggestion is quietly built from.

The fix isn't to skip the passes. It's to treat each one as a second opinion you evaluate, not an edit you apply by default. Read what a pass flags, decide which points are real and which aren't, and reject the rest deliberately rather than by habit. A flagged sentence you disagree with and keep unchanged isn't a failure of the process. It's the process working, because the point was never to defer to the model. It was to see your own blind spots faster than you'd see them alone.

The writing is still yours to be responsible for, claim by claim, sentence by sentence, after every pass above has had its say.

Do you need Prompt Architects to run any of this?

No. Every prompt above runs in any chat window, on any plan, and costs nothing beyond the time it takes to read what comes back. That's deliberate. The value in this piece is the method, not a feature gate.

What Prompt Architects is actually built for is narrower than "editing your writing." It's a tool for building and storing prompts, including editing prompts like the ones above, so you're not retyping a half-remembered version of one every time you start a new draft. Our own free plan includes five prompt enhancements a day, forever, no card required, per our FAQ page. A tagged, searchable library to keep the templates you actually reuse starts on the Pro plan; the free plan keeps a running history instead of a curated one. Neither is the point of this piece, and neither edits a single word of your draft for you. The general prompt library is a separate, larger resource if editing prompts are only part of what you're after.