The brief is open in one tab, an empty doc in the other. You paste the topic into a chat box, get back five paragraphs that could sit under anyone’s byline, then spend an hour rewriting what you asked for to save time.

The common writing challenges AI tools solve are not really solved by the tool. They are solved by what you hand it.

The speed gain is real. Noy and Zhang, publishing in Science in July 2023, gave 453 professionals real writing tasks of 20 to 30 minutes each and had the results scored blind. ChatGPT access cut completion time by 40 percent and raised quality scores by 18 percent.

One caveat from the authors matters here: those tasks did not require precise factual accuracy about real companies or customers. Put fact-checking back in and the gain shrinks.

What changed since then is capacity. Context windows are now measured in hundreds of thousands of words, models reason before answering, and many search the live web. Hand one a full brief and the source material it came from, and the first draft stops being the interesting part.

Each section below gives you the same four things: what to hand the model, a prompt you can paste, the thing it will still get wrong, and the check before you publish.

1. Writer’s Block and the Blank Page

A model will always fill a blank page. Getting it to fill the page with something only you could have written is a different job, and it starts before the prompt.

What this actually looks like

The doc is empty and the deadline is Thursday. You type “write a blog post about [topic]” and the first paragraph opens by explaining that the topic is increasingly important for modern businesses. It reads fine. It says nothing you could not have predicted, because what you asked for was the statistical average of everything ever written about that subject, and that is exactly what arrived.

What to hand it before you ask for anything

• The brief, not the topic. Your angle, your reader, and the one thing the piece has to prove.
• The raw material you already have. Call notes, a support ticket, the half-formed argument in your notebook, the numbers. Paste it in as source material instead of summarizing it from memory into the instruction.
• Three things you already believe. The claims you would make in a pub argument about this topic. This is what stops the model reaching for the average take.
• What you do not want. The openings you have already seen on this topic this month, named so it cannot hand them back to you.

None of that reaches the model unless you paste it. If you are still short of an angle at this stage, that is a different job with its own method, covered in how to generate blog topic ideas with AI.

The prompt

Verbalized Sampling (Zhang, Manning and co-authors, arXiv 2510.01171, October 2025) is a training-free prompting method that asks a model for several candidates with a probability attached to each. The paper reports roughly 1.6 to 2.1 times more diverse output on creative writing tasks with no measured quality cost.

Here is my brief: [paste the angle, the reader, and the one thing this
piece has to prove].

Here is raw material I already have: [paste call notes, customer emails,
support tickets, your own bullet points, the data].

Generate 5 different opening angles for this piece. For each one give me:
1. The angle in one sentence
2. The first two sentences as they would actually be written
3. A probability, from 0 to 1, for how likely that angle is to be the
   most common way this topic gets opened

Do not smooth them toward each other. I want five genuinely different
approaches, and I want the obvious ones labeled as obvious.

Read the lowest-probability options first. The highest-probability one is, by definition, the phrasing already flooding the training data.

What it still gets wrong

Post-training alignment pulls a model toward the most typical, most familiar-sounding phrasing available to it, a form of mode collapse. Asked for one opening, a model gives you the most expected opening it knows. No adjective in your prompt fixes that, because it is what the default behavior optimizes for.

The second failure is quieter. It will write a confident, well-formed opening for an angle you have not actually decided on yet, and a good sentence wrapped around a bad angle is much harder to spot than a bad sentence.

The check before you publish

Read your opening line and ask whether any of the pages currently ranking for this topic could publish it unchanged. If the answer is yes, it is not your opening. Fastest version: paste your first two sentences directly beneath the first two sentences of the top result.

The blank page is not the problem. The first plausible answer is.

2. Grammar, Clarity and Readability

A model will make any paragraph read better. The risk is that it also makes the paragraph say something you did not write.

The draft is clean and still wrong

You paste in a section, ask it to tighten this up, and get back something shorter and smoother in which a hedge has quietly become a claim. “Some teams report” is now “teams report”. A rough number has picked up a decimal place. Nothing looks wrong, which is the problem, because your eye is checking the prose and the prose is fine.

Hand it the source, not just the draft

• The draft and the material it came from, in the same prompt, clearly marked. Context windows are now big enough to hold the draft, the transcript, the data and your style rules at once, which was not true of the chat box most advice on this topic still describes.
• Your style rules as constraints, not adjectives. Banned words, a sentence-length ceiling, no dashes, contractions kept.
• The one thing the section has to land. A clarity edit with no target just averages the prose toward the middle. There is more on grounding a draft in real inputs in how to write AI content.

The prompt

Below is my draft and the source material it was built from.

SOURCE MATERIAL:
[paste the notes, transcript, data or research the draft came from]

DRAFT:
[paste the draft]

Edit the draft for clarity and readability only. Rules:
1. Do not add any fact, number, name, date, quote or citation that is not
   in the SOURCE MATERIAL above.
2. If a sentence is unclear because the underlying fact is missing, do not
   fix it. Flag it as [MISSING: what you would need].
3. Do not strengthen a hedge into a claim. "Often" stays "often".
4. Return the edited draft, then a change log listing every sentence you
   changed and why, one line each.

The change log is the part that pays. It converts a diff you would otherwise have to eyeball into a list you can scan in a minute.

Then edit in three separate passes rather than one combined review.

What it still gets wrong

Ask a model to support a claim and it will produce a citation with the right shape and a plausible year, whether or not the paper exists. An audit of 111 million references across 2.5 million papers on arXiv, bioRxiv, SSRN and PubMed Central put a conservative estimate of 146,932 hallucinated, non-existent citations in 2025 alone (arXiv 2605.07723, May 2026). They appeared more often in manuscripts carrying the linguistic signatures of AI-assisted writing.

It happens in newsrooms with editors. Ars Technica retracted an AI-assisted story that contained fabricated quotes, calling it a serious failure of its standards. The New York Times ran a correction after a freelance reviewer’s AI-assisted draft reproduced language and details close to a Guardian review of the same book.

The check before you publish

Trace every number, name, date and citation in the edited version back to something you handed it. Anything that appeared during the edit gets deleted or verified, with no third option. The fast version: search the edited draft for digits, then for capitalized names. On a 1,500-word piece that is a 60-second pass.

A clarity pass may change how a sentence moves. It may not change what the sentence is claiming.

3. SEO Without the Keyword Stuffing

The fastest way to make an AI draft worse for search is to tell it your target keyword.

Where “optimize this for SEO” goes wrong

You name the phrase, you ask for optimization, and it comes back in every H2, in the opening line, and once inside a sentence where it does not belong. The model took a literal instruction literally. Behind that sits the worry that brings most people here, that Google penalizes AI writing at all. It does not, and its own wording says what it targets.

Google also describes E-E-A-T (experience, expertise, authoritativeness, trustworthiness) as a set of signals its systems use, not a direct ranking factor.

Hand it the SERP, not the keyword

• The three pages currently ranking, their H2s and H3s, plus a line on what each covers. Add your own page if this is a refresh.
• The People Also Ask questions, pasted as they appear.
• The intent in one sentence. What the person typing this wants to do next.
• The keyword as context, not as an instruction. “This page needs to be findable by people searching for X” behaves differently from “optimize for X”.

The prompt

This page needs to be findable by people searching for [phrase].
Do not optimize for that phrase, do not count keyword density, and do not
repeat the phrase to hit a target.

Here are the three pages currently ranking for it:
[paste the H2s and H3s of each, plus one line on what each one covers]

Here are the questions people also ask:
[paste the PAA questions]

Do two things:
1. List the questions a searcher would still have after reading all three
   of those pages.
2. Propose an outline for a page that answers those questions, in the
   order a reader would ask them. Mark every section where I would have to
   supply something those pages cannot: my own data, a screenshot, a
   customer's words, a worked example.

Those marked sections are the only part of the page nobody else can generate at scale, which is what Google’s wording points at. The mechanics of working the phrase in afterwards sit in how to write SEO optimized content using AI tools.

What it still gets wrong

A model cannot see the SERP unless you show it. It also cannot tell a phrase that belongs in a sentence from one it has been told to hit. Stuffing is a negative signal under Google’s spam policies, so the literal instruction produces the opposite of what you asked for.

A web-connected model narrows the first half of that and not the second. It summarizes what the ranking pages say rather than noticing what none of them answers, and that gap is the only thing worth writing.

The check before you publish

Search the finished body copy for your target phrase. If it appears more than once every 200 words, it was stuffed rather than written in. Then ask the harder question: what does this page answer that the top three do not? If it has no answer, the page is scaled content with extra steps, whatever produced it.

Google’s line is not drawn between human and machine. It is drawn between pages that add something and pages that add volume.

4. Audience Targeting and Brand Voice

Adjectives do not describe your voice to a model. Pairs of sentences do.

Why “friendly and professional” produces nothing

You write “friendly, professional, conversational, authoritative” into the prompt and get back an opener with an exclamation mark and a sentence about why the topic matters right now. Those words mean different things to different people and almost nothing to a model.

Audience targeting has the same hole in it. A chat model has no access to your customers or your analytics. Competitors write that AI “analyzes demographics and behavior”. It does not, unless you paste the demographics and the behavior in.

Contrastive pairs, not a style guide

• 3 to 5 pairs, not 3 samples. Each pair is the generic version of an idea sitting next to your published version of it, and the set should span formats: a blog intro, an email opener, a social caption, so the range of the voice is represented rather than one register. Showing only good copy teaches the model to write something vaguely like this. Showing the pair teaches it to discriminate.
• Real customer language, verbatim. Two or three lines lifted from support tickets, reviews or call transcripts, in the customer’s words rather than your summary of them. That is also what makes personalized content land.
• A per-brand home for all of it. Global custom instructions are one shared slot per account and break the moment you serve a second client. Keep each voice in a project-level instruction set instead.

The prompt

Below are pairs of copy. GENERIC is how this idea usually gets written.
OURS is how we write. Match OURS. Do not average the two.

PAIR 1
GENERIC: [the flat version]
OURS: [your published line]

PAIR 2
GENERIC: [the flat version]
OURS: [your published line]

PAIR 3
GENERIC: [the flat version]
OURS: [your published line]

Here is how our customers describe this problem in their own words:
[paste 2 or 3 verbatim lines from tickets, reviews or calls]

Now write [asset] about [topic] in the OURS voice. Constraints: keep
contractions, keep first person where we use it, no exclamation marks,
sentences under 25 words where possible. After the draft, list every place
you were unsure whether it matched OURS.

What it still gets wrong

Tom van Nuenen had three frontier models rewrite 300 personal narratives and measured 13 linguistic markers across the results (“Voice Under Revision”, arXiv 2604.22142, April 2026). The direction was consistent. Function words, contractions and first-person pronouns went down; vocabulary diversity, word length and punctuation elaboration went up. The rewritten texts converged toward each other and became harder to match back to their original author.

The same paper describes what that looks like inside the sentence. Narration moves from embedded to distanced, and reasoning that ran as an explicit causal chain comes back compressed into a tidier abstraction. Your argument survives. The working out that made it recognizably yours does not.

Prompts written specifically to preserve a writer’s voice reduce the magnitude of the drift. They do not eliminate its direction. A voice prompt buys you a smaller correction, and the correction runs the same way every time, toward the flatter and more polished register.

The check before you publish

Count the contractions and the first-person pronouns in your original next to the model’s version. If both dropped, the voice drifted, whatever the prompt said. Then read one paragraph out loud, because the longer words and the elaborated punctuation are audible before they are visible.

Contrastive pairs get you close enough that the edit is a pass rather than a rewrite. They do not get you to the point where you can skip the pass.

5. The Repetitive Work Eating Your Week

The job worth automating is not the article. It is the small things you do after the article is finished.

The Friday afternoon problem

The post is done. Now there are three meta descriptions, four social variants, a newsletter blurb, alt text for five images and a note for the team channel. Each one takes five minutes, and none of them takes five minutes, because every one of them requires reloading the whole article into your head first. The switching is what costs you, once per item.

Hand it the finished piece, not the topic

• The finished asset in full. Every derivative is a compression of something that already exists and has already been fact-checked, which is what makes this the safest AI job on the list and the natural place to start repurposing content.
• The hard constraints per channel, written as numbers. Character limits, the claims legal will not allow, whether you use hashtags at all.
• A closed set of facts. State plainly that nothing may appear in a derivative that is not in the source text.
• The house rules you always forget to repeat. No emoji, no exclamation marks, no question-mark openers.

The prompt

Below is a finished, published piece. Everything you produce must come
from it. Do not add a statistic, a claim or an example that is not in the
text below.

ARTICLE:
[paste the finished piece]

Produce, each clearly labeled:
1. Three meta descriptions, under 155 characters each, each leading with a
   different benefit
2. Four social posts: one for LinkedIn (under 1,200 characters, no
   hashtags), one for X (under 280 characters), one newsletter preview
   (two sentences), one internal note (one line)
3. Alt text for every image referenced in the piece

Rules: no emoji, no exclamation marks, no hashtags, and do not reuse the
same opening structure twice across the set. If you are not certain of a
current character limit, say so instead of guessing.

What it still gets wrong

A model will tell you a platform’s character limit, field length or posting rule with complete confidence and no way of knowing whether it changed last month. Platform specs move faster than any training cutoff, and this is the one category where the output looks like a fact rather than a draft, so nobody checks it. That is the whole reason the prompt above tells it to flag an uncertain limit instead of guessing.

Batch output also converges. Ask for four social posts in one go and you get four posts with the same rhythm and the same opening move, which is the normalization effect from the voice section showing up again at small scale. The explicit “do not reuse the same opening structure” line helps, and it only half works.

There is a second reason the closed-facts rule matters most here. A derivative travels without the article attached, so a number that drifted during the compression gets read as a standalone claim by people who never open the post.

The check before you publish

Paste each derivative into the real field and read the count there, not the one the model reported. Then scan for any number that is not in the source article and delete it. Both halves take under a minute for a full set, and they are the only two ways this task fails.

You still review everything. What disappears is the context switching, and reviewing a finished set takes less time than the switching did. If what you actually want is a comparison of the tools that run this kind of batch work, that is a separate piece: the best AI copywriting tools.

When AI Is the Wrong Tool for the Job

Four jobs on this page get worse the moment a model touches them.

• Original reporting. If the fact does not exist yet because nobody has been asked, no model can retrieve it, and the gap gets filled with something plausible. A Wyoming reporter was caught using AI to invent quotes attributed to real, named people.
• Anything whose value is that a specific person is saying it. A founder’s letter, a condolence note. The value is provenance, which a model cannot supply.
• A first draft of something you do not yet understand. MIT Media Lab’s “Your Brain on ChatGPT” study (arXiv preprint, June 2025) had 54 participants write essays under three conditions: ChatGPT, a search engine, no tool. The ChatGPT group showed the weakest brain connectivity and struggled to quote back what they had just produced, a pattern the researchers call cognitive debt. A draft you did not think through leaves you unable to defend it in the meeting, or notice what it got wrong.
• Sensitive or regulated copy. Medical, legal, financial, HR, anything with a compliance review attached. The cost of one confidently wrong sentence there is not an editing pass.

Where that line falls is the subject of AI vs content writers. The five methods on one screen:

The four jobs above the table are not slower with a model. They are worse with one.

Frequently Asked Questions

Does AI-written content rank on Google?

Yes. Google’s spam policy targets scaled content abuse, defined as “many pages are generated for the primary purpose of manipulating search rankings and not helping users” (policy page last updated 15 May 2026). Nothing in that wording penalizes AI assistance as such. What it describes is volume published without value, which is why the SEO section above spends its effort on the gap between your page and the three currently above it.

Can AI detectors tell if I used AI to write something?

Not reliably, and the errors land on human writers. Liang, Yuksekgonul, Mao, Wu and Zou at Stanford, in “GPT detectors are biased against non-native English writers” (Patterns, volume 4, article 100779, 2023), found more than half of TOEFL essays by non-native English speakers were misclassified as AI-generated, while the same detectors were near-perfect on US eighth-grade essays. Use a detector as a pre-publish QA layer, never as something to beat: best AI detectors.

Do I have to disclose that I used AI?

Not for ranking. Google’s generative AI guidance says sharing how a piece was made “can help give your readers more context” and suggests you “consider adding information on how your content was created” (page last updated 10 December 2025). That is a recommendation about reader trust, not a Search requirement. Whether you disclose is an editorial call, and the case for making it sits in AI copywriting ethics.

How much editing does an AI first draft actually need?

There is no honest percentage, so use the pass order instead. Structure and boilerplate move fastest, claims and voice take the most work, and the three passes in section 2 exist so you are never doing all three at once. Remember the Noy and Zhang caveat too: their measured gains came from tasks that did not demand precise factual accuracy, so the fact pass is exactly where some of the saved time goes back.

Does using AI make you a worse writer over time?

The MIT Media Lab EEG study (“Your Brain on ChatGPT”, June 2025) is the sharpest evidence available, and its fourth session is the part to read. Participants who had drafted with ChatGPT were switched to writing unaided, and the switch did not fully restore the engagement the earlier sessions had lost. The mitigation is cheap: think the argument through before the model sees it, which is what the “what to hand it” step in every method above is really for.

What’s the best prompt for writing a blog post with AI?

There is not one, and hunting for it is the mistake. Prompt libraries circulate because a prompt is easy to copy and a brief is not. Every prompt on this page would fit in a text message, and every one of them does nothing on its own, because the material stacked above the instruction is what moves the output. Two writers running the identical prompt on different inputs get two different articles.

Where This Leaves You

Five methods, five short prompts, and in every case the output was decided by what sat above the instruction: the brief, the source material, the contrastive pairs, the live SERP, the finished asset. That part is also the cheapest to fix, because the inputs are things you already own.

The checks did not change. All five are still manual and still yours, and none of them can be handed to the thing being checked. Models got much faster at producing a draft and no better at knowing whether the draft is true, or whether it sounds like you.

If the drafting end of that work is what eats your week, Instacopy generates blog, ad, email and social copy from templates in 25+ languages and 20+ tones. It gets you to a draft faster, and every check above still belongs to you.

Pick one method from this article and run it on your next piece, with the check attached. One method with its check beats five prompts without them.