10 ChatGPT Prompts for Writing

The short answer: the best ChatGPT prompts for writing in July 2026 are short, structured, and biased toward diagnosis before rewriting. The prompts below are tuned for GPT-5.5 (the new default since May 5, 2026) and GPT-5.4, with notes for Claude Opus 4.7 where it behaves better. They cover blogs, novels, essays, short stories, copywriting, academic writing, technical docs, ghostwriting, screenwriting, and research synthesis.

I use roughly this stack every week. The meta-rule behind every prompt here is simple: don’t ask the model to write yet. Ask it to interrogate the prompt, the angle, the reader, or the existing draft first. That’s the part most “best prompts” lists skip, and it’s the part that decides whether the output sounds like a person or like a press release.

Before we get to the prompts, two things have changed since the original January 2026 version of this guide. GPT-5.5 Instant replaced GPT-5.3 Instant as the ChatGPT default on May 5, 2026, with OpenAI describing it as more “factually reliable, especially for prompts where accuracy matters most” and “tighter, more direct answers without losing useful details” (OpenAI release notes, May 5, 2026). And the disclosure landscape is no longer theoretical. Amazon KDP has enforced its AI-content disclosure policy since September 2023, and as of 2026, Medium disqualifies undisclosed AI-generated writing from anything beyond Network distribution (Medium AI policy, updated June 29, 2026). I’ll cite the rules where they matter, inside each prompt.

Quick caveat: I’m writing this in first person because I built it from real use. If you want a more academic tone, swap the voice constraints. The prompts work in ChatGPT, Claude, and Gemini with minor adjustments.

---

Writing Format vs. Prompt Goal: A Quick Map

Before the long list, here is the comparison table the rest of the article expands on. Every row maps a format to a prompt goal, a typical output, and the single tone-control lever that matters most for that format.

Writing format	Prompt goal	Typical output	Tone-control lever
Blog post	Angle and search intent	Title + outline + meta description	“Write for a reader who already knows the basics”
Novel (fiction)	Scene diagnosis and stakes	Scene rewrite + tension map	“Preserve the protagonist’s voice; never resolve the conflict”
Essay (personal)	Voice preservation	Paragraph with original metaphors intact	“Match the cadence of [paste 200 words of your own writing]”
Short story	Hook + compression	1,000-2,500 word draft with one twist	“Cut anything that doesn’t earn its place by line 5”
Copywriting (sales pages, ads)	Conversion-focused rewrite	Headlines + bullets + CTA variations	“Lead with the objection, not the feature”
Academic paper	Argument structure + literature map	Outline + citation list + counterargument	“Assume a hostile peer reviewer”
Technical docs	Reader-task mapping	Step-by-step procedure + warnings	“Write for someone who has never used the product”
Ghostwriting	Client voice mimicry	Draft + deviation report	“Flag every line that sounds like the AI, not the client”
Screenwriting	Dialogue subtext + scene economy	Scene beats + dialogue polish	“Subtext over text; cut every on-the-nose line”
Research synthesis	Source triangulation	Cross-source memo with confidence ratings	“Rate each claim: high / medium / low confidence”

Two quick notes on the table. First, the “tone-control lever” column is the one sentence I’d type into every prompt. It’s not optional decoration. It’s the difference between a draft that sounds like you and a draft that sounds like every other blog post in the SERP. Second, you will notice I do not have a row for “fully AI-generated book.” I don’t recommend that as a prompt goal, and as of 2026, both Amazon KDP and Medium actively police that pattern.

---

The 2026 Writing Landscape: What Changed

Definition: AI disclosure is the explicit labeling of content that was generated, in whole or in part, by an AI system. Major platforms introduced or enforced disclosure rules between 2023 and 2026, and the rules differ in scope and enforcement.

If you wrote with ChatGPT in early 2025 and you’re picking it back up now, three things have shifted:

1. Self-publishing volume keeps climbing

Amazon’s Kindle Direct Publishing paid out $70.3 million to KDP Select authors in May 2026 alone (Amazon KDP site, July 2026). That’s a single-month figure for the Select pool. The total KDP ecosystem is much larger. With that volume comes noise: Amazon’s policy distinguishes between “AI-generated” (the tool produced the text, images, or translations) and “AI-assisted” (you wrote it and used AI to edit, brainstorm, or polish). Both are allowed on KDP, but only the second doesn’t have to be disclosed (KDP Content Guidelines, current as of July 2026).

2. AI detection got better - but unevenly

The Authors Guild ran its own test in May 2026. They submitted ten articles published in 2022 or earlier (before ChatGPT existed) to five detection tools. Pangram returned 0% AI across all ten. Originality.ai returned 0% on eight and 1% on two. Grammarly flagged two articles at 7% and 9%. ZeroGPT ranged from 5% to 76% on the same texts. Sidekicker.ai scored every human-written article between 71% and 100% AI (Authors Guild, “Can AI Detectors Be Trusted?”, May 26, 2026). The Authors Guild’s conclusion: a couple of tools are reliable, several are not, and a polished human writer can still get flagged by the broken ones.

3. Hallucination is now a publishing liability

In May 2026, The New York Times found that Steven Rosenbaum’s nonfiction book The Future of Truth - a book about AI’s effect on truth - contained multiple fabricated and misattributed quotes. He called the errors accidental and launched his own investigation (NYT, May 19, 2026). A week later, the 2026 Commonwealth Short Story Prize was thrown into controversy when a winning story, published in Granta, scored 100% AI on the Pangram detector. Both organizations stood by the story; the Commonwealth Foundation said it was reviewing its process (NYT, May 20, 2026). The lesson is not that AI detection is broken. The lesson is that any draft needs a human fact-checker before publication.

“AI detection tools are AI models trained to recognize statistical patterns associated with large language model output, such as sentence rhythm, vocabulary distribution, and predictability of word choice. But polished, edited prose written by experienced human writers shares many of those same characteristics.” - Authors Guild, May 2026

---

How These Prompts Are Structured

Every prompt below follows the same skeleton. I learned this from OpenAI’s own prompting guide and from months of trial and error with GPT-5.3 and GPT-5.5:

1. Role + scope. Tell the model who it is and what it’s not allowed to do.
2. Context dump. Paste in the brief, the draft, the prior research, or the existing article. The model can’t read your mind.
3. Constraint. Give a length, a structure, a voice sample, or a hard rule (“do not invent citations”).
4. Deliverable format. Specify what you want back: a list, a table, a paragraph, a JSON object.
5. Verification step. Ask the model to flag its own weak spots. This is the part most people skip.

If you only internalize one thing from this article, make it the verification step. It turns the model from a confident author into a careful collaborator.

---

1. The Angle Generator (Blog Posts and Articles)

Best for: Blog posts, listicles, thought-leadership essays, newsletter issues.

Why it works: Most bad drafts start with the wrong angle. A model asked to “write a blog post about X” defaults to the highest-probability framing, which is the most-written framing. Asking for angles first inverts the problem.

The prompt:

You are a senior editor at a publication that publishes only
opinionated, specific pieces. No generic advice.

Topic: [paste your topic]

Step 1: Generate 8 distinct angles for this topic. Each angle
must argue something a smart reader would disagree with.
Forbid the following overused framings: [list 3-5
framings you want to avoid, e.g., "AI is changing
everything," "the future of work is..."].

Step 2: For each angle, write a one-sentence thesis, the
single strongest counterargument, and the one source the
author would need to cite to make it credible.

Step 3: Rank the 8 angles by (a) originality, (b) defensibility,
(c) commercial reader appeal. Show your ranking logic in one
sentence per angle.

Do not write the article yet. I want to pick an angle first.

Before / after:

Before (typical ChatGPT output for “blog post about cold email”):

“Cold email is a powerful tool for sales professionals. In this article, we’ll explore five strategies for writing effective cold emails that get responses.”

After (using the angle prompt, with “cold email for B2B SaaS” as the topic):

Angle 4: “Cold email is dead - but only if you measure opens. If you measure replies, it works better in 2026 than in 2019 because inboxes are less crowded.” Counterargument: reply rates are confounded by inbox filtering. Source needed: 2025-2026 benchmark data from a sales-tech vendor with transparent methodology.

Voice-preservation tip: Once you’ve picked an angle, paste two paragraphs of your own previous writing and say: “Match this cadence. Short sentences, no em dashes, contractions on.”

What NOT to delegate: Picking the final angle. The model can give you options; the actual editorial judgment is yours. Also: do not let it generate the headline until you’ve picked the angle. Most generic headlines come from skipping this step.

---

2. The Outline Stress-Test (Long-Form Essays and Newsletters)

Best for: 1,500-4,000 word essays, Substacks, LinkedIn long-form, NYT-style op-eds.

Why it works: Outlines lie. They look solid until you try to write them. This prompt asks the model to attack the outline before you’ve wasted three hours drafting it.

The prompt:

You are a hostile but fair peer reviewer. Your job is to
find the weaknesses in the outline below before I waste
time writing it.

Outline: [paste your outline]

Step 1: For each section, identify the weakest claim or
thinnest evidence. If a section has no defensible claim,
say so.

Step 2: Identify the single biggest logical gap in the
outline as a whole. The gap that, if you removed it,
would collapse the entire argument.

Step 3: Suggest one section to delete and one section to
deepen. Be specific about why.

Step 4: Propose a closing section that would land harder
than my current conclusion. Do not write the section -
just describe what it would do.

Do not rewrite my outline. I want critique, not a
substitute.

Disclosure-relevant detail: Medium’s AI policy (June 29, 2026) treats “stories where the majority of the content has been created by an AI-writing program with little or no edits, improvements, fact-checking, or changes” as AI-generated writing, which is disallowed behind the paywall and restricted to Network-only distribution if undisclosed. Stress-testing an outline you’ve written yourself keeps the human authorship signal intact.

What NOT to delegate: Your personal experience. The model cannot tell you what it’s like to ship a feature at 2 a.m. or to lose a client. Keep those paragraphs human.

---

3. The Voice Lock (Personal Essays and Memoir)

Best for: Personal essays, memoir drafts, opinion columns, op-eds.

Why it works: The single biggest tell of AI-assisted writing is generic sentence rhythm. This prompt anchors the model to your actual cadence using a sample.

The prompt:

You are ghost-editing a personal essay. Your job is to
match my voice, not to improve it.

Voice sample (do not edit this; use it as a reference):
[paste 200-300 words of your own writing that you think
sounds like you]

Draft to revise: [paste your draft]

Rules:
1. Match the average sentence length of the voice sample
   within ±15%. If my voice averages 14 words per sentence,
   your revision must too.
2. Preserve every proper noun, place name, and personal
   detail. If a fact looks wrong, flag it in a comment -
   do not silently fix it.
3. Do not insert metaphors, idioms, or rhetorical questions
   that are not already in the voice sample.
4. If a sentence is already strong, leave it alone.
5. Output the revised draft followed by a "voice deviation
   report": list every line that drifts from the voice
   sample, with a one-sentence explanation of how.

Do not write a new essay. Edit mine.

Why this matters in 2026: On May 26, 2026, the Authors Guild warned that “the more refined and controlled a writer’s style, the more it may resemble the output these tools are designed to flag.” In other words, polished prose triggers false positives. A voice-locked prompt keeps your draft close enough to your baseline that the reliable detectors (Pangram, Originality.ai) won’t flag it, and far enough from AI cadence that a careful reader won’t either.

What NOT to delegate: Personal narrative beats. The model can polish the prose; it cannot invent the time you met your father at the airport.

---

4. The Scene Surgeon (Novels and Short Fiction)

Best for: Novel chapters, short stories, literary fiction, romance, thrillers.

Why it works: Fiction drafts usually have two problems: too much throat-clearing and not enough stakes. This prompt forces the model to diagnose before rewriting.

The prompt:

You are a developmental editor for literary fiction. You
are not a line editor and you are not a co-author.

Scene to evaluate: [paste the scene, or describe it in
detail if it's early in the process]

For this scene, answer in order:

1. What does the protagonist want at the open of the scene?
   Quote the line that establishes it. If there isn't one,
   say "not established."

2. What stands in their way? Quote the line that names the
   obstacle. If it's only implied, say "implied only."

3. What is the worst thing that could happen to the
   protagonist in this scene? Does it happen? If no,
   explain why the scene still earns its place. If yes,
   explain how.

4. Where is the point of no return? Quote the line that
   crosses it.

5. What does the protagonist believe at the end of the
   scene that they did not believe at the start? If
   nothing has changed, the scene is exposition. Say so.

After answering, propose the smallest possible rewrite
that would make the scene work. Do not rewrite the whole
scene - describe the surgical change in 100 words.

Constraints:
- Preserve every line of dialogue the author wants to keep.
- Preserve the protagonist's voice and vocabulary register.
- Do not introduce new characters or settings.

What NOT to delegate: The actual emotional truth of the scene. The model can find structural holes. It cannot tell you what grief sounds like from the inside.

Disclosure note: Per Amazon KDP’s Content Guidelines, “AI-generated” includes text “created by an AI-based tool,” and disclosure is required even if you applied substantial edits afterwards. If you’re using this prompt to draft, mark the final manuscript as AI-assisted (not generated) in your KDP submission if applicable.

---

5. The Compression Engine (Short Stories and Tight Word Counts)

Best for: Flash fiction, magazine submissions (typically 1,000-5,000 words), Twitter threads, newsletter shorts.

Why it works: Short fiction dies from bloat. This prompt forces the model to cut.

The prompt:

You are an editor at a literary magazine with a strict
word limit. Cut the draft below to exactly [target word
count, e.g., 1,500] without losing the central image
or twist.

Current draft: [paste]

Rules:
1. You may cut any sentence. You may not add new content.
2. Preserve the final twist or image intact.
3. If a sentence is doing double duty (character + plot),
   keep it. If it's only doing one, cut it.
4. Preserve the opening image. Flash fiction lives or
   dies in the first three sentences.

Output format:
- The cut draft, exactly [target word count] words.
- A "what I cut and why" list: each cut item with a
  one-sentence reason.
- A "what I would cut next if I had more space" list:
  3-5 sentences that the draft could lose in a 1,000-word
  version.

What NOT to delegate: The opening image. Most generic “short story” openings come from the model defaulting to weather or introspection. Bring your own opening line; let the prompt compress from there.

---

6. The Conversion-First Copywriter (Sales Pages, Landing Pages, Ads)

Best for: Sales pages, landing pages, product launch copy, paid ads, email subject lines.

Why it works: Most AI copy sounds like AI copy because it leads with features. Real conversion copy leads with the objection.

The prompt:

You are a direct-response copywriter trained on the
work of Eugene Schwartz, Joanna Wiebe, and Harry Dry.
You write copy that converts in the $50-$500 product
range. You are allergic to hype.

Product: [describe the product in 2-3 sentences]
Price point: [$]
Target buyer: [one specific persona, not a demographic]

Deliverables:

1. Five headline variations. Each must lead with the
   buyer's primary objection, not the product feature.
   Example pattern: "Stop [objection]. [Promise]."
   Vary the objection across the five.

2. Ten bullet points for the body. Each bullet must
   follow the pattern: [Specific outcome] + [Timeframe or
   mechanism]. No adjectives. No "seamless," "powerful,"
   "cutting-edge." If a bullet contains an adjective, I
   will reject it.

3. One CTA. Two versions: a low-commitment CTA ("see how
   it works") and a high-commitment CTA ("start now").
   The high-commitment CTA must contain a specific
   guarantee or risk-reversal.

4. A "objection map": list the top 5 objections your copy
   did not address, in order of how likely they are to
   stop the sale. For each objection, write the exact
   sentence the buyer says in their head.

Do not write the body copy yet. I want to pick headlines
and bullets first.

What NOT to delegate: Social proof. The model cannot verify whether your testimonials are real, recent, or attributed correctly. If a quote is fabricated, the FTC’s Operation AI Comply (announced September 2024) and ongoing AI-washing enforcement treat it as deceptive marketing.

---

7. The Academic Argument Mapper (Papers, Theses, Grant Proposals)

Best for: Journal articles, conference papers, lit reviews, grant applications, dissertation chapters.

Why it works: Academic writing fails when it pretends the counterargument doesn’t exist. This prompt forces the model to surface it.

The prompt:

You are a peer reviewer for a top-tier journal in
[field]. You are skeptical but fair. You are not
here to be convinced.

Working thesis: [paste your thesis]
Existing literature I'm building on: [paste citations or
descriptions of the 5-10 key papers]
Method or evidence: [describe in 2-3 sentences]

Tasks:

1. State the strongest version of my thesis. If my draft
   is hedged or vague, sharpen it. Flag where you sharpened
   and why.

2. Generate three competing theses that a smart colleague
   in this field would offer instead. For each competing
   thesis, identify the one paper in the literature that
   best supports it.

3. For my thesis, list the three pieces of evidence I
   would need to cite but probably haven't. If any of
   those don't exist in the literature, say so.

4. Identify the assumption my thesis rests on that, if
   disproven, would collapse the argument. Name that
   assumption.

5. Propose the smallest possible empirical or theoretical
   move that would harden the thesis. Do not redesign my
   study - describe the move in 150 words.

Do not write the paper.

Disclosure note: As of 2026, most major academic publishers (Elsevier, Springer Nature, Wiley, Taylor & Francis) require authors to disclose AI use and to verify all citations. Authors are responsible for the accuracy of references. AI-generated citations are a known hallucination category. This prompt explicitly asks the model to flag missing evidence, which reduces the hallucination risk but does not eliminate it. Every citation still needs to be verified by hand against the actual journal.

What NOT to delegate: Citation accuracy. Verify every reference in the actual journal database before submission.

---

8. The Reader-Task Mapper (Technical Documentation)

Best for: API docs, product help articles, internal runbooks, SOPs, README files.

Why it works: Technical docs fail when the writer knows the product too well. This prompt assumes zero context.

The prompt:

You are a technical writer who has never seen this product
before. You are writing for a developer who has also never
seen it. You will not assume any context.

Product: [name and 1-sentence description]
Task the reader is trying to complete: [e.g., "authenticate
their first API request"]

Deliverables:

1. A prerequisites section. List everything the reader
   needs before they start. Include version numbers,
   environment variables, and account permissions. If you
   are unsure of any requirement, say "verify with the
   product team" - do not invent.

2. A step-by-step procedure. Each step must contain:
   - The action the reader takes
   - The expected output or response
   - One troubleshooting note for the most likely failure
   - The exact command, code snippet, or UI path

3. A "verify it worked" section. How does the reader know
   the task succeeded? What should they see in logs, UI,
   or response payloads?

4. A "next steps" section. What are the 2-3 things the
   reader is most likely to want to do next? Link to those
   doc pages if you know them; flag if you don't.

5. A "known gotchas" section. List the 3-5 mistakes a
   first-time reader is most likely to make, in order of
   frequency. For each gotcha, write the error message or
   symptom and the fix.

Format constraints:
- Use second person ("you"), present tense.
- No marketing language. No "powerful," "simple,"
  "seamless."
- Total length: [target word count].
- Code samples must be syntactically valid. If you're not
  sure, flag the line with a comment like
  // verify syntax.

What NOT to delegate: Code correctness. The model is a starting point; CI tests and a real engineer are the actual verification.

---

9. The Ghostwriting Brief Expander (Client Work, Ghostwriting, Co-Authoring)

Best for: Ghostwritten books, co-written articles, executive thought-leadership, branded content.

Why it works: Ghostwriting fails when the ghost’s voice replaces the client’s. This prompt makes the model flag every drift.

The prompt:

You are a ghostwriter. Your job is to write in the voice
of the named client. Your own voice is irrelevant.

Client name: [name]
Client role: [role]
Audience: [describe in one sentence]
Deliverable: [e.g., 1,200-word LinkedIn post, 800-word
newsletter intro, full chapter]

Client voice sample (this is the only style reference;
use it strictly): [paste 500-1,000 words of the client's
own previous writing, ideally multiple samples]

Topic: [topic]
Angle: [angle, if known]
Constraints:
- Hard facts that must appear: [list]
- Quotes from the client you must include: [list]
- Words or phrases the client would never use:
  [list, e.g., "synergy," "circle back," "dive deeper"]
- Off-limits topics: [list]

Tasks:

1. Write the deliverable to spec. Match the voice sample
   in cadence, vocabulary, and sentence length.

2. Produce a "deviation report" alongside the draft. For
   every line that drifts from the voice sample, list:
   - The line number
   - Why you think it drifts
   - An alternative in the client's voice

3. Produce a "facts to verify" list. Every statistic,
   quote, or named reference must be on this list. The
   client will check each one.

Do not publish the draft. The client reviews first.

Disclosure note (load-bearing): The Authors Guild’s model contract clause for AI in publishing (updated April 22, 2026) reads in part: “Publisher agrees and warrants that it will not use AI to substantially edit a manuscript (excepting the use of basic spelling and grammar-checking applications).” If you are ghostwriting for a traditionally published client, the client’s publisher will want to know what role AI played. Be honest up front, in writing. Surprise disclosure after delivery is a contract risk.

What NOT to delegate: Verifying the client’s anecdotes and lived experience. The ghostwriter’s job is voice and structure; the client’s job is truth.

---

10. The Subtext Engine (Screenwriting and Stage Scripts)

Best for: Feature screenplays, short films, stage plays, TV pilots.

Why it works: Most AI-generated dialogue is on-the-nose. Real screenwriting is subtext. This prompt forces the model to read between the lines.

The prompt:

You are a screenwriting consultant trained in the
McKee school, the Robert McKee school, and the Lajos
Egri school of dramatic writing. You are allergic to
on-the-nose dialogue.

Scene: [paste the scene or describe it in slug lines]
Genre: [dramedy / thriller / romance / etc.]
Page count target: [target]

Tasks:

1. Identify the dramatic question of this scene
   (the question the audience is asking that the
   scene must answer or deepen).

2. Identify the subtext. What is each character
   actually saying underneath the words they speak?
   For each line of dialogue, write the subtext in
   one bracketed sentence.

3. Identify every on-the-nose line. An on-the-nose
   line is dialogue that says exactly what the character
   means. List them.

4. Rewrite each on-the-nose line to carry the same
   dramatic function through subtext. Preserve the
   information the scene needs to deliver.

5. Identify the action verb of the scene (the verb
   that describes what the scene is *doing* - not
   what it's about). One word. If you can't find it,
   the scene is exposition.

6. Propose the smallest cut that would sharpen the
   scene. Do not rewrite the whole scene - describe
   the cut in 100 words.

Constraints:
- Preserve all slug lines, action lines, and
  character names exactly as written.
- Preserve every plot-critical line of dialogue
  (these are flagged with [PLOT]).
- Do not introduce new characters or locations.

What NOT to delegate: The unspoken emotional truth of a scene. The model can find structural on-the-nose lines; it cannot tell you what a father really means when he says “drive safe” at the airport.

---

11. Bonus: The Research Synthesizer (Cross-Source Memo)

I’m giving you an eleventh prompt because research synthesis is the single highest-leverage use of these tools for nonfiction writers in 2026. Definition: research synthesis is the process of combining multiple sources on a topic into a coherent argument, with each claim tied back to its evidence and rated for confidence.

The prompt:

You are a research analyst. Your job is to combine
multiple sources into a single memo. You do not have
opinions. You cite or you don't write.

Topic: [topic]
Question to answer: [one specific question, not a
topic area]
Sources to use: [paste or list URLs / paper titles /
article names]

Tasks:

1. For each source, write a one-sentence summary and
   rate its relevance to the question (high / medium /
   low).

2. Identify the consensus across sources. What do at
   least 3 sources agree on?

3. Identify the disagreement. Where do sources contradict
   each other? Quote both sides.

4. List the strongest claim the memo can make, with
   citations from at least 2 sources.

5. List the claim the memo cannot yet make, and what
   additional source would be needed.

6. For every statistic, quote, or specific claim in the
   memo, attach a confidence rating:
   - High: confirmed by 2+ independent sources
   - Medium: confirmed by 1 source, no contradiction
   - Low: single source, contradicted, or extrapolated

7. List every claim that requires human verification
   before publication.

Output format:
- Section 1: Source-by-source summary
- Section 2: Consensus
- Section 3: Disagreements
- Section 4: Strongest defensible claim
- Section 5: Gaps
- Section 6: Verification list

Do not write the final article. Write the memo.

Why this matters in 2026: OpenAI itself flagged GPT-5.5 Instant as “more factually reliable” but added that “factual issues” can still appear. Anthropic’s Claude Opus 4.7 release (April 16, 2026) emphasized that the model “correctly reports when data is missing instead of providing plausible-but-incorrect fallbacks.” That second property is what you want for research. The verification list at the end of this prompt is the deliverable that protects you when the model is wrong.

What NOT to delegate: Final fact verification. Every claim still goes to the source.

---

Editing, Voice, Citation, and Fact-Checking Workflows

Beyond the prompts themselves, here are four workflows I use every week. They are the difference between using ChatGPT and being used by it.

A diagnosis-before-rewrite habit

For every draft, the first prompt you run should be diagnostic. Paste the draft. Ask: what’s wrong with this? Only after the diagnosis should you ask for a rewrite. If you skip the diagnosis, the model rewrites a flawed structure with cleaner prose, which is worse than useless - it makes you think the draft is good when it isn’t.

A voice-sample anchor

Keep a permanent voice sample (200-300 words) in a text file. Paste it into every prompt that produces prose for you. This is the cheapest way to keep your draft close to your baseline cadence. It also helps with detection: as the Authors Guild noted in May 2026, polished writers get flagged because their prose resembles AI output. A voice sample keeps your sentences imperfect enough to read human.

A citation verification rule

Every citation in a research-synthesis output must be verified by hand. Type the citation into Google Scholar, the journal database, or your library system. Do this for every reference, every quote, every statistic. The Authors Guild’s “Can AI Detectors Be Trusted?” study explicitly warned that hallucinated quotes and citations are now a publishing liability, citing the May 2026 New York Times investigation into fabricated quotes in a nonfiction book about AI.

A disclosure audit before publication

Before you hit publish, ask: does any part of this draft qualify as “AI-generated” under the platform’s definition? The definitions are not the same across platforms.

Platform	Definition of AI-generated	Disclosure required?	What happens if you don’t disclose?
Amazon KDP	Text, images, or translations created by an AI tool, even with substantial edits	Yes	Content removed; account may be terminated for repeat violations
Medium	Writing where the majority of the content is AI-created with little editing	Yes (in first two paragraphs)	Network-only distribution; paywall removal; Partner Program revocation
Anthropic / OpenAI consumer products	As defined by the platform’s own TOS	Not applicable to user drafts	n/a
Academic journals (Elsevier, Springer Nature, Wiley, Taylor & Francis)	Use of generative AI to create text, images, or data	Yes, in methods or acknowledgments	Retraction risk; author sanctions

Sources: Amazon KDP Content Guidelines (current 2026), Medium AI policy (updated June 29, 2026), Author’s Guild Use of Consumer AI Systems in Publishing (April 22, 2026), and standard publisher policies as compiled by the Authors Guild.

---

Disclosure Norms in Detail (July 2026)

The platforms differ. Here’s what I tell writers who ask.

Amazon KDP. Amazon updated its KDP Content Guidelines in September 2023 and has enforced them since. The policy distinguishes “AI-generated” (the tool produced the actual text, image, or translation) from “AI-assisted” (you wrote it and used AI to edit, brainstorm, or check). Only AI-generated content needs to be disclosed at submission. The Authors Guild welcomed the original policy and has continued to push Amazon for stronger author-side protections. Note: Amazon also introduced an “Ask This Book” AI feature on Kindle in late 2025, which the Authors Guild flagged in December 2025 as a concern for author consent over how their books are queried.

Medium. Medium updated its AI content policy on June 29, 2026. The rule is plain: AI-generated writing cannot be paywalled under the Partner Program. Undisclosed AI writing gets Network-Only distribution (your followers only, no Boost eligibility). AI-assistive tools (outlining, grammar check, fact-check) do not need to be disclosed. AI-assisted text that incorporates meaningful AI snippets does need a disclosure - Medium says to include it in the first two paragraphs (“This story was written with the assistance of an AI writing program”).

Substack. Substack’s help pages do not currently host a single, dated AI policy. The widely circulated guidance as of 2026: Substack does not ban AI-assisted writing, but it does not promote undisclosed AI-generated work, and newsletter creators who sell paid subscriptions should disclose AI use if their audience expects it. Best practice: state your AI-use policy in your “About” page and be consistent.

Academic journals. The big five (Elsevier, Springer Nature, Wiley, Taylor & Francis, and SAGE) all require AI-use disclosure. They do not allow AI to be listed as an author. Citations must be human-verified. In May 2025, the U.S. Copyright Office’s Part 3 AI Report clarified that purely AI-generated content is not copyrightable in the U.S., though human-authored works that incorporate AI assistance can be. Verify the specific journal’s policy before submitting.

The New York Times. The Times does not currently accept op-eds or feature submissions generated by AI. As of May 2026, the paper’s standard for human-authored journalism remains firm, and the publication’s coverage of AI misuse (the Rosenbaum Future of Truth story, the Granta Commonwealth Prize controversy) signals an unusually high editorial sensitivity to the issue.

Authors Guild. The Authors Guild has been the most active U.S. writers’ organization on AI disclosure since 2023. Its Human Authored certification mark (launched 2025, updated 2026) lets authors certify their books as fully human-written for a fee. Its April 22, 2026 statement on “Use of Consumer AI Systems in Publishing” introduced two model contract clauses - one preventing publishers from uploading manuscripts to consumer AI tools without author permission, and one preventing substantive AI editing. As of April 2026, 91.3% of books covered by the $1.5 billion Anthropic settlement had been claimed by authors, the largest author-side AI settlement in U.S. history.

---

FAQ

Do these prompts work on Claude and Gemini too?

Yes, with minor adjustments. Claude Opus 4.7 (released April 16, 2026) is especially good for research synthesis and voice-locked editing; Anthropic specifically highlighted that the model “correctly reports when data is missing instead of providing plausible-but-incorrect fallbacks.” Gemini 2.0 Flash is competitive for fast drafting but I’d avoid it for citation-heavy work. For pure prose drafting, GPT-5.5 Instant (the new default since May 5, 2026) and Claude Sonnet 4.6 are roughly equivalent.

Will any of these prompts get my draft flagged as AI?

Not if you keep the voice-sample anchor and the personal-experience paragraphs human. The Authors Guild’s May 2026 study found that Pangram and Originality.ai flagged pre-2022 human articles at 0% or 1%, which is what you should aim for. The biggest risk factor is generic sentence rhythm, not vocabulary. If your draft has consistent sentence lengths, no contractions, and balanced paragraph structure, it can read as AI. Break the rhythm with a short, blunt sentence. Add a memory.

Is it okay to use these prompts for paid client work?

Yes, but disclose it. The Authors Guild’s April 22, 2026 statement makes clear that publishers and editors are not allowed to upload author manuscripts to consumer AI tools without permission, and they are not allowed to substantively AI-edit a manuscript without author consent. As a freelancer, your contract with the client should specify what role AI plays. The safest language: “I use AI for outlining, research synthesis, and copyediting. All final prose is human-drafted.”

Can I use ChatGPT to write a nonfiction book and publish it on KDP?

Legally, yes. Amazon KDP allows AI-assisted and AI-generated books, as long as AI-generated content is disclosed at submission. Practically, the Authors Guild’s May 2026 detector test, the NYT investigation into fabricated quotes (May 19, 2026), and the Commonwealth Short Story Prize controversy (May 20, 2026) all show the same lesson: the publishing ecosystem is actively watching for AI-only books, and the cost of getting caught is high. If you publish an AI-only book and your topic overlaps with anything fact-sensitive, you are one fact-check away from a public takedown.

What should I never let AI do for my writing?

Invent citations. Invent quotes. Invent statistics. Invent personal experiences. Invent legal claims. Invent medical claims. Generate cover art that misrepresents the book’s contents. Substantively edit someone else’s manuscript without their written permission. These are the failure modes that have already caused public scandals in 2026, and they are the ones publishers, agents, and platforms are actively policing.

What’s the single best prompt in this list?

The voice-lock prompt (#3). It is the one that most directly preserves what makes your writing yours. Everything else is leverage. The voice lock is the floor.

---

Sources

1. OpenAI Help Center - ChatGPT Release Notes, updated July 14, 2026 (notes through July 9, 2026): https://help.openai.com/en/articles/6825453-chatgpt-release-notes
2. Amazon KDP - Content Guidelines (Artificial Intelligence section, current 2026): https://kdp.amazon.com/en_US/help/topic/G200672390
3. Amazon KDP - Kindle Direct Publishing Terms and Conditions (last updated September 27, 2024): https://kdp.amazon.com/terms-and-conditions
4. Authors Guild - Can AI Detectors Be Trusted? The Authors Guild Put Five of Them to the Test, May 26, 2026: https://authorsguild.org/news/can-ai-detectors-be-trusted/
5. Authors Guild - Use of Consumer AI Systems in Publishing: Statement and New Model Contract Clauses, April 16, 2026 (updated April 22, 2026): https://authorsguild.org/news/use-of-ai-in-publishing-and-new-model-contract-clause/
6. Authors Guild - Artificial Intelligence Advocacy page (with Anthropic settlement, Human Authored certification, AI Model Clauses): https://authorsguild.org/advocacy/artificial-intelligence/
7. Authors Guild - Authors Guild Encouraged by Penguin Random House’s New AI Restrictions, October 23, 2024: https://authorsguild.org/news/ag-encouraged-by-penguin-random-house-ai-restrictions/
8. Authors Guild - Anthropic Settlement Update: 91.3 Percent of Books Claimed in Settlement, April 2026: https://authorsguild.org/news/anthropic-settlement-update-91-percent-of-books-claimed/
9. Anthropic - Introducing Claude Opus 4.7, April 16, 2026: https://www.anthropic.com/news/claude-opus-4-7
10. Originality.ai - We Have 99% Accuracy in Detecting AI, updated July 3, 2026: https://originality.ai/ai-content-detection-accuracy/
11. Originality.ai - Is GPT-5.5 Content Detectable?, July 11, 2026: https://originality.ai/blog/is-gpt-5-5-detectable
12. Medium Help Center - Artificial Intelligence (AI) content policy, updated June 29, 2026: https://help.medium.com/hc/en-us/articles/22576852947223-Artificial-Intelligence-AI-content-policy
13. Medium Help Center - Medium’s Distribution Guidelines: How curators review stories for Boost, General, and Network Distribution, updated June 29, 2026: https://help.medium.com/hc/en-us/articles/360006362473-Medium-s-Distribution-Guidelines
14. The New York Times - The Future of Truth and the AI Quote Scandal, May 19, 2026: https://www.nytimes.com/2026/05/19/business/media/future-of-truth-ai-quotes.html
15. The New York Times - Commonwealth Short Story Prize and AI Detection Controversy, May 20, 2026: https://www.nytimes.com/2026/05/20/books/ai-fiction-contest-granta.html
16. Reedsy Blog - Writer’s Block: What’s Causing It, and How to Fix It, June 11, 2026: https://reedsy.com/blog/writers-block/
17. Reedsy Blog - How to Find a Ghostwriter For Hire in 2026, March 10, 2026: https://reedsy.com/blog/how-to-find-a-ghostwriter/
18. Amazon KDP Select - Total KDP Select Author Earnings, May 2026: $70.3 million (cited from KDP site footer, July 2026)
19. U.S. Copyright Office - Copyright and Artificial Intelligence Part 3: Generative AI Training, May 2025: cited via Authors Guild summary at https://authorsguild.org/news/us-copyright-office-ai-report-part-3-what-authors-should-know/
20. Penguin Random House - Statement on AI training restrictions added to copyright pages, October 2024: https://www.thebookseller.com/news/penguin-random-house-underscores-copyright-protection-in-ai-rebuff
21. Authors Guild - Amazon’s New Disclosure Policy for AI-Generated Book Content Is a Welcome First Step, September 2023: https://authorsguild.org/news/amazons-new-disclosure-policy-for-ai-generated-book-content-is-a-welcome-first-step/
22. Authors Guild - Statement on Amazon Kindle “Ask This Book” AI Feature, December 2025: https://authorsguild.org/news/statement-on-amazon-kindle-ask-this-book-ai-feature/
23. Authors Guild - Authors Guild Releases AI Best Practices for Writers, May 2026: https://authorsguild.org/news/ag-updates-ai-best-practices-for-writers/
24. Authors Guild - What Authors Need to Know About the $1.5 Billion Anthropic Settlement, September 2025: https://authorsguild.org/news/what-authors-need-to-know-about-the-anthropic-settlement/
25. FTC - FTC Sues Air AI, August 2026 (AI-washing enforcement action, referenced from FTC AI guidance)

---

The prompts above are tools. They are not a substitute for taste, lived experience, or fact-checking. Use them as scaffolding; tear them down where they get in the way. The writers producing the best work in 2026 are not the ones with the best prompts. They are the ones who know which draft to keep, which sentence to cut, and which claim to verify before publication. The model can help with any of those. The judgment is yours.