I keep a plain text file called prompts_that_failed.txt. It's longer than most of what I've published this year. Every entry is a prompt that sounded reasonable and produced something I couldn't use: a blog intro that opened with a rhetorical question, a chapter summary that invented a character, ten headlines with the same shape.
The prompts that worked came from somewhere else. Not from prompt libraries, and not from the "act as a world-class copywriter" school. They came from treating the model less like a writer and more like a fast, literal assistant who has read everything and remembers nothing about me. Once I stopped asking it to be good and started giving it things to work against, the output changed.
What follows is the short list that survived. Each tactic has a vendor document or a peer-reviewed study behind it, or a specific thing that happened when I used it. I've left out the advice you've already read: be specific, give context, break big tasks into small ones. All true, and all on the first page of every search result, which is why this article is about the rest.
If you only read this box • Long material goes at the top of the prompt, the question at the bottom. • The first output is a probe, not a draft. Read it to find out what the model assumed. • The biggest quality jumps came from editing prompts, not drafting prompts. • Every factual claim a model writes needs its own pass. Budget for it. |
Quick reference
The nine tactics at a glance
| # | Tactic | What it fixes | Setup |
|---|---|---|---|
| 1 | Document first, instruction last | The model losing track of a long draft | 1 minute |
| 2 | Voice pack plus a written style guide | Output that sounds like every other AI post | 20 minutes, once |
| 3 | Interview before draft | Generic drafts built on generic assumptions | 5 minutes per piece |
| 4 | Reverse outline | Repeated points and buried arguments | 2 minutes |
| 5 | Structure over step-by-step | Bloated prompts that fight reasoning models | 3 minutes |
| 6 | Ask for the fifth idea | Angles that match everyone else's | 2 minutes |
| 7 | Feed it rejected drafts | Vague "make it better" loops | 3 minutes |
| 8 | Make it a reader | Not knowing where readers stop | 2 minutes |
| 9 | Pull the claims out | Invented facts sliding into the final | 5 minutes |
The tactics, in the order I'd learn them
1. Put the document first and the instruction last
Backed by: Anthropic's prompting documentation and its 2023 long-context experiment.
Most people paste their instructions, then the draft. Anthropic's guidance says to do the opposite for long material: document at the top, question at the end. In their tests, putting the query last improved response quality by as much as 30 percent on long, multi-document inputs. Their 2023 long-context experiment added a second move: ask the model to pull out relevant quotes before it answers.
I do this even for 1,500-word drafts. The model stops answering a half-remembered version of the instruction and works from the text in front of it.
Prompt to try [Paste the full draft here] Above is a 2,400-word feature draft. First, quote the three sentences that carry the argument. Then tell me which section could be cut without losing any of them. |
2. Build a voice pack, then make the model write your style guide
Backed by: OpenAI's reasoning guidance on when examples help, and a lot of trial and error.
A single "write in my voice" line does nothing. Five to eight of your own paragraphs do a lot, but only if the model reads them the right way. What worked: paste the samples, ask for a written description of the patterns (sentence length, how paragraphs open, what I never do), correct that description by hand, then save it as the standing brief for every piece.
The correction step is the point. The first style guide it wrote for me said I "favour vivid imagery." I don't. I cut that line, added "opens with a concrete object, never a question," and the next draft opened with a coffee cup. OpenAI's guidance says to try a prompt without examples first and add them when needed. For voice, they're always needed.
Prompt to try Here are six paragraphs I wrote. Describe the patterns you see in sentence length, paragraph openings, punctuation and vocabulary. List anything I never do. Write it as a style guide I can reuse. Do not write any new prose yet. |
The voice pack lives on a sticky note now, because the first version the model wrote was wrong. Photo by RDNE Stock project via Pexels.
3. Ask for the interview before the draft
Backed by: what happens every time I skip it.
The generic AI article exists because the model was never told anything specific. The fix is to make it ask. Before any draft, I have it interview me: six questions, one at a time. The questions expose its assumptions. When I asked for a piece on freelance rates, its second question was "which country?" I hadn't thought to say.
Answer in fragments; the model will smooth them out. When it drafts, it has your numbers, your examples and the one story only you can tell.
Prompt to try I want to write a 1,200-word newsletter on [topic] for [reader]. Before drafting, ask me the six questions you'd need answered to write it well. Ask them one at a time and wait for my reply. |
4. Reverse outline your own draft
Backed by: decades of editors doing it by hand.
This is the tactic I'd keep if I could keep one. Paste a finished draft and ask for one line per paragraph stating what job it does. Then ask which paragraphs do the same job. The model does in seconds what takes an afternoon by hand, and it doesn't get bored by paragraph 30.
On a 4,000-word feature I'd been stuck on, the reverse outline found two paragraphs, 900 words apart, making the same point with different examples. I'd been staring at that draft for two weeks.
Prompt to try Write a reverse outline of the draft above: one line per paragraph, stating the job that paragraph does for the reader. Flag any two paragraphs that do the same job. Do not rewrite anything. |
5. Use structure instead of "think step by step"
Backed by: OpenAI's reasoning best practices and Anthropic's extended-thinking guidance.
For years the advice was to tell the model to think step by step. OpenAI's documentation for its reasoning models now says the opposite: those models reason internally, so asking them to explain their thinking is unnecessary. Keep prompts simple and separate the parts with delimiters such as headings or tags. Anthropic's advice on extended thinking points the same way, favouring general instructions over prescribed steps.
For writers, that means spending prompt space on labelled sections rather than coaching. My prompts got shorter and the results more predictable.
Prompt to try <draft> [paste] </draft> <audience> Freelance designers, three to ten years in, who read on a phone. </audience> <constraints> Keep every number. Maximum 900 words. No rhetorical questions. </constraints> <return> The edited draft only. </return> |
6. Ask for the fifth idea, not the first
Backed by: Doshi and Hauser, Science Advances, 2024.
A 2024 study in Science Advances by Anil Doshi and Oliver Hauser found that writers given AI-generated story ideas produced work rated as more creative and better written, with the biggest lift among less experienced writers. The catch: those stories were more similar to each other than stories written without help. Individually better, collectively narrower.
The way around it is to refuse the first answers. Ask for eight angles, rule out the first three because every other writer will use them, and pick from what's left. Or ask what a writer who has never read a listicle on the subject would notice first.
Prompt to try Give me eight angles for a piece on [topic]. Assume the first three that come to mind are what every other writer will use, so mark those and give me five that aren't. For each of the five, name the one detail that makes it specific. |
7. Feed it the drafts you rejected
Backed by: the model needing a target, not an adjective.
"Make it better" gives the model nothing to aim at. Rejected paragraphs do. When I paste a paragraph I cut and say why, the next attempt avoids that failure. Three rejected versions with reasons teach it more about my taste than any list of tone words.
This also fixes the vocabulary problem. If a draft keeps reaching for the same tired word, I paste the sentence, ban the word and its cousins, and ask for the sentence rebuilt around the concrete thing it was pointing at.
Prompt to try Here are three openings I wrote and cut, each with the reason I cut it. Write a fourth that doesn't fail in any of those three ways. Keep the same facts. |
8. Make it a reader, not a writer
Backed by: the drop-off point being the same one I already suspected.
The most useful role I've assigned isn't "expert editor." It's a specific reader: the person the piece is for, with their actual patience. I describe them in two sentences and ask where they'd stop reading and why. The paragraph it flags is usually the one I'd been avoiding. I run this before the line edit, so I fix the drop-off point instead of polishing sentences nobody will reach.
Prompt to try Read the draft above as [describe the reader in two sentences]. Tell me the exact sentence where you'd stop reading, and why. Then tell me the one thing you'd want to know that the draft doesn't say. |
9. Pull the claims out and check them separately
Backed by: Noy and Zhang, Science, 2023, and one near miss of my own.
Models write facts with the same confidence whether they're right or invented, and the invented ones read fine. So I don't fact-check prose. I ask for a list of every claim in the draft, each marked as coming from my material or from the model's own knowledge, and verify the second group myself.
This is the tactic that eats the time savings. In the MIT experiment by Shakked Noy and Whitney Zhang, published in Science in 2023, professionals with ChatGPT finished writing tasks 40 percent faster and quality rose 18 percent. But those tasks didn't require factual accuracy, and Noy said real-world gains would shrink once you count fact-checking.
Prompt to try List every factual claim in the draft above as a numbered table: the claim, whether it came from the material I gave you or from your own knowledge, and how confident you are. Do not fix anything yet. |
Which tactic for which job
Where I'd start, by tactic number
| Writing job | Start with | Add once it's working |
|---|---|---|
| Blog post or article | 3, 4, 9 | 6, 8 |
| Novel chapter | 2, 7 | 4, 8 |
| Email newsletter | 1, 3 | 9 |
| Client copy | 2, 8, 9 | 7 |
| Personal essay | 4, 7 | 6 |
What real writers say Two novelists ran their own experiments in public. Two surveys asked the wider field. Rie Kudan, winner of the 2024 Akutagawa Prize Kudan told reporters that around 5 percent of her winning novel, Tokyo Sympathy Tower, came directly from ChatGPT. The book has an AI as a character, and she used the tool's replies for that character's lines. When the AI didn't say what she expected, she wrote her own reaction into the protagonist instead. Reported by The Japan Times, summarised by Smithsonian Magazine, January 2024. Curtis Sittenfeld, novelist In 2024 Sittenfeld agreed to a New York Times test that set a story she wrote against one ChatGPT produced "in her style" from the same brief. She found the AI version so dull she wouldn't have finished it without the assignment, and summed it up in five words: "there's just something missing." The New York Times, August 2024, as reported by Commonweal. What the surveys found 45% of 1,229 authors surveyed by BookBub in 2025 used generative AI somewhere in their work. 48 percent didn't and didn't plan to. 22% of 787 respondents to the UK Society of Authors' January 2024 survey had used it in their work; 31 percent had used it to brainstorm. BookBub Partners survey, 2025; Society of Authors survey, April 2024 release. |
Mistakes I kept making
Each cost me at least a week before I noticed.
× Treating the first output as a draft It's a probe. Read it for what the model assumed about the reader and the facts, then fix the prompt before you fix a sentence. | × Asking for "engaging" or "compelling" The model hears "add adjectives." Say what should happen in the reader instead. |
× Writing 400-word prompts Structure beats length. If a prompt needs paragraphs of coaching, the task is too big. Split it. | × Letting it write the first sentence I write the opening myself now. Everything after it sounds more like me, because the model matches what it's given. |
The verdict, after months of this
I still write the first draft myself more often than not. What changed is everything around it. The reverse outline runs on every draft over a thousand words. The interview runs before anything I'm unsure how to start. The claim list runs before anything goes to an editor, and it has caught a statistic the model had rounded into a different number.
The drafting tactics are more fragile. A voice pack gets me maybe 70 percent of the way to a paragraph I'd sign; the rest is still me rewriting. Rejected drafts help, but only if I'm honest about why I cut something, which takes more thought than typing "make it better." On the days I don't have that thought in me, the model writes like a model.
If you take one thing: stop asking the model to be a good writer and start giving it things to push against. Your draft, your reader, your discards. The output gets specific when the input does. And keep your own file of prompts that failed. Mine is still the most useful document I own.
