AI Writing Prompts That Actually Work (and Why Yours Don’t)

A prompt is a recipe, and AI is a line cook who has memorized every dish ever written down. Hand that cook a note reading “make food” and you get the blandest average of everything: edible, forgettable, identical to what everyone else gets. Hand the same cook a recipe with the dish, the portion, the ingredients to use and the ones to leave out, and the output has shape. The cook didn’t get smarter. Your instruction did.

I lose patience fast with people who type six words into ChatGPT, get mush back, and announce that AI can’t write. Of course it can’t write from six words.

The tool did exactly what it was told, and the person blaming it never said what they wanted. Give it a real instruction and it’ll do real work for you. Ask it to do your thinking for you and you’ll get back something nobody wants to read.

Why is most AI writing output terrible?

AI writing tools like ChatGPT train on enormous datasets of existing text. Give them a vague prompt and they produce the statistical average of everything written on that topic. That’s why AI output sounds generic. It’s generic: the average of millions of documents compressed into a few paragraphs.

That average runs on passive voice, corporate jargon, hedge words, and the throat-clearing openings professional editors cut on sight. When a draft opens with “In today’s fast-paced world,” nobody told the AI what to say, and paying for a fancier model won’t cure that. Fix the prompt. Give the tool specific constraints, a defined audience, a clear purpose, and examples of what you want, and the output moves from unusable to workable.

Workable, not publishable, and the difference matters. AI output is a starting point. I develop that further in what AI does well and where it wrecks your book.

Watch: Google's 9 Hour AI Prompt Engineering Course In 20 Minutes, Tina Huang

What are the five elements of a good prompt?

A prompt that produces usable output has five elements. Role tells the AI who it’s writing as.

“Write as a financial advisor explaining retirement planning to small business owners” produces different output than “write about retirement planning,” because the role sets expertise level, vocabulary, and perspective. Audience answers who’s reading. A blog post for first-time entrepreneurs reads differently than a white paper for venture capitalists, and naming the audience forces the AI to calibrate complexity, tone, and what the reader already knows.

Format sets the shape: an email, a chapter outline, a product description, a speech, a case study. Naming it stops the AI from falling back on its favorite structure, usually a five-paragraph essay with a generic conclusion.

The last two elements are where most people fail. Constraints tell the AI what to avoid, and telling it what not to do often works better than telling it what to do. “No jargon, no passive voice, no sentences over 20 words, no opening with ‘In today’s world'” kills the worst AI habits before they appear. Back in late 2024, when I wanted a quick sanity check, my prompt was short: does this make sense? Don’t give me critique, just tell me yes or no and flag any glaring problems. Without that line, the models of the day would hand me 15 pages of stuff.

Examples close the gap. If you have a sample of the style you want, include it. AI is excellent at pattern-matching, and showing it a paragraph in your target voice and asking it to continue moves the output from generic to specific on its own.

Most people skip constraints and examples, because writing them forces you to decide what you want, and that decision is the one part the machine can’t make for you. My AI-enhanced writing handbooks, more than forty of them, are built on exactly this kind of prompt. They help a writer build a sharper protagonist or a tighter plot, and the writer still writes the book.

I have Claude write my image prompts. I tell it what I want in the picture, and it hands me a prompt for Leonardo. The first image is never quite right, so I give Claude the image and tell it what’s off. It writes a new prompt. After a few rounds of that, I get the image I wanted.

What do ten strong prompts look like?

These show the difference between vague instructions and structured ones, and each includes role, audience, format, and constraints.

1. Business book chapter outline. “Create a chapter outline for a book about scaling a consulting practice from solo to seven figures. The audience is experienced consultants earning $150K-$300K who want to grow but don’t want to manage a large staff. Each chapter should address a specific growth obstacle with a concrete solution. No motivational filler. No generic business advice. Every chapter title should promise a specific outcome.”

2. Executive bio. “Write a 150-word professional bio for a CFO who spent 20 years in manufacturing before moving to tech. The bio should emphasize the unusual career transition as a strength. Tone should be confident but not arrogant. Third person. No buzzwords like ‘passionate,’ ‘visionary,’ or ‘thought leader.’ End with something specific about their current role, not a generic statement about their commitment to excellence.”

3. Email sequence for a service business. “Write a three-email nurture sequence for a high-end interior design firm targeting homeowners who just purchased homes over $1 million. Email one introduces the firm’s approach. Email two addresses the most common mistake new homeowners make with renovations. Email three offers a free consultation. Tone is knowledgeable and direct. Each email should be under 200 words.”

4. Blog post opening. “Write three different opening paragraphs for a blog post about why most business plans fail. Each opening should use a different technique: one starts with a specific statistic, one starts with a brief story, one starts with a counterintuitive claim. No opening should use the phrase ‘in today’s world,’ ‘it’s no secret,’ or ‘have you ever wondered.’ Each opening should be under 60 words.”

5. Case study. “Write a 300-word case study about a mid-size accounting firm that reduced client onboarding time from 3 weeks to 4 days by restructuring their intake process. Use the structure: problem, approach, specific result. Include at least two specific numbers. Tone is professional but not stiff. No jargon. The reader should finish understanding exactly what changed and why it worked.”

6. Book back cover copy. “Write back cover copy for a memoir by a former combat medic who became a trauma surgeon. The book covers both careers and the psychological cost of each. Tone should be serious but not melodramatic. Under 150 words. The copy should make the reader feel the tension between saving lives and the toll it takes. Don’t use the words ‘journey,’ ‘inspiring,’ or ‘powerful.'”

7. LinkedIn post. “Write a LinkedIn post about why hiring for ‘culture fit’ often means hiring people who look and think like the existing team. The post should be under 150 words, written in first person, conversational tone. Include one specific example. No hashtags. No emoji. End with a question that invites genuine responses, not engagement bait.”

8. Speech opening. “Write the opening 90 seconds of a keynote speech for a cybersecurity conference. The speaker is a former hacker who now runs a security consulting firm. The opening should grab attention without relying on a generic ‘imagine this’ scenario. Use a real-world example of a security breach that the audience would recognize. Conversational tone. No slides referenced.”

9. Product description. “Write a product description for a handmade leather journal targeted at writers who prefer analog tools. The description should emphasize the tactile experience. The feel of the leather, the weight of the pages, the sound of the binding. Under 100 words. No superlatives like ‘best,’ ‘finest,’ or ‘premium.’ Let the sensory details sell the product.”

10. Query letter. “Write a query letter for a thriller novel about a forensic accountant who discovers that a major pharmaceutical company has been laundering money through a children’s hospital charity. The letter should follow standard query format: hook, brief synopsis, comparable titles, author credentials. Under 300 words. The hook should establish stakes in the first two sentences. Comparable titles should be recent thrillers published within the last five years.” The same specificity applies whether you’re drafting leadership content or a romance novel.

What should you do with AI writing output?

Treat every piece of output as a draft you still have to rewrite. Check every fact, cut the filler, and put it back in your own voice. People who skip that step and publish whatever the machine hands them are the reason readers now roll their eyes at anything that smells like AI, and they’re making it harder for everyone who uses the tool well.

From the podcast

Martin Ricketts watches people chase every new tool instead of getting better at the one they have, and he’s direct about it on Leaders and Their Stories: Instead of this tool or that tool, the prompt is infinitely more important than which tool you’re using. It’s like when I say this thing, tactics matter more than tools. How you use the tool is the really important thing. The prompt is the part you control and the part that transfers to whatever tool you use next year.

Even good prompts produce output that needs work. AI doesn’t know your voice, your audience’s specific concerns, or the details that make content credible. Treat the output the way you’d treat a rough draft from a research assistant: useful raw material that needs a professional writer’s judgment before it can be published. My own prompts stay simple. I ask for a report, and when it misses I type back in plain words: that’s not exactly what I wanted, focus more on this, make it 500 words, you’ve used too many fancy words for business people.

Check every fact. AI fabricates statistics, invents quotes, and cites sources that don’t exist. If the output includes a number, verify it. If it includes a name, verify it. If a claim sounds too perfect, it probably is. Strip the AI voice. Even with good prompts, AI falls back on patterns readers now recognize: throat-clearing openings, hedge words like “it’s important to no​te,” synonym cycling, and conclusions that restate the introduction. Cut those and put specifics in their place.

Back in 2024, I used ChatGPT to make sense of meeting transcripts. I’d paste in the whole transcript with one prompt: make sense of this. A two-hour meeting ran past the length limit, so I gave it the transcript in two pieces. Then I’d interrogate it for the to-dos and the agreements. You have to be careful, though, because sometimes it lies, so I checked everything it gave me. These days I use Claude.

The Chicago Sun-Times learned this in public. On May 18, 2025, its Sunday paper carried a syndicated summer insert called Heat Index, and it included a reading list of 15 books. Ten of them don’t exist. Isabel Allende never wrote a climate novel called Tidewater Dreams, and Percival Everett never wrote The Rainmakers. The same list ran in at least one edition of The Philadelphia Inquirer.

Freelance writer Marco Buscaglia had generated the list with AI for King Features, a Hearst syndication unit, and never checked it. “Huge mistake on my part and has nothing to do with the Sun-Times,” he told NPR, adding, “It’s on me 100 percent.” According to the Sun-Times, King Features ended its relationship with him, and the paper didn’t charge print subscribers for that edition.

I’ll give him credit for owning it. It still cost him the client, and his name is now tied to those fake books in every story about AI making things up. A few minutes checking each title would have caught all ten.

Skip the fact check and your name goes on the invented statistic. When a reader catches it, nobody blames the software. They blame the author, and they stop trusting everything else on the page.

Then add what AI can’t: your expertise, your experience, your opinion, and the details that come from knowing the subject. AI produces structure. Insight is still your job. This is the same discipline behind making AI-generated content feel real and avoiding the phrases that mark writing as machine-made. The AI writing hub covers the rest, and if you want a book that uses AI the right way with your voice intact, that’s the work I do.

If your AI output is garbage, look at what you typed before you blame the software. A lazy prompt gets you lazy writing, and publishing lazy writing under your own name is a choice you made, whatever tool you used. Take ten minutes to write a real instruction, then do the rewrite. Skip either one and you’re adding to the pile of filler readers already hate.

From the author

The Birth of the Augmented Human
Good prompts are one piece of using AI properly. I wrote The Birth of the Augmented Human about how AI, used correctly, amplifies your thinking instead of replacing it.

Frequently Asked Questions