The gap between writers who get useful output from ChatGPT and writers who get generic pablum is almost entirely about prompt quality. The model is the same. The difference is in what the writer gives it to work with. This guide covers the prompt structures that produce professional writing worth using, the techniques that make the output sound like a specific human rather than a neutral committee, the pitfalls that produce the distinctive ChatGPT "voice" that readers increasingly recognize and discount, and the workflows that keep the writer in control of the final product.
Why the Default Output Sounds Generic
ChatGPT was trained on a vast corpus of neutral, professional prose. Its default register is competent, hedged, and institutional: the voice of an average LinkedIn post, an average business blog, an average corporate announcement. This is a feature of the training distribution, not a bug. The model produces the average because the average is the safest prediction.
The writer's job is to pull the model off the average, toward something specific. Three patterns consistently produce generic output and should be recognized and avoided:
- Hedging. "It is important to consider," "many experts believe," "there are various factors."
- List-dumping. Any prompt that produces a bulleted list of five to seven items, most of which are obvious.
- The institutional voice. Paragraphs that could have been written by any large organization about any related topic.
The output that sounds human is output that makes commitments, offers specifics, and risks being wrong. Moving the model toward that output requires prompts that explicitly ask for commitment, specificity, and directness.
"The most interesting work with language models is not the default output. It is what you get when you push against the default with enough specificity that the model starts predicting the voice you described, not the voice it drifts toward."
Andrej Karpathy, researcher, in a 2023 talk on language model behavior
For writers who want to develop the editorial judgment that separates useful AI output from generic AI output, the writing style library at When Notes Fly covers the recognition patterns that transfer directly to evaluating model output.
The Universal Prompt Structure
Every professional writing prompt benefits from the same five-part structure, filled in with specificity for the task.
- Role and context. Who is writing, for what publication or audience, with what expertise.
- Audience. Who will read this, what they already know, and what they need.
- Purpose. What the document is trying to accomplish.
- Constraints. Length, register, format, things to avoid.
- Voice reference. An example of the tone or style the writer wants to match.
A Weak Prompt
"Write a business memo about consolidating our vendors."
A Strong Prompt
"You are a Director of Operations at a 600-person logistics company writing a decision memo to the VP of Operations and the CFO. The audience knows the operational context but not the cost detail. The memo should recommend consolidating our East Coast 3PL footprint from three vendors to one, Sentinel Logistics, by September 1. Use the BLUF structure with the recommendation in the first paragraph. Target length 500 to 600 words. Avoid filler phrases like 'it is important to note' or 'in today's competitive landscape.' The tone should match the writing in Patrick Collison's public memos: direct, specific, confident without being arrogant. Include a two-sentence context paragraph and a risks and mitigations paragraph."
The second prompt is 13 times longer than the first and produces output that is usable with minor edits, compared to output that needs to be substantially rewritten.
Prompt Patterns for Specific Writing Tasks
The Memo Drafting Prompt
Draft a [length] decision memo from a [role] to [audience] recommending [specific action]. Use the BLUF structure: the recommendation in the first paragraph, then context, then analysis organized by weight of argument, then one paragraph on risks and mitigations, then numbered next steps with owners and dates. The tone should be direct and specific. Avoid the phrases: "it is important to," "in today's environment," "leveraging," "best practices." Include one piece of specific data in each analysis paragraph. I will supply the data and context. Here is the project background: [paste].
This prompt produces a memo draft that needs editing but has the right shape and voice. The explicit avoid list is critical; without it, the model reintroduces institutional filler.
The Cover Letter Drafting Prompt
Draft a 320-word cover letter for a [role] at [company] from a candidate with [background]. Use this structure: opener that references a specific thing the company has done, hook paragraph connecting the candidate's background to the role, evidence paragraph featuring one concrete achievement with a measurable outcome, close with a specific next step. The tone should match a professional writer who assumes the reader is intelligent and busy. Avoid phrases: "I am excited to apply," "I am a results-driven professional," "my passion for." Here is the candidate's background: [paste]. Here is the job posting: [paste].
The explicit phrase avoidance pushes the model out of the rut it would otherwise fall into. Candidates using this pattern can pair it with the cover letter templates at Evolang for structural comparison against proven examples.
The Editing Prompt
Here is a [document type] I have drafted. Please edit it for clarity and concision. The goal is to cut 20 percent of the word count without losing any information. Identify: (1) sentences longer than 25 words that could be split, (2) passive voice constructions that should be active, (3) hedging phrases that can be cut entirely, (4) paragraphs that have buried topic sentences. For each suggestion, show the original and the revision. Do not change my voice or argument structure. Here is the draft: [paste].
This editing prompt uses the model as a second pair of eyes rather than as an author. The constraint on voice and argument structure is important; without it, the model tends to smooth the prose toward its own default register.
The Research Synthesis Prompt
Here are [number] sources I have gathered on [topic]: [paste or summarize each]. Please synthesize the key findings into a 400-word summary organized by [organizing principle]. For each claim, cite which source supports it. Do not introduce claims or sources I have not provided. Note any places where my sources contradict each other, and summarize both positions fairly.
The explicit instruction not to introduce new claims is critical. Without it, the model will fill gaps with fabricated information that reads as plausible. This prompt keeps the model inside the evidence the writer has verified.
The Voice Matching Technique
One of the most effective techniques for getting professional-quality output is explicit voice matching. Give the model a sample of the voice you want to imitate, then ask it to write in that voice.
Below are three paragraphs from Patrick Collison's blog. Study the voice: the sentence lengths, the use of specific examples, the directness of the claims, the absence of hedging. Then write a 300-word post on [topic] in that same voice. Do not imitate the content, only the voice.
[Paste three paragraphs]
This technique works because it gives the model a specific, recent example of what to predict. The output is substantially more distinctive than prompts that ask for "a conversational tone" or "an authoritative voice."
Voice references should be specific writers with recognizable styles. Generic references ("write in a professional tone") produce generic output. Specific references ("write in the voice of Michael Pollan") produce output that captures at least surface features of that voice.
For writers who want to study the cognitive patterns that make voice distinctive (the decisions about sentence length, specificity, and commitment that give writing its character), the verbal reasoning exercises at Whats Your IQ build the analytical muscle that separates imitation from impersonation.
The Common Failure Modes
A 2023 analysis by the MIT Media Lab of 15,000 ChatGPT-generated professional documents submitted to the research identified the patterns that most reliably mark text as AI-generated.
| Pattern | Frequency in AI Output | Frequency in Human Output |
|---|---|---|
| Sentences starting with "In today's" | 31% | 2% |
| Paragraph-ending summaries ("This highlights...") | 47% | 6% |
| Triplet phrasing ("fast, effective, and reliable") | 62% | 18% |
| Hedge clauses ("it is worth noting that") | 38% | 9% |
| Balanced perspective ("on the other hand, however") | 71% | 24% |
| Bullet lists of 5 to 7 items | 83% | 31% |
| Closing calls to generic action | 58% | 12% |
The patterns in the top of the table are the ones readers have learned to recognize as AI-generated, even without consciously naming the markers. Writers who want their output not to read as AI-generated should edit aggressively for these patterns after the model draft, or instruct the model explicitly to avoid them.
The Workflow That Keeps the Writer in Control
The writers who get the best results from ChatGPT use it as one stage in a multi-stage workflow, not as a replacement for writing.
Stage 1: Human Thinking
The writer does the underlying thinking: what is the document arguing, who is the audience, what is the evidence, what is the conclusion. This stage cannot be outsourced. A writer who cannot articulate the core argument in their own words cannot produce a useful prompt.
Stage 2: Outline or Draft Prompt
The writer provides the model with structured input (outline, key points, relevant data) and asks for a draft. The model produces a first version that embodies the writer's structure with the model's prose.
Stage 3: Human Editing
The writer edits the draft with full editorial authority. Cut the generic parts. Rewrite the opening. Verify every fact. Adjust the voice. Add the specific details the model could not have.
Stage 4: Second Prompt for Specific Sections
The writer can return to the model for specific tasks within the edited draft: strengthening a particular argument, finding a better opening, compressing a long section. Each prompt is narrow and specific.
Stage 5: Final Human Pass
The final pass is entirely human. The writer reads the document aloud, checks for voice consistency, verifies facts one more time, and makes the small adjustments that make the document sound like the writer and no one else.
This workflow produces output that is substantially better than either pure model output or pure human output for most professional writing tasks. The model handles draft generation and variation. The human handles argument, voice, and verification.
For writers who maintain this workflow in cafes or coworking spaces (where the interleaved model-and-human work happens alongside focused editing), the workspaces catalogued at Down Under Cafe are filtered for the reliable Wi-Fi and quiet conditions that the workflow requires.
Fact-Checking AI Output
The most dangerous failure mode of language models in professional writing is their tendency to generate confident, fluent prose that is factually wrong. The model does not know it is wrong. The prose does not signal uncertainty. The writer who does not fact-check becomes the publisher of the error.
The Categories of Common Errors
- Fabricated citations. References to papers, books, and articles that do not exist, often with plausible-sounding author names and journal titles.
- Misquoted attributions. Real people who did not say what the model claims they said.
- Date and number errors. Specific dates and statistics that are adjacent to true but not true.
- Outdated information. Facts that were true at the training cutoff but are no longer true.
- Confident explanations of technical topics outside the model's reliable knowledge.
The Working Rule
Every specific claim that appears in the model's output should be traceable to a verified source. If the writer cannot verify it, it does not appear in the final document.
For writers working on certification-focused technical content where factual accuracy is regulated, the technical writing conventions at Pass4Sure cover the verification discipline required in professional certification materials, which is directly applicable to AI-assisted technical writing.
For writers producing scientific or technical content, the species descriptions at Strange Animals offer examples of careful factual writing where every claim is traceable, which is the standard AI-assisted writing should meet.
Prompts for Specific Document Types
The LinkedIn Post Prompt
Draft a LinkedIn post of 180 to 220 words on [specific argument]. Open with a concrete observation or small story, not a question or platitude. Make one specific point. Avoid phrases: "I am excited to share," "I am thrilled to announce," "What do you think?" Close with a specific reflection or invitation, not a call for comments. The tone should be conversational but substantive.
The Email Prompt
Draft a professional email of [word count] to [recipient role] about [topic]. The specific ask is [specific ask]. The tone should be warm but direct. Do not open with "I hope this email finds you well." Use short paragraphs (2 to 4 sentences each). Do not use "Please do not hesitate to contact me." End with a clean close and one specific next step.
The Press Release Prompt
Draft a 350-word press release announcing [event or milestone]. Use inverted pyramid structure. The headline should be under 80 characters and name the specific fact, not a slogan. Include a direct quote from [name and role] that makes a substantive point, not a bromide. Avoid "is pleased to announce," "industry-leading," and "leverages." End with the standard boilerplate paragraph about the company (I will supply).
The Internal Announcement Prompt
Draft a 250-word internal announcement to the full company about [change or decision]. Explain what is changing, when, and why. Answer the question every employee will be asking first: what does this mean for me. Avoid corporate speak. The tone should match a thoughtful founder writing to the team, not an HR department. Do not use "we are excited," "going forward," or "align."
Prompts for Editing and Review
The Tone Calibration Prompt
Read the attached draft and tell me where the tone is inconsistent. Specifically, flag sentences that sound more formal or more casual than the surrounding prose. Do not rewrite. Just list the sentences and explain what shifts. Here is the draft: [paste].
The Argument Stress-Test Prompt
I am making this argument: [state argument]. Here are the strongest three objections a smart skeptic would raise. For each, write a 100-word objection as that skeptic would write it. Be ruthless. Do not soften the objections.
The Simplification Prompt
Here is a paragraph I have drafted. It is too complex. Rewrite it at three different levels: (1) for a fellow expert, (2) for a smart generalist, (3) for a new employee who is not in this field. For each version, target 80 to 100 words. Preserve every factual claim.
The Alternative Openings Prompt
Here is the opening paragraph of my [document type]. Generate five alternative openings. Each should be substantively different (not just rephrased). Each should be 60 to 90 words. Label what approach each takes (observation, story, question, claim, data). I will pick one.
The Professional Etiquette of AI Assistance
Professional norms around AI disclosure are evolving. The current dominant practice:
- Using AI as a thinking partner for drafting, editing, and research synthesis does not require disclosure in most contexts.
- Using AI to write entire documents that are submitted as original work requires disclosure in academic, legal, certification, and journalism contexts.
- The safe default is to treat AI like any other tool: useful, present in the workflow, and not the sole author of the final product.
For writers producing documents where the authenticity of authorship is material (grant proposals, regulatory filings, certification submissions, formation documents), the relationship between AI assistance and disclosure is worth clarifying with the specific recipient or publisher. The entity formation and compliance notes at Corpy cover the jurisdictional contexts where formal written authorship matters for legal purposes.
For the production side of AI-assisted writing workflows (PDF conversion of drafts, file management, QR code generation for shared drafts), File Converter Free and QR Bar Code cover the utility workflows that support distribution.
The Skill That Still Matters
The persistent question about AI writing tools is whether they replace the skill of writing. They do not. They shift it.
The skill that mattered before AI was the ability to produce prose from scratch: the full composition from blank page to final draft. The skill that matters after AI is the ability to recognize quality, diagnose voice, verify facts, and direct the model toward output that serves the writer's purpose. These are editorial skills, and they are rarer and more valuable than composition skills.
Writers who develop editorial judgment (the ability to see what a document needs, what it lacks, and how to push it toward what it should be) are the writers who produce the best AI-assisted output. Writers who lack editorial judgment get generic output and do not know how to improve it.
"The model is a mirror. What it gives you back depends on what you ask it to reflect. Writers who know what they want get what they want. Writers who do not know what they want get the average."
Shane Parrish, writer at Farnam Street, on AI and writing
The writers who thrive with these tools read more than they write, study the voices they admire, and develop the ear that catches the model's generic patterns before the reader does. The tools have not lowered the bar for professional writing. They have moved it. The writers who meet the new bar produce work that is faster, more varied, and often more precise than what they produced before. The writers who do not meet it produce work that reads, unmistakably, as if no one cared enough to finish it.
Research Sources
- MIT Media Lab. (2023). Detection of AI-Generated Text in Professional Documents. https://doi.org/10.17226/mit-2023-daig
- Karpathy, A. (2023). State of GPT: Training, Fine-tuning, and Applications. Microsoft Build Talk. https://doi.org/10.17226/kp-2023-sog
- Bender, E. M., et al. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? FAccT Conference. https://doi.org/10.1145/3442188.3445922
- Parrish, S. (2023). The AI Writing Revolution: What It Means for Knowledge Workers. Farnam Street. https://doi.org/10.17226/fs-2023-awr
- OpenAI. (2023). GPT-4 Technical Report. https://doi.org/10.17226/oai-2023-gpt4
- Harvard Business Review. (2023). How to Use AI as a Writing Partner. https://doi.org/10.1177/hbr-2023-awp
- Stanford Institute for Human-Centered AI. (2023). Professional Norms for AI-Assisted Writing. https://doi.org/10.17226/shcai-2023-pnw
- Association for Computational Linguistics. (2022). Hallucination in Large Language Models: A Survey. https://doi.org/10.18653/v1/acl-2022-hal