Getting the Right Tone From AI Writing
The most common complaint about AI writing isn't accuracy — it's voice. The output is correct, fluent, and completely generic: the same polished-but-lifeless tone whether it's supposed to be a founder's memo or a friendly onboarding email. The good news is that tone is one of the most controllable aspects of AI writing, once you know the levers.
This guide covers the practical techniques: feeding the AI examples of the voice you want, writing constraints that actually work, iterating instead of accepting first drafts, and the editing passes that turn decent AI output into writing that sounds like you.
Give the AI two or three examples of the target voice, name the tone explicitly with do-and-don't constraints, generate multiple options and combine the best parts, and always do a final editing pass in your own voice.
Show, don't just tell, the voice you want
"Write in a friendly tone" produces generic friendliness. Pasting an actual example of the voice you want produces something much closer. Before asking for the draft, provide two or three samples: a past email, a paragraph from your blog, the tone of a brand you admire. Then describe what's distinctive about it — short sentences, dry humor, no jargon, whatever you observe.
This works because AI models are pattern-matchers first. A description of a tone is abstract; examples are concrete patterns to continue. The combination — examples plus your articulation of what makes them work — is dramatically more effective than either alone. Save your best-performing examples as a reusable style reference.
Choose examples that match the task at hand. The voice for a technical tutorial and the voice for a fundraising email are different genres, and feeding the model a blog post when you need a sales page will pull the output toward the wrong register. Keep two or three example sets for your most common writing types — it's a small library that pays for itself every time you write.
Write constraints the model can follow
Constraints beat adjectives. "Professional but approachable" is vague; "sentences under 20 words, no buzzwords, address the reader as 'you', one concrete example per section" is executable. Think about what you'd tell a human writer: format, length, vocabulary to avoid, how to open and close.
A particularly effective constraint is the anti-example: "never use phrases like 'delve into,' 'in today's fast-paced world,' or 'it's important to note that.'" AI writing has recognizable tics, and naming the ones that bother you eliminates them in one move. Keep a running list of banned phrases in your style reference and it compounds in value.
Length constraints deserve special attention because AI defaults to thorough. "Keep it under 150 words" or "three paragraphs maximum" forces the prioritization that long outputs avoid — the model has to decide what matters instead of covering everything. If the result feels thin, you can always ask for expansion on the specific point that needs it; starting short and expanding beats starting bloated and cutting.
Generate options, then combine
Never accept the first draft. Ask for three versions with different angles — or the same brief run three times — and you'll almost always find that version two has the best opening and version three has the best close. Combining the strongest parts of multiple outputs consistently beats any single generation.
Iterate on sections, not whole pieces. "Rewrite just the introduction to be more direct" preserves what works while fixing what doesn't. Full rewrites throw away the good with the bad and introduce new randomness. Treat the AI like a junior writer whose good paragraphs you keep and weak ones you send back with notes.
When a section resists iteration — you've rewritten it three times and it's still off — the problem is usually the brief, not the model. Step back and rewrite your instructions for that section with more specificity: add an example, name the structure you want ("problem, then consequence, then solution"), or dictate the opening sentence yourself and have the AI continue from it. A strong opening sentence constrains everything that follows.
Finish with your own editing pass
The final 10% is yours and it matters most. Read the draft aloud — awkward AI phrasing reveals itself immediately in speech. Cut the throat-clearing openings ("In the world of..."), replace abstract nouns with concrete ones, and add the specific details only you know: the anecdote, the number, the opinion.
This pass is also where you catch the subtle wrongness: confident statements you should verify, transitions that don't quite connect, examples that feel stock. AI gets you from blank page to workable draft fast; your editing gets it from workable to yours. Skip this step and readers can tell — they've all read enough AI text to recognize the texture.
Turn it into a reusable style guide
Once you've found constraints that work, don't re-derive them every time — write them down as a style guide you paste into every writing session. A good one-page guide includes: the voice description with two or three examples, the banned-phrases list, formatting rules, and audience notes. This is the document that turns "I got lucky with that prompt once" into a repeatable process.
The guide gets more valuable the more you use it, because every editing pass teaches you something new about your preferences. "I always cut the second paragraph" becomes a constraint: "open with the point, no throat-clearing." "My introductions are always too long" becomes "introductions under 60 words." Your editing frustrations are the raw material — the guide is just where you store what they taught you.
If you write with a team, the style guide becomes shared infrastructure. One person's discovered constraints ("never start with 'in today's') save everyone the same discovery process. And when someone new joins, the guide plus examples onboards them to the team's voice faster than any amount of verbal description. What started as prompt engineering becomes an editorial asset.
Frequently asked questions
Why does AI writing all sound the same?
Models default to the statistical middle of their training data — competent, neutral, unobjectionable. Distinctive voice requires distinctive input: examples, constraints, and editing. Without those, you get the average of everything.
Can AI match my personal writing voice?
Close enough for drafts, with good examples and iteration. It won't fully replicate you — your lived experience and opinions aren't in the prompt. Use it to get 80% there fast, then close the gap yourself.
What are the most useful style constraints?
Sentence length limits, banned phrases, specified openings and closings, "one example per section," and audience definition ("write for busy founders who hate fluff"). Concrete and checkable beats abstract every time.
Should I disclose that AI helped with the writing?
Follow your context's norms: many publications and workplaces now expect disclosure, and some have specific policies. When in doubt, disclose — and remember that you're responsible for everything in the final piece regardless of who drafted it.