AI for writing
Keep a consistent voice and style
Voice is a pattern of choices tied to purpose and authorship, not a celebrity name or a handful of adjectives.
Before you start
Why this matters
Choose two short pieces you wrote for different situations. Underline choices that make them sound like you: sentence rhythm, directness, examples, humor, level of formality, or how you address the reader. Which choices are stable across both pieces, and which changed because the audience or purpose changed? Write three observable style rules without using “professional,” “engaging,” or “authentic.”
1Learn the idea
Read
Distinguish voice, tone, and style
These terms overlap, but separating them improves prompts and review.
Voice is the recognizable stance and pattern across a body of writing: curious, evidence-first, direct, reflective, or practical. Tone is the attitude appropriate to a specific moment: reassuring during a disruption, neutral in an incident report, energetic in an invitation. Style is the set of observable language and formatting choices that creates voice and tone.
“Sound confident” is ambiguous. “State the recommendation in the first paragraph, use active verbs, acknowledge one limitation, and avoid hype” can be reviewed. Convert labels into actions.
AI models tend toward generic smoothness: balanced paragraphs, familiar transitions, broad claims, and predictable conclusions. Without constraints, repeated use can flatten the distinct decisions that make writing useful. The answer is not adding more decorative personality. It is making your real communication standards visible.
Read
Create a voice card
A voice card is a short, maintainable guide. Include:
- Stance: how the writer relates to reader and subject;
- Principles: recurring priorities such as evidence before assertion;
- Sentence behavior: typical length, variation, directness, and point of view;
- Vocabulary: preferred terms, terms requiring definition, and words to avoid;
- Examples: patterns for openings, explanations, transitions, and endings;
- Tone range: how the voice changes for celebration, uncertainty, disagreement, or bad news;
- Anti-patterns: habits that would make the text feel wrong.
For example:
STANCE
Practical peer: informed but not all-knowing.
DO
Lead with the reader's task. Use concrete examples. Name uncertainty.
Mix concise statements with explanatory sentences. Prefer "you" where useful.
DON'T
Use hype, fake quotations, rhetorical questions in every opening, or
"In today's fast-paced world." Do not call a step easy without evidence.
TONE RANGE
For risks, become calmer and more precise, not dramatic.
Keep the card short enough to use. A fifty-page guide may be a reference manual, but it is a poor prompt.
Read
Use examples without imitation
Examples communicate patterns better than adjectives, but choose them responsibly. Use your own writing or material you are allowed to process. Remove confidential and personal information. Do not request “write exactly like” a living author, colleague, or public figure. That substitutes imitation for understanding and can misrepresent identity.
Instead, extract high-level features: restrained metaphors, short openings, concrete verbs, explicit limitations, or a question-and-answer structure. Ask the model to explain which features it applied. Then decide whether those features serve this document.
One sample is not a complete voice. It may represent an unusual audience or mood. Select several approved excerpts and distinguish stable patterns from situation-specific choices. Never assume a model can authenticate authorship or perfectly reproduce a person.
Read
Run style as a separate pass
First settle purpose, structure, claims, and clarity. Then apply the voice card with a preservation contract:
Revise this clear, verified draft using the voice card.
Preserve facts, logic, uncertainty, conditions, quotations, terminology,
citations, and requested action. Do not add anecdotes, opinions, or evidence.
Return:
1. revised text;
2. a brief list of voice-card choices applied;
3. any sentence where style and precision conflicted.
Separating passes reduces the chance that a tone request rewrites substance. Review every conflict. A punchier sentence may overstate causality. A warmer sentence may add praise the author does not mean. A more decisive ending may convert an option into a commitment.
Read
Adapt without becoming inconsistent
Consistency does not mean sameness. The same organization may write a tutorial, incident update, policy explanation, and campaign page. Build a stable core plus context controls.
The stable core might require direct openings, respectful language, evidence-linked claims, and honest uncertainty. Context controls determine formality, detail, urgency, and emotional temperature. During a service outage, clarity and timing matter more than wit. In an onboarding guide, examples and encouragement may be useful.
Define tone boundaries. “Warm” does not permit invented empathy or familiarity. “Urgent” does not permit panic or false deadlines. “Persuasive” does not permit hiding trade-offs. “Simple” does not permit removing important conditions.
Read
Audit for style drift
After several sections are assembled, inspect:
- changes in point of view or formality;
- multiple labels for the same concept;
- repeated generated phrases and transition clichés;
- sudden changes in sentence rhythm;
- unjustified certainty or enthusiasm;
- examples that sound unlike the author or audience;
- endings that summarize without advancing the reader’s task.
Search repeated phrases, but judge them in context. Necessary terminology should stay consistent. Accidental formula—“It’s important to note,” “Ultimately,” or repeated three-item lists—can be revised.
The final test is not whether a classifier calls the text human. Such detectors are unreliable and do not establish authorship. Ask whether you can defend every choice, whether the voice supports the reader, and whether the document truthfully represents its author. Disclose AI assistance when policy or context requires it.
Continue learning · glossary & guides
- How do voice, tone, and style differ?
- Why are observable rules stronger than adjectives?
- How can examples be used without imitating a person?
- Why should style follow clarity and verification?
- What does a stable core plus context controls provide?
- Glossary: responsible AI · Cheatsheet: prompt recipe