Build Practical AI Systems
How to Build a Tone-of-Voice Reference File for AI
Build a short AI tone-of-voice reference from observable rules, approved examples and clear boundaries, then document one human-reviewed paragraph test.
Build a tone-of-voice reference file as a short set of observable writing rules. Give it a narrow scope, add two approved examples with notes, then record boundaries, a counterexample, an owner and a version.
This is a human-maintained reference, not permanent model training or a substitute for editorial approval. Use it beside the task and factual sources. Test one paragraph, keep the comparison and let a person approve its meaning, evidence and tone.
What is an AI tone-of-voice reference file?
An AI tone-of-voice reference describes how writing should sound for a defined use. It gives the AI-assisted process and human reviewer the same visible reference.
Here, voice means relatively stable writing characteristics. Tone means how they adjust to a situation. A service notice may be more concise than an article while both remain clear and measured. This is a practical distinction, not a universal definition.
The file has a narrower job than a general AI marketing context file.
| Input | Its job |
|---|---|
| Tone-of-voice reference | Reusable choices about expression |
| Task instruction | What the tool should do now |
| Source material | Facts, notes or copy to use or examine |
| Output template | Required fields or document shape |
A style example may contain facts, but that does not make it an approved factual source for a new draft.
The six-part structure below is an editorial starting method created for this lesson, not a provider standard or universal optimum.
1. Choose a narrow scope
Start with one useful kind of writing, such as educational B2B website copy. Record the audience and the purpose of that copy.
An all-channel file can collect conflicting rules. Begin narrowly, then add a labelled variation only when real work requires one.
2. Select two authorised examples
Choose two short samples that represent acceptable work within the scope. Two is a practical starting point here, not a universal optimum.
Record why each sample belongs. One might show an answer-first explanation; another a practical sequence. Remove unnecessary sensitive or time-bound details.
Use writing your organisation owns or has permission to reuse. UK government copyright guidance says original written work can receive automatic protection and that copyright restricts copying and adaptation. This is not legal advice, but do not use a competitor or creator's distinctive writing.
3. Turn traits into observable rules
Words such as professional, friendly and authentic leave room for many interpretations. Pair each chosen trait with visible choices.
| Trait | Do | Avoid |
|---|---|---|
| Clear | Answer the question early and explain necessary terms | Long scene-setting before the point |
| Measured | Qualify claims and preserve uncertainty | Guarantees, hype and false certainty |
| Practical | Give a concrete action or decision | Advice with no usable next step |
Add only rules that matter to the task and that a person can check in the text.
Google's current prompt guidance says examples can show patterns in phrasing, scope and format. Anthropic's current guidance also describes examples as signals for tone and structure and advises relevant examples that reduce accidental patterns. These are provider-specific recommendations, not a cross-tool promise.
4. Add boundaries and a counterexample
Record language to avoid, claims the tool must not invent, and the point where a person must review the work.
Add one short counterexample and explain why it fails. A line such as “Transform your marketing overnight with revolutionary AI that guarantees flawless content” is off-tone because it uses hype and unsupported certainty.
A compact sample file
A real reference should use authorised material. This excerpt is fictional teaching copy, not a real brand file or tested configuration.
Title: Tone-of-voice reference
Scope: Educational B2B website copy
Audience: Marketers beginning to use AI
Purpose: Explain one practical action without hype
[Voice rules]
- Clear: answer early; explain necessary terms.
- Measured: qualify claims; preserve uncertainty.
- Practical: give a concrete next step.
[Preferred and avoided patterns]
- Prefer: "can help", "test", "check", "approved".
- Avoid: "game-changing", "revolutionary", guarantees.
[Positive examples]
1. "AI can help a marketing team compare options. It does not decide which claim is true. A person checks the evidence and approves the wording."
Why: short explanation, measured claim, human control.
2. "Start with one repeated task. Record the input, expected output and review point. Test it before adding more steps."
Why: direct verbs, practical sequence, limited scope.
[Counterexample]
"Transform every campaign overnight with flawless AI content."
Why it fails: hype, universal scope and unsupported certainty.
[Hard boundaries]
- Do not invent claims, figures, quotations or results.
- Do not copy distinctive wording from an example.
- A person approves final meaning, evidence and tone.
Owner: [name or role]
Version: 0.1
Reviewed: [date]Use the file with the task and sources
Give the AI tool three visibly separate inputs:
- The current task, audience, channel and required output.
- The relevant part of the tone-of-voice reference.
- The approved factual source material.
Ask it to preserve the source meaning, follow the named rules and flag gaps rather than inventing details. How you supply the material depends on the current tool and approved process.
Test one paragraph and record the comparison
Keep the task and source paragraph the same. If practical, create one output without the reference and one with it. Record the date, file version, tool and model if known, both outputs, checks, human decision and next change.
Generated text can vary between runs, so one comparison does not prove a general improvement.
Illustrative comparison, not a model test: Both paragraphs below were constructed for this article.
| Check | Before review | After human review against the sample file |
|---|---|---|
| Paragraph | Unlock game-changing growth with our revolutionary AI workflow, guaranteed to transform every campaign. | An AI workflow can help a small marketing team repeat one defined task. Its usefulness depends on the inputs, review steps and work being tested. A person still approves the final decision. |
| Claim strength | Guarantee | Qualified statement |
| Specificity | Vague promise about every campaign | One defined task |
| Hype or vocabulary | Game-changing and revolutionary | Removed; plain description |
| Human responsibility | Missing | Explicit approval point |
The record makes the decision visible. It does not claim that AI produced the revision or that the file caused a measured result.
What should stay out of the file?
Leave out changing campaign instructions, unverified claims, private notes, credentials and weak examples. Do not turn the file into a complete brand manual or knowledge base.
If two rules conflict, resolve the conflict rather than asking the AI to choose silently. If an example becomes outdated or no longer represents approved writing, remove it through a deliberate review.
Your next step
Choose one narrow task and create the six parts using two authorised samples. Compare one paragraph with and without the reference, record the version and make one deliberate change after review.
Further Reading
You Might Still Be Wondering...
Frequently asked questions
No. A brand guide may cover identity, messaging, design and many channels. This is a smaller writing reference for one defined scope.
There is no universal length. Keep the smallest version that makes the important rules, examples and boundaries inspectable. Split it when unrelated channels create conflicts.
Two are a useful starting point for this exercise because they can show limited variation. Add or remove examples according to the real task and observed output. More examples are not automatically better.
Do not assume so. Keep stable voice rules where they genuinely apply, but label channel-specific adjustments. Test each important use case separately.
AI can suggest candidate patterns in approved writing, but a person should select the samples, define the rules, remove restricted information and approve the file.
No. Output can vary by tool, model, task, source material and update. Check the final meaning, evidence, tone and suitability every time.