How to Prompt for Tone When You Have No Examples to Give

The ideal way to describe tone to an AI model is through examples. Show it three paragraphs you consider representative of your voice and tell it to write in that style. The model can imitate patterns it can see far more accurately than patterns described in abstractions.

Most writers who are starting out with AI-assisted workflows do not have a clean sample document ready. They have published work, but not in a format that is easy to paste into a brief. Or they are writing in a new voice, for a new project, and do not yet have work in that voice to show.

When examples are not available, the description of tone has to do more work. The description can be made specific enough to be useful, but it requires thinking about the problem from a different angle.

What not to do

Do not describe the tone in positive terms only. "Direct," "engaging," "conversational," and "professional" are the same description that 70 percent of content briefs contain. They do not narrow the output space meaningfully. A model given only those terms will produce the statistical center of all writing described that way, which is unremarkable.

Positive descriptions of tone are also difficult to make specific without examples. What is the difference between "direct" and "blunt"? Between "conversational" and "informal"? Without an example to anchor the distinction, the model has to guess.

Using negatives to narrow the range

Negative constraints are more effective than positive ones when no examples are available. "Not academic" excludes a large category of writing style. "Does not use phrases like 'it is important to note' or 'one might argue'" gives the model something concrete to avoid. "Does not balance every claim with a counterargument of equal weight" produces different output than "opinionated."

A set of five specific negative constraints will narrow the output space more than ten positive descriptions. The model knows what not to do, and not doing those things leaves a smaller range.

Useful negatives for a practical, direct voice: no hedging phrases, no balanced-view framing on questions that have a right answer, no passive voice constructions where an active one is available, no introductory sentences that describe what the piece is going to argue, no bullet-point summaries at the end.

Describing sentence-level patterns

Sentence length patterns are something models can imitate from description when they cannot imitate from examples. "Short declarative sentences for main points. Longer sentences for explanation and qualification. Single-sentence paragraphs used occasionally for emphasis." That is specific enough to produce a noticeable effect.

You can also describe what is not there: "No subordinate clauses in the opening paragraph. No sentences that begin with a conjunction followed by a noun clause." These constraints sound like copy-editing rules, but they produce real changes in the rhythm of the output.

Describing the assumed reader

The register of a piece is largely determined by what the writer assumes the reader already knows and what the writer thinks the reader needs. Describing the reader specifically produces different output than describing the tone abstractly.

"Writes to someone who is already doing the work and does not need convincing that the work matters" produces a different register than "beginner-friendly." "Assumes the reader has tried basic approaches and is looking for what comes next" is a different register than "comprehensive guide."

The reader description is tonal information even though it appears to be audience information. The two are inseparable. Get the reader right and the register follows.

When none of this is enough

If the description-based approach is still producing output that requires extensive tonal correction, the sample document is worth the time to put together. Take three to five existing pieces you have written, clean them up into a single document, and put that document in the system prompt as an example corpus.

The description approach works well enough for writers who are not trying to preserve an existing voice but are building a new one for a new project. For preserving an established voice, examples are worth the setup cost.