Brand voice guidelines
Every template you can find for this is a blank container. A document with headings and nothing under them, a spreadsheet behind an email gate, a workbook you copy to your own drive. A blank is easy to make and does nothing on its own. The evidence: a two-year-old forum thread asking what to put in one still ranks.
So this page publishes the shape itself, field by field, with types and defaults. The schema is released under CC0, so you can take it and build your own thing with it without asking.
This page is the document half. Brand voice for AI covers the files each tool reads. If you would rather have this filled in than described, the tool will do it in about five minutes.
What is in a tone of voice document that a model can act on?
A useful document splits into two halves that get confused constantly. One half is orientation: what the organisation does, who it serves, what it wants to be to them. That half is written for people and it belongs in the document your team reads.
The other half is instruction: settings and prohibitions that decide individual sentences. That half is what a model can act on, and in most published guides it is a page or two out of fifty.
The test for which half a line belongs to is whether a checker could decide, without discussion, whether a given sentence breaks it.
One openly licensed guide has been put through that test in public. A real published style guide, compiled into machine-readable rules shows which of its lines survived the conversion and which would not compile at all, which is a more honest demonstration than an invented example.
Other published guides are collected the same way. The guides that are public, checked and dated have their mechanical rules pulled out and their dead links reported, so you can see how much of a fifty-page document is actually instruction.
Field types and defaults
| Field | Type | Default |
|---|---|---|
| brand_name | String, the named subject of every file | Required |
| brand_what | One sentence, 140 characters | Required |
| audience_primary | One of six reader types | Required |
| tone_formality | Forced choice, goes in as a worked example | Required |
| tone_humour | Forced choice, goes in as a worked example | Required |
| tone_irreverence | Forced choice, goes in as a worked example | Required |
| tone_energy | Forced choice, goes in as a worked example | Required |
| traits_never | List of prohibitions | Required |
| mech_person | We, I, by name, or mostly you | Required |
| mech_contractions | Always, never, or depends on surface | Required |
| locale_variant | One of six English variants | Required |
| mech_exclamations | Permitted, rationed, or forbidden | Rationed |
| mech_serial_comma | Boolean | Follows the locale |
| mech_heading_case | Sentence, title, or lower | Sentence |
| mech_em_dash | Permitted or forbidden | Permitted |
| lex_substitutions | Pairs: say this, not that | Empty |
| lex_mandated | Terms that must appear verbatim | Empty |
The eleven required fields are the ones asked first and they are enough on their own to produce a working file. Everything below them in the table is optional and refines that file. The optional fields come from two further rounds, for up to thirty-six questions in all. The page on the core questions gives the reasoning behind each required one.
Which fields are rules, and which are decoration?
The mechanical fields are rules. Person, contractions, locale, punctuation and vocabulary each decide something in every sentence, and two people applying them independently will agree.
Everything expressed as an adjective is decoration in this specific sense: it cannot be checked, so it cannot be enforced, so a model can appear to honour it while doing whatever it was going to do anyway. That does not make it worthless to a human reader. It makes it inert as an instruction.
The questionnaire exists to sort each line onto one side or the other. The core question set, word for word, with the reasoning for each records the question formats that were rejected as well as the ones that survived.
Why do adjectives not survive the trip into a model?
Because an adjective is a compression of decisions somebody already made and the model has to decompress it with no access to what was compressed. Warm might mean contractions and second person to you and emoji to the model.
The fix is not to avoid the word. It is to write down the behaviour that made you choose it. Confident usually decompresses to no hedging words. Warm often decompresses to contractions, second person, and short sentences. Write the decompression and the adjective becomes unnecessary.
This is also why the four tone questions here are pairs of written paragraphs rather than sliders: the paragraph you choose goes into your file as a worked example. An example needs no decompressing.
How long should it be?
Shorter than you think: fifteen binding rules. Anything past that is kept but marked non-binding. That is not a limitation of this tool. It was designed that way.
Compliance with a set of constraints degrades as the constraints stack, and vendors say so in their own documentation. Anthropic states that longer instruction files consume more context and reduce adherence and that if a rule keeps being ignored the file is probably too long. Cursor's anti-pattern list names copying entire style guides and tells you to use a linter instead.
Every documented length limit is collected here, including which of the widely quoted figures turn out not to exist.
Which parts a human reads and which parts only a machine reads
Both documents should exist and they should not be the same file. The human document holds positioning, audience, rationale and examples with commentary. The machine document holds the settings and prohibitions, in the format each tool reads, with nothing else in it.
Keeping them separate is what stops the machine file growing until it stops working. The manifest shows both, and every format the rules get rendered into.
Whichever of the two you circulate, somebody has to own it once the first week is over. Who owns the file, who reviews it and what a model release does to your rules are easy to leave until last. Settle them early.
What we left out, and why
- Brand archetypes. A diagnostic of perceived personality, not a source of writing rules.
- A tone spectrum with sliders. The most requested visual in the category and the format with the worst response quality for this audience.
- A voice score. There is no defensible one: within a single author's register, paragraph-level readability varies by several grades, so any per-paragraph verdict is noise larger than the effect it claims to measure. The limits page explains that refusal and the others.
- Anything that would push you past the fifteenth rule. A field whose only effect is to make the file longer makes the file worse.
Where these come from
Every claim above is quoted from one of these, and each was read on the date beside it. If one of them has changed since, the page is wrong and we would like to know.
- Anthropic: instruction length and adherence checked 2026-08-16
- Cursor: rules and the copy-a-style-guide anti-pattern checked 2026-08-16
- Mailchimp content style guide checked 2026-08-16
- GOV.UK style guide checked 2026-08-16
The schema on this page is what the tone of voice generator fills in and it hands you the completed document plus every machine-readable rendering of it.
Advertisement.
Also from Best Served Cold.
