The template
It works in ChatGPT, Claude, Codex, or any tool that accepts written instructions.
# AI-ready brand voice guidelines
Version: 1.0
Owner: [person accountable for final approval]
Last reviewed: [YYYY-MM-DD]
## 1. Purpose
Write for [primary reader] so they can [specific next understanding or action].
## 2. Source hierarchy
When sources disagree, use this order:
1. Current approved instruction for this message
2. Current approved writing in this channel
3. Verified facts and approved claims
4. Older examples that still match the work
5. Generic style guidance
If a higher source is missing or conflicts with a lower source, pause and ask. Do not invent the answer.
Accuracy, privacy, and the claim rules in section 5 override this order. An example sets style. It never supplies a fact.
## 3. Content Origin Brief
Before drafting, state:
- Origin: [the real note, fact, request, or earlier message this draft is based on]
- Who is speaking: [person or team]
- Who is reading: [specific reader]
- Why this matters now: [real trigger]
- What the reader should understand or do: [outcome]
- Evidence available: [links, facts, examples]
- Channel and constraint: [email, post, page, memo, length]
- Earned detail: [one sourced detail that makes it specific]
If there is no real origin, ask for one or write a plain factual version. Do not invent a stake, memory, opinion, or feeling.
## 4. Voice decisions
- Prefer: [specific choices shown in approved examples]
- Avoid: [specific phrases, tones, or structures]
- Sentence rhythm: [short, varied, direct, other]
- Level of certainty: [what needs evidence or a caveat]
- Personal detail: [what is appropriate and what is private]
## 5. Claims and boundaries
- Use only verified facts supplied for this message.
- Mark an inference as an inference.
- Do not add a personal anecdote, result, client story, or product claim without a source.
- Ask when the sender, audience, promise, or evidence is unclear.
## 6. Channel rules
For [channel], use [structure or required details].
Keep [required opening, call to action, or format].
Do not use [channel-specific failure mode].
## 7. Review before sending
Check:
- The reader and outcome are still clear.
- Every material claim has a source.
- The draft sounds like the approved examples because of choices, not copied phrases.
- The call to action matches the actual next step.
- No phrase from the Avoid list appears.
- The draft could not be sent unchanged to ten unrelated readers.
- The owner named above approved this exact final version before it was sent or published.
## 8. Maintenance
After a meaningful edit, record what changed and why. Review this guide after [cadence] or when the approved writing changes.Set it up once
- 1. Choose a home where your AI can reliably read the guide, such as a project instruction or a shared document.
- 2. Add three to five current approved examples. Link to them or paste the relevant parts.
- 3. Write the rules you can check in a draft. Replace adjectives with observable choices whenever you can.
- 4. Name who can approve changes.
Safe dry run
Use a low-risk internal note or draft that has no new product, customer, financial, medical, legal, or employment claim.
Expected output: the sources and assumptions, then either a draft of about 120 words or one question about the missing fact.
Illustrative example
This is a fictional scenario, included to show the shape of a useful brief. Replace it with your own evidence and approved writing.
Brief
A small product team is writing a customer update after fixing a recurring export error. The reader needs to know what changed, what still needs watching, and how to report a problem.
What a good draft preserves
It leads with the customer impact, states the verified change, avoids claiming the issue can never recur, and gives the reader one honest next step. It does not manufacture an apology or a success story.
Filled-in voice decisions (fictional)
- Prefer: open with what changed for the customer.
- Avoid: "We're excited to announce".
- Rhythm: one idea per sentence.
- Certainty: every number links to its source.
Three places that make it yours
Source order
Decide what wins when a sales deck, a founder note, and an older blog post say different things. This prevents confident guesswork.
Channel choices
A voice is not a single length or format. Set separate rules for email, a post, a page, and a high-stakes update when they truly differ.
Review trigger
Define when the model must pause: uncertain claims, sensitive recipients, a missing source, or a promise that changes the stakes.
When the draft still feels off
| Signal | What to change |
|---|---|
| The draft uses the right adjectives but sounds generic. | Replace adjectives with two approved examples and a specific choice about openings, evidence, or sentence rhythm. |
| The model makes confident claims without enough proof. | Move claim rules above style rules. Require a source label or one question before it drafts. |
| The guide grows every time someone reviews it. | Remove rules that do not change a real draft. Keep a short change log instead of preserving every old preference. |
Review standard
- The guide names a real reader, outcome, and approval owner.
- A draft can point to the evidence behind every material claim.
- The guide tells the model when to ask rather than make a choice for you.
- A fresh sample feels more specific after the guide, not merely more polished.
Make the voice compound
A Second Brain keeps your context, preferences, instructions, and reusable skills in your own AI work, so the next draft starts from them.