Most teams do not have a brand voice problem. They have a brand voice file problem. The brief says "friendly professional," the model writes the same LinkedIn-flavored paragraph everyone else gets, and the marketing lead spends Friday afternoon rewriting twenty AI drafts back into English. The fix is not a better adjective. It is a reusable constraint system: voice tokens, a do/don't list, an annotated example bank, a scoring rubric, and a refresh cadence that stops the file from rotting.
The short answer
A brand voice guide for AI is a constraint system, not a personality essay. You token your voice on numeric scales, lock a forbidden-phrase list, bank annotated examples, score every draft on a rubric, and refresh the file quarterly with one named owner. Paste the tokens and two gold examples into the style field of any tool (or any model), keep the full bank in your wiki for scoring, and use the rubric to decide what ships. A score of 80 or higher is a reasonable threshold for ordinary content; P0 pages such as home, pricing, or legal need a full human read.
The ILAA AI Rewriter is the natural place to start: paste your ledger, run a rough draft through it, and score the result. The Ad Copy Generator and SEO Meta Generator are also consumers of the same file. The same guide works across the GPT-5.6 family, Claude Fable 5, and the Gemini 3.5 agentic stack, because you are feeding the model rules, not training it.
Why "friendly professional" breaks (and what AI actually needs)
Large language models are trained on a lot of "friendly professional." To them, it is a prior: clear but warm, confident but not brash, no jokes, no jargon, end with a forward-looking sentence. That prior is also the default voice of every SaaS homepage, every LinkedIn thought leader, and every AI-generated welcome email. It is not wrong. It is empty.
The problem is that adjectives collapse into averages. "Friendly" pushes the model toward exclamation marks and first names. "Professional" pulls it back toward formal hedges and passive constructions. "Helpful" adds a bullet list. The result is a paragraph that no human would write and no reader would remember.
What AI needs instead are behavioral constraints. Tell the model what to do in language it can execute. For example:
- Open with the answer, not a warm-up. (Not "we are excited to announce.")
- Use "you" at least once in the first two sentences.
- Name one concrete object, number, or constraint per paragraph.
- Never say "seamless," "unlock," "empower," or "game-changer."
- Write sentences between 8 and 22 words, with one long sentence allowed per section.
Those rules are not poetic. They are executable. That is the difference between a brand brief and a brand voice guide that AI can follow.
If you are coming from the rewriter workflow post, think of this guide as the lock you install before you run the FAST system. Speed without voice discipline just ships mush faster. The broader tool-to-task map is in our guide to the best free AI writing tools in 2026.
Voice vs tone: the one distinction models collapse
Mailchimp's famous content style guide made the distinction cleanly: you have the same voice all the time, but your tone changes depending on the situation and the reader's emotional state. Voice is the constant. Tone is the situational overlay.
Most AI tools treat voice and tone as the same slider. A "Professional" preset tries to do both. A "Casual" preset tries to do both. The result is that a support apology and a launch announcement sound like siblings, which is wrong in both directions. A good brand voice guide encodes them separately:
- Voice tokens never change: formality level, sentence band, forbidden phrases, CTA verbs, punctuation rules.
- Tone overlays change by channel and moment: support is calmer and direct; launch copy is a little brighter; legal pages are cooler and exact.
Nielsen Norman Group breaks tone into four spectra: humor (serious to funny), formality (formal to casual), respectfulness (respectful to irreverent), and enthusiasm (matter-of-fact to enthusiastic). Those four dimensions are a useful foundation for tokens, because they are concrete enough to score. The article credits NN/g for the framework; the AI-specific additions below are ours.
The TOKEN, BANK, SCORE, REFRESH system
This is the operating system. Each step prevents a specific failure:
| Step | Artifact | Prevents |
|---|---|---|
| TOKEN | Numeric voice tokens, forbidden phrases, register map | Adjective mush and model default drift |
| BANK | Gold / near-miss / fail examples with annotations | "Sounds like us" with no proof |
| SCORE | 0 to 100 rubric, ship threshold | Vibes-based quality assurance |
| REFRESH | Cadence, owner, triggers, versioned file | Stale PDF that nobody updates |
The system is deliberate and a little boring. That is the point. Brand consistency at scale is not a creative breakthrough every Monday. It is the same constraints applied thousands of times. The AI does the variation; the guide does the guardrails.
Step 1: TOKEN your voice
Start with five to eight tokens. Any more and the model starts ignoring them; any fewer and you are back to adjectives. Every token needs four parts:
- Name. The dimension, such as formality, directness, or claim_temperature.
- Scale. A 0 to 10 score. Zero is not bad; it is just the opposite pole.
- Behavioral sentence. The rule the model can execute, not a mood description.
- Anti-example. A sentence that violates the token, so the model knows where the line is.
Here is a starter token table. The first four are adapted from NN/g's tone dimensions; the rest are AI-specific additions we have found useful for 2026.
| Token | Score | Behavioral rule | Anti-example |
|---|---|---|---|
| formality | 3/10 | Write like a competent colleague at a standup, not a keynote. | "We are delighted to present our comprehensive solution." |
| humor | 2/10 | Dry observations only. No jokes, memes, or exclamation-heavy cheer. | "Invoicing doesn't have to be a snooze-fest!" |
| respectfulness | 7/10 | Take the reader's work seriously. No snark about competitors or users. | "Unlike legacy tools that still live in 2010..." |
| enthusiasm | 4/10 | Confident but measured. Let the feature do the bragging. | "Supercharge your workflow with cutting-edge AI!" |
| directness | 8/10 | Open with the answer, not a warm-up sentence. | "In today's fast-paced business landscape, efficiency matters." |
| warmth | 7/10 | Use "you" early and often. Acknowledge the reader's situation. | "Acme enables businesses to streamline processes." |
| claim_temperature | 3/10 | Hedge claims that could be regulated or audited. "Helps" over "guarantees." | "Guarantee 10x ROI in 30 days or your money back." |
| specificity | 8/10 | Every paragraph names one concrete object, number, or constraint. | "Our powerful platform delivers robust results for teams of all sizes." |
Notice the anti-examples. A token without an anti-example is just an adjective. The anti-example is what makes the token teachable.
After the core tokens, add formatting and lexical locks:
- Sentence band: typical 8 to 22 words; one long sentence allowed per section.
- Forbidden phrases: a denylist of words that flag instant rewrite. (See the template below for a starter list.)
- CTA verbs: the exact verbs you use for action links. Start, Open, Download, Send, Compare, not Unlock, Supercharge, or Skyrocket.
- Emoji policy: none, sparse, or channel-only. Define it.
- Punctuation policy: no em dash, Oxford comma yes/no, max one exclamation per email.
- Register map: tone overlays by channel (site, email, support, social, legal, incident).
The register map is where voice meets tone. It does not change the tokens; it dials them. For example, support keeps directness high but drops enthusiasm to 2; a product launch keeps directness high but allows enthusiasm to 6 for one paragraph; legal pages move formality to 7 and claim_temperature to 1.
Step 2: BANK annotated examples
Models copy labeled examples better than they follow abstract rules. Your example bank should be short and annotated. We recommend a minimum viable bank of seven pieces:
- 3 gold: ready to ship as-is, with annotations explaining why.
- 2 near-miss: almost right, with a clear fix note.
- 2 fail: typical AI slop for your brand, with the violation highlighted.
Every example needs a channel, a job, the text, and token evidence. Here is the format:
| Channel | Job | Text | Why it scores | Token evidence |
|---|---|---|---|---|
| Homepage | Explain the product in 50 words | "Late invoices kill studio cash flow. Northline sends reminders, tracks partial payments, and exports clean reports so you get paid without chasing clients." | Names the pain, names three concrete features, no hype verbs. | directness 8, specificity 8, claim_temperature 3 |
| Support reply | Tell a user their CSV import failed | "Your CSV import failed because the Amount column includes text. Remove the dollar signs, save as CSV, and re-upload. We will retry automatically." | Direct cause, exact fix, no apology theater. | directness 9, warmth 6, formality 3 |
| Email launch | Announce a new reports feature | "You can now export a cash-flow report for any date range. Open Reports, pick your range, and download CSV or PDF. No new plan required." | Leads with the new thing, gives exact steps, no surprise exclamation. | directness 8, enthusiasm 5, specificity 8 |
The annotations are the teaching signal. Without them, the model sees three good paragraphs and tries to clone them. With them, the model sees the underlying pattern.
Near-miss and fail examples are even more valuable. A near-miss might be a perfectly grammatical paragraph that uses "empower" once. The fix note says: "Replace 'empower' with a concrete verb or remove it." A fail might be a paragraph that opens "In today's fast-paced financial landscape, businesses need seamless solutions..." The annotation says: "Every token violated. Rewrite from the reader's problem."
Guard against one common mistake: do not let the bank become a thesaurus of your past copy. If every gold example is from 2023, the model will write like your 2023 self. Replace one gold and one fail every refresh.
Step 3: SCORE every draft
Brand fit is subjective until you make it a rubric. A rubric also protects the writer. It turns "I do not like it" into "It scored 64 because it used three forbidden phrases and buried the lede."
Here is a 100-point rubric you can start with. Adjust weights to match your risk profile; regulated brands should raise claim_temperature weight, and consumer brands might raise warmth.
| Criterion | Points | Pass signal | Fail signal |
|---|---|---|---|
| Token match | 25 | All scored dimensions land in the band; register map respected. | Two or more tokens out of band. |
| Lexical locks | 20 | No forbidden phrases or brand landmines. | Any forbidden phrase appears. |
| Example-bank rhythm | 15 | Sentence length and cadence match gold examples. | Reads like generic AI cadence. |
| Claim temperature / accuracy | 15 | Claims are provable or properly hedged. | Absolute promise, invented stat, or regulatory risk. |
| Channel fit | 10 | Tone overlay matches the channel. | Launch voice in a refund email. |
| Specificity | 10 | Concrete object, number, or constraint per paragraph. | Adjectives without anchors. |
| Human residue | 5 | Reads like a letter from a knowledgeable person. | Summary-bot cadence. |
Ship threshold is 80. That is not a law; it is a useful line. If a piece scores below 80, either revise it against the tokens or flag a human. If it scores below 60, rewrite from scratch.
Tier your human review by risk. P0 pages (home, pricing, product, legal, incident communications) get a full human read. P1 pages (top blog posts, key landing pages) get a spot check of the score and a skim of the copy. P2 pages (long-tail social variants, internal notes) can ship on rubric score alone with random sampling. The point is not to read everything; the point is to read what matters.
If you want a second opinion, run the draft through the ILAA Grammar Checker after the voice gate. Grammar polish should happen after brand fit, not instead of it.
Step 4: REFRESH before it rots
A guide that never changes becomes a liability. Your product changes, your audience changes, the models change, and what sounded fresh in January sounds like everyone else by August. The refresh step is what separates a living system from a PDF in a folder.
| Element | Rule | Why it matters |
|---|---|---|
| Owner | One named human: brand lead, content lead, or editorial manager. | Committees do not refresh files. One owner does. |
| Cadence | Quarterly calendar review, 30 minutes. | Forces a regular look even when nothing is on fire. |
| Triggers | New product line, new ICP, model-family change, 3+ rubric failures in one week, legal/compliance change, acquisition, rebrand. | Catches drift before it scales. |
| Ritual | Voice autopsy: score 5 recent AI outputs, replace 1 gold and 1 fail example, update 1 token if needed. | Keeps the bank grounded in current reality. |
| Versioning | File name includes date: voice-guide-v2026-08.md. | Prevents people from pasting the wrong version. |
The voice autopsy is the key ritual. Pull five recent pieces of AI output that shipped. Score them against the rubric. Discuss one thing the model got right, one token it ignored, and one phrase that crept in. Then replace one gold example and one fail example. The whole exercise takes half an hour and pays for itself in the first avoided rewrite.
If you are using the guide across multiple models (say GPT-5.6 Sol for long content, Claude Fable 5 for careful edits, and Gemini 3.5 Flash for quick variants), the same tokens apply. What changes is how often each model needs the reminder. Sol may drift less because of its long context; Flash may drift more because it is tuned for speed. The refresh step catches both.
Worked example: from "friendly professional" to a working guide
Let us make this concrete with a fictional brand. Northline is a B2B invoicing and cash-flow tool for independent design studios and small agencies. Their old brief says: "friendly professional, helpful, modern, not too salesy." That is exactly the kind of brief that makes models produce slop.
Here is how you convert it into a tokened guide.
The vague brief
"Our voice is friendly professional. We want to sound helpful and modern, but not too salesy. We are a B2B invoicing tool for design studios. Avoid jargon."
The working guide (Northline)
Purpose: We sound like a studio operations lead who has been burned by late invoices, not a SaaS mascot. We respect the reader's time and their anxiety about money.
Core tokens:
- formality 3: write like a competent colleague at a standup.
- directness 8: open with the answer, not a warm-up.
- warmth 7: use "you" in the first two sentences.
- humor 2: dry observations only; no forced jokes.
- enthusiasm 4: confident but measured.
- claim_temperature 3: hedge regulated claims; say "helps" not "guarantees."
- specificity 8: every paragraph names a concrete object, number, or constraint.
Forbidden phrases: seamless, unlock, empower, game-changer, cutting-edge, next-gen, revolutionize, world-class, synergy, elevate your, in today's, leverage, utilize, robust.
CTA verbs: Send the invoice, Turn on reminders, Export the report, Compare plans.
Register map:
- Homepage: directness 8, enthusiasm 5.
- Support replies: directness 9, enthusiasm 2, warmth 6.
- Launch emails: directness 7, enthusiasm 6.
- Legal pages: formality 7, claim_temperature 1, warmth 4.
The job: write a 50-word homepage lede
We asked an LLM to write the same job twice: once with no guide, once with the Northline guide pasted into the instructions.
Output without the guide
"In today's fast-paced business landscape, managing invoices can be a challenge. Northline empowers design studios to streamline their billing workflow with a seamless, user-friendly platform. Unlock faster payments, reduce manual work, and take control of your cash flow with our cutting-edge invoicing solution built for modern agencies."
Rubric score: 31/100. It uses six forbidden phrases ("In today's," "empower," "seamless," "unlock," "cutting-edge," "take control of"). The first sentence is a warm-up. There is no concrete object. Claim_temperature is violated by "faster payments" with no qualifier.
Output with the guide
"Late invoices kill studio cash flow. Northline sends automatic reminders, tracks partial payments, and exports clean reports so you spend less time chasing clients and more time billing real work."
Rubric score: 88/100. It opens with the answer (directness 8), names three concrete features (specificity 8), uses "you" early (warmth 7), avoids the forbidden list (lexical locks 20/20), and hedges with "less time" rather than guaranteeing speed (claim_temperature 3). It loses a few points on example-bank rhythm because it is a new sentence pattern, but the cadence is close enough to ship.
The difference is not prompt engineering magic. It is that the second output had constraints. The first output had an adjective.
The copy-ready Brand Voice Guide template
Copy this template into a Markdown file or a Notion page. Fill in the brackets. Keep it under 1,500 words total; longer guides get ignored. The examples section will grow over time, but the tokens and forbidden list should stay tight.
---
# BRAND VOICE GUIDE
Company: [Your company name]
Owner: [One named human]
Version: voice-guide-vYYYY-MM
Updated: [date]
---
## 1. Purpose
We sound like [one-sentence persona].
Our voice builds trust by [specific outcome, not an adjective].
## 2. Core voice tokens
Each token: name | 0-10 score | behavioral rule | anti-example
- formality [0-10]: [rule] | Anti: [sentence]
- humor [0-10]: [rule] | Anti: [sentence]
- respectfulness [0-10]: [rule] | Anti: [sentence]
- enthusiasm [0-10]: [rule] | Anti: [sentence]
- directness [0-10]: [rule] | Anti: [sentence]
- warmth [0-10]: [rule] | Anti: [sentence]
- claim_temperature [0-10]: [rule] | Anti: [sentence]
- specificity [0-10]: [rule] | Anti: [sentence]
## 3. Formatting and lexical locks
- Sentence band: [e.g. 8-22 words typical; one long sentence per section allowed]
- Forbidden phrases: [comma-separated list]
- CTA verbs: [Start, Open, Download, Compare, etc.]
- Emoji policy: [none / sparse / channel-only]
- Punctuation policy: [no em dash; Oxford comma yes/no; max exclamations per email]
- First sentence rule: [e.g. answer first, no warm-up]
- "You" rule: [e.g. at least once in first two sentences]
## 4. Register map (tone overlays)
- Homepage: [which tokens move and by how much]
- Blog posts: [overlay]
- Support replies: [overlay]
- Launch emails: [overlay]
- Social posts: [overlay]
- Legal / compliance: [overlay]
## 5. Example bank
### Gold examples (ready to ship)
| Channel | Job | Text | Why it scores | Token evidence |
|---|---|---|---|---|
| [channel] | [job] | [text] | [why] | [tokens] |
| [channel] | [job] | [text] | [why] | [tokens] |
| [channel] | [job] | [text] | [why] | [tokens] |
### Near-miss examples (fix notes)
| Channel | Job | Text | Fix | Why it almost works |
|---|---|---|---|---|
| [channel] | [job] | [text] | [fix] | [notes] |
| [channel] | [job] | [text] | [fix] | [notes] |
### Fail examples (what slop looks like for us)
| Channel | Job | Text | Violation | Score loss |
|---|---|---|---|---|
| [channel] | [job] | [text] | [violation] | [points] |
| [channel] | [job] | [text] | [violation] | [points] |
## 6. Scoring rubric
| Criterion | Points | Pass | Fail |
|---|---|---|---|
| Token match | 25 | Tokens in band | 2+ tokens out of band |
| Lexical locks | 20 | No forbidden phrases | Any forbidden phrase |
| Example-bank rhythm | 15 | Matches cadence | Generic cadence |
| Claim temperature | 15 | Provable or hedged | Absolute or risky |
| Channel fit | 10 | Right register | Wrong register |
| Specificity | 10 | Concrete object/number | Empty adjectives |
| Human residue | 5 | Sounds like a letter | Summary-bot cadence |
Ship threshold: 80/100.
P0 pages (home, pricing, legal, incident): 100% human review.
P1 pages: spot-check score + skim.
P2 pages: ship on score, random sample.
## 7. Refresh rules
- Owner: [name]
- Quarterly review: [scheduled date]
- Event triggers: [new product, new ICP, model change, 3+ rubric failures in a week, legal change]
- Voice autopsy: 5 recent outputs, replace 1 gold + 1 fail, update 1 token if needed
This template is intentionally plain. A beautiful template that nobody fills in is useless. A plain one that gets used is the whole job.
How tools consume the guide
A guide is only valuable if the team actually uses it. The right size for a daily prompt is smaller than the full file. We recommend two versions of the guide:
- Prompt version: the tokens, forbidden phrases, and two gold examples. This is what you paste into a tool's style field or a model's system prompt.
- Reference version: the full bank and rubric, kept in a shared wiki or doc for scoring.
The ILAA AI Rewriter is the natural first consumer. It already supports style matching. Paste the prompt version into the instructions, drop in a rough draft, and run it. Then score the result on the rubric. The rewriter is free, requires no signup, and works well for blog posts, emails, and support replies.
The ILAA Ad Copy Generator is the next consumer. Ad copy needs tighter length constraints, but the same tokens and forbidden list apply. Give it the brand token and the CTA verb list, and you get variants that stay on voice without sounding like every other ad in the auction. This is a tool consumer, not the paid-ads workflow itself; the Aug 19 ad-copy sibling post will cover Google and Meta asset mechanics in depth.
The ILAA SEO Meta Generator is also a consumer. Feed it the brand token and forbidden phrases so titles and descriptions do not slip into "cutting-edge SEO platform" territory. The full meta-scaling system is in the Aug 18 meta post; treat this guide as the voice layer underneath it.
Across the major 2026 model families (GPT-5.6 Sol/Terra/Luna, Claude Fable 5, and the Gemini 3.5 agentic stack), the same ledger works. You do not need three different voice religions. You need one file, pasted consistently.
Platform analog: Meta Brand Memory
In June 2026, Meta announced Brand Memory inside a new AI Creative Solution at Cannes Lions. The idea is that the system learns a brand's identity and tone from existing ads and posts, then applies that learning to generative ad variants inside Ads Manager. It is a platform-level attempt to solve the same consistency problem this post is about.
Brand Memory is useful, but it is not a replacement for a company-owned voice guide. Three reasons:
- Scope. It lives inside Meta ads. It does not cover your website, your emails, your documentation, your support replies, or your product copy.
- Training data. It learns from ads you have already run. If your old ads were already generic, Brand Memory will scale that generic work efficiently.
- No rubric or refresh cadence. A platform feature does not replace a documented scoring system and a named owner who updates the file.
Think of Meta Brand Memory as a consumer of your brand rules, the same way the ILAA Ad Copy Generator is. The source of truth is still your ledger. The platform applies it.
Common failure modes
After building and refreshing a few of these guides, the same mistakes show up repeatedly. Here is how to spot and avoid them.
| Failure | Symptom | Fix |
|---|---|---|
| The 40-page PDF | Nobody reads it; operators paste the old prompt. | Cap the prompt version at 1,500 words. Keep the full bank separate. |
| Adjective-only guide | Every output still sounds like LinkedIn. | Add anti-examples and numeric scales. |
| No forbidden list | "Seamless" and "unlock" keep leaking in. | Keep a denylist and treat any hit as a 0 on lexical locks. |
| No fail examples | Team cannot agree what bad looks like. | Add 2 fail examples with explicit violations. |
| No owner | File gets updated when someone is annoyed, not on a schedule. | Name one human and a quarterly calendar slot. |
| No version date | People paste last year's file. | Use voice-guide-vYYYY-MM naming. |
| Gold-line cloning | Model copies exact phrases from gold examples. | Rubric should penalize clones; refresh examples. |
| Voice theater | Team argues about "on brand" instead of scoring. | Use the rubric. Disagreements become arithmetic. |
Frequently asked questions
What is an AI brand voice guide?
An AI brand voice guide is a short, machine-readable constraint file that tells an AI model how your brand sounds. It includes voice tokens on numeric scales, a list of forbidden phrases, annotated gold and fail examples, channel-specific tone overlays, a scoring rubric, and a refresh cadence. The goal is to make AI drafts consistently on-brand without rewriting the same personality paragraph every time.
How is a voice guide different from a tone slider or a ChatGPT custom instruction?
A tone slider changes mood in one dimension. A custom instruction is usually a short paragraph the model can forget or dilute. A voice guide separates constant voice from situational tone, gives numeric scales, locks forbidden phrases, includes an example bank with annotations, and pairs with a rubric that decides what ships. It is an operating system, not a setting.
What are voice tokens?
Voice tokens are named, scored, behavioral rules that describe how your brand writes. A good token has a name, a numeric scale, a behavioral sentence that explains what the score means, and an anti-example. For instance, directness 8 with the rule "open with the answer, not a warm-up" is a token; "professional but friendly" is not.
How many examples does an example bank need?
Start with three gold examples that are ready to ship as-is, two near-miss examples with clear fix notes, and two fail examples that show typical AI slop for your brand. Annotate every example with the channel, the job, and the token evidence. More examples help, but annotations matter far more than volume.
How do you score AI copy for brand voice?
Use a 100-point rubric. A practical starting split is: token match 25 points, lexical locks and forbidden phrases 20 points, example-bank rhythm 15 points, claim temperature and accuracy 15 points, channel fit 10 points, specificity 10 points, and human residue 5 points. A score of 80 or higher is a reasonable ship threshold for ordinary content; P0 pages such as home, pricing, or legal require a full human read.
How often should you refresh a brand voice guide?
Review the guide quarterly as a minimum, and immediately after a major trigger such as a new product line, a new target audience, a model-family change, a compliance change, or three rubric failures in one week. Name one owner and version the file with a date. A stale guide is worse than no guide because it ships stale-sounding copy with a false sense of quality.
Can Meta Brand Memory or a custom GPT replace a written guide?
No. Meta Brand Memory, announced in June 2026, learns your visual and tonal identity from existing ads and applies it to new ad variants inside Meta Ads Manager. A custom GPT can carry a system prompt. Both are useful consumers of your brand rules, but neither covers owned channels such as your website, emails, documentation, or support replies. Neither gives you a rubric or a refresh cadence. Your company-owned voice ledger is still the source of truth.
Which ILAA tools can consume a brand voice guide?
The AI Rewriter at /tools/ai-rewriter is the natural home for the guide: paste your tokens, forbidden phrases, and two gold examples into the style field and run rough drafts through it. The Ad Copy Generator and the SEO Meta Generator are also consumers: feed them the brand token and forbidden list so headlines and meta tags stay on voice. The Grammar Checker can do a final polish pass after the voice gate.
Build the file this week
You do not need a brand summit. You need one hour, one owner, and a plain document. Pick your five most important tokens, write one behavioral sentence and one anti-example for each, list ten forbidden phrases, collect three good examples with annotations, and run one recent AI draft through the rubric. That is the minimum viable voice guide. Everything else is iteration.
The goal is not to make AI sound creative. The goal is to make AI sound like you on your best day, repeated at scale. Constraints do not kill creativity; they kill the generic middle. Once the guide is in place, tools like the AI Rewriter stop being slop machines and start being editors that know your house style. A voice guide also feeds the direct-answer craft covered in our AEO writing guide.
Ready? Open the ILAA AI Rewriter. Paste the template above, add one rough draft, and score the result. Free, no signup, no card. The first rubric score will tell you exactly where to start.
Tool and model names checked August 2026. Re-verify before publishing.