Why your brand voice doc isn't working for AI

Most B2B SaaS voice docs read like a personality test. “We’re confident but humble. Professional but conversational. Bold but never arrogant.” The intent is good. The output, when you feed those words into Claude or ChatGPT and ask it to write a post, is generic.

The shape of the doc is the issue. Most voice docs describe how the brand should feel. AI agents can’t translate feelings into writing decisions, so the output reads hollow.

The voice docs I see working for AI-powered content engines look different. They name specific writing decisions the brand has made and applies every time. The agent reads decisions and produces consistent output. The reader feels the personality on the other end without you ever using the word “confident.”

What’s wrong with personality-based voice docs

Try this experiment. Open your current voice doc. Hand it to Claude with this prompt: “Write a 400-word LinkedIn post about pricing AI products, in this voice.” Read what comes back.

If your voice doc is personality-based, the output will feel hollow. It’ll be technically correct. It’ll be the kind of polished generic you’ve seen ten thousand times on LinkedIn. The personality words don’t translate into a specific written choice the model can make. Without those choices, the model defaults to its training data, which is the polished generic mean of the internet.

Joanna Wiebe at Copyhackers has been writing about this gap for years, before AI tools existed. The “voice” that works in real copy is a set of decisions about what to do and what to refuse to do. Wiebe’s frameworks are full of explicit rules: “lead with the customer’s pain, not the product.” That’s a decision an agent can follow. “Be authentic” is not.

What a decision-based voice doc actually contains

Five sections. Each one a list of decisions, not adjectives.

Sentence-level rules. What the writer always does and never does at the sentence level. Always uses contractions. Never starts a sentence with “Honestly.” Always cuts adverbs. Never writes more than two sentences in a row over twenty words. Always uses straight quotes, never curly.

Word-level rules. A banned-words list. The corporate verbs you keep flagging in drafts. The consultant-speak that pattern-matches to AI slop. The specific replacements you make every time. This is the section most voice docs skip entirely. It’s also the highest-impact section for AI output.

Structure rules. How a piece is shaped, regardless of topic. Open with an observation, never a question. Always include at least three external citations. Always close with a specific action the reader can take. Subheads are sentence case, never title case.

Reference voice samples. Three to five actual pieces the founder has shipped that capture the voice at its best. The agent reads these and matches the patterns it sees there, not the patterns it imagines from the adjectives. The richest source of these samples is the founder’s own customer calls, processed through the customer-call-to-content pipeline so the voice in the doc matches the voice on the calls.

Common failure modes. What the voice tends to slip into when the writer is tired or the topic is generic. Examples include the corporate voice (overly formal, no first person), the LinkedIn voice (motivational, vague), and the AI voice (em dashes, contrarian reversals, banned words). Naming the failure modes helps the agent self-correct.

David Perell has written about the difference between “writing in your voice” and “writing about your voice.” Most voice docs do the latter. The good ones do the former by example, with hard rules.

A before-and-after

A typical personality-based instruction in a voice doc:

“Our voice is direct and confident. We don’t waste the reader’s time. We’re peer-to-peer with senior operators.”

An AI agent reads that, nods, and produces “We strategically partner with leading enterprises to maximize operational impact.” Direct? Sure. Confident? Sure. Useful? No.

The decision-based version:

• Use “you” not “our customers”
• Banned-words list (the corporate verbs and AI tells you flag every time)
• Maximum two sentences in a row over twenty words
• Always name a specific person or company when citing a position
• Cite at least three external sources per long-form piece
• Close every piece with a specific action

Same writer, totally different output. The agent has somewhere to go.

This is the same shift I covered in AI agents are your new employees. Treating the agent like a contractor (vague brief, hope for the best) produces contractor results. Treating it like a junior employee with a structured brief produces compounding output.

How to rewrite your voice doc this week

Three concrete moves:

1. Pull your current voice doc. Count the adjectives. If it’s mostly adjectives, it’s a personality doc. Scrap it.
2. Pull five pieces of writing the founder is genuinely proud of. Read them out loud. Note every decision: word choices, sentence rhythms, structural patterns. Write those decisions down as rules.
3. Add a banned-words list. Start with the obvious AI tells you flag in your own drafts. Add anything you cringe at when you see it. This list is the fastest part of the doc to write and one of the most powerful.

Test the rewritten doc the same way you’d test a hire. Hand it to Claude with three different topic prompts. Read what comes back. If the output reads like the founder, you’ve got a working voice doc. If it still reads generic, your rules aren’t specific enough yet. Tune until it sounds right.

Justin Welsh has noted that the operators producing the most consistent content in 2026 are the ones who’ve turned their voice into a system the agent can execute. The voice doc is that system.

If you want help converting your brand voice into a working spec your content engine can actually use, book a call. Most of the content engine work I do starts with this exact rewrite.

Jared Castronova is the founder of JAC Growth Marketing, where he builds AI-powered GTM systems for B2B companies.