Why AI writing sounds off-brand (and what I built to fix it)
I woke up one morning, opened my inbox, and saw that a new email had gone out to our entire user base overnight. Title case on every heading. I'd asked for sentence case over and over. A few scrolls down, emojis scattered through the body copy. We'd agreed on no emojis, multiple times, in writing.
This wasn't a one-off. I spent six years building Uxcel, a UX education platform, and brand consistency started breaking the moment we scaled past two people. Marketing wrote with energy. Product wrote with precision. Support wrote with warmth. Sounds fine on paper. In reality, we sounded like five different companies depending on which email or help article you happened to read.
The standard advice is to write brand guidelines. We did. Everyone does. Here's what nobody tells you: 95% of organizations have them, and only about a quarter actually use them. They end up in a Google Doc that nobody opens after week one.
Then AI showed up and turned a manageable problem into a structural one. Overnight, every person on the team had a writing assistant producing perfectly grammatical, confidently structured, completely generic content. A new writer that sounded like nobody and everybody at the same time. I tried prompt engineering, custom GPTs with uploaded docs, the whole playbook. Nothing stuck. The AI forgot its instructions constantly, there was no version control across team members, and every session felt like starting over.
My co-founder Gene summed it up: "The most pain for me is aligning this all the time, because it always breaks."
The root cause is baked into how these models work. Every LLM predicts the most statistically likely next word, which is the linguistic equivalent of always ordering the most popular dish on the menu. Safe, predictable, never distinctive. Cranking up temperature adds randomness, not your brand's voice. And the reinforcement learning process that makes these models "helpful" actively averages out anything distinctive.
So I built Brivvy. It breaks your brand voice down into two layers: tones that shape personality across five dimensions, and rules that enforce mechanics across 30+ parameters, everything from heading case to contraction frequency to whether you use em dashes. Your voice becomes something a machine can read and follow, not a PDF it ignores. It works inside Claude, ChatGPT, Cursor, Windsurf, and GitHub Copilot through the Model Context Protocol, so your rules are just there every time you write. No more uploading and no more re-prompting.
If you've ever looked at AI output and thought "this doesn't sound like us," I'd love to hear what you've tried. That exact frustration is why Brivvy exists.
For me, the biggest miss with AI writing tools is that they completely miss contextual relationships. I write one way to a coworker on Slack, totally differently to a client on email, and in a completely different register when I'm writing on behalf of an organization to a broad audience vs. as a member of one. But there's STILL a common thread running through all of it, and that balance is really hard to strike with AI. Also, as a freelancer working across multiple brands, sometimes voices start bleeding into each other. Several times I'll catch AI tools writing for one client like another client. A tool that actually solves for this would 100% simplify things for a lot of us dealing with the same problem
this really resonates, especially the part where brand guidelines exist but no one actually uses them
AI definitely made this worse by scaling inconsistency instead of fixing it
the tones + rules layer is an interesting way to make voice enforceable instead of just documented
you could also try testing this in a structured way across different teams to see how consistency improves
prize pool just opened at $0, so timing is interesting
The brand guidelines problem is older than AI — but you're right that AI made it structural. Before, inconsistency came from different people making different judgment calls. Now it comes from one tool that defaults to "statistically average," which is actually harder to fix because it feels consistent even when it's off-brand.
The insight about RLHF actively averaging out distinctiveness is the most underrated part of this. It's not that the model ignores your style guide — it's that it was literally trained to sand down anything unusual. Prompt engineering fights against the model's core objective.
Curious how Brivvy handles the gap between content types — a support article and a marketing email should feel like the same brand but not sound identical. Is that a single voice config or multiple?
Multiple voices. You create separate voices per context, support, marketing, leadership, whatever you need. Each one has its own tone settings, rules, and formatting preferences. They all live in the same workspace so they feel like the same brand, but they behave differently depending on what you're writing. The RLHF point is spot on too, you're not just fighting bad prompts, you're fighting the model's training objective. That's why we built it as infrastructure rather than a prompt layer.
The brand voice problem is underrated. I ran into something similar building an AI productivity tool — getting the model to classify intent correctly without sounding generic took most of my debugging time. How did you approach the training data?
We didn't approach it from the training data side. Instead of trying to fine-tune or retrain anything, we built a constraint layer that sits on top of whatever model you're using. You define the tone, rules, and formatting, and those get delivered through MCP at generation time. So it's model-agnostic, no training data needed. Might be a different angle than what you're solving for, but for brand voice specifically it's been a much faster path than trying to get the model itself to behave.
AI sounds "off-brand" because most people use it to generate noise, not to build Architecture.
The real disconnect isn't just in the words—it's in the lack of Visual Trust and Sovereignty. I’ve seen countless founders polish their AI copy while building on rented platforms that can vanish tomorrow (exactly what happened to my Medium infrastructure recently).
A brand isn't just a tone of voice; it’s an ecosystem. I moved my focus from "writing better prompts" to building a Bunker where my visual authority (85mm) and my data delivery are 100% mine.
You can fix the voice, but if you don’t own the stage, you’re just a well-spoken tenant.
Interesting take -especially the “well-spoken tenant” part. Do you think owning the “stage” actually changes how users perceive the message, or is it more about long-term control? Feels like most founders focus on the message itself, not where it lives.
The "well-spoken tenant" line is great...ownership matters. We're focused on a slightly different piece of the puzzle, making sure that wherever you publish, the voice is actually yours. Platform sovereignty is important, but so is making sure the words themselves aren't just the statistical average. Both problems need solving.
The email going out with title case after you'd asked for sentence case multiple times - that's such a specific pain, i felt that.
The google doc graveyard for brand guidelines is so real. The problem was never the guidelines, it was that nobody had a reason to open them mid-workflow.
the MCP angle is interesting - making the rules live inside the tools people already use rather than being a separate step is the right instinct. how are you handling it when someone's tone genuinely needs to flex, like support vs marketing copy?
This is a very accurate summary of the problem. We have faced the exact same challenge while building KortexMail. It is difficult to keep the AI from falling back into that generic, robotic tone. Moving brand guidelines out of a static document and directly into the daily workflow is a smart solution.
That's exactly the shift. Static brand guidelines in a PDF or Notion doc sound good in theory, but nobody's copying rules into every AI session. Moving them into the workflow so they're always present is what actually makes them stick. Cool to see you're dealing with the same thing on the email side with KortexMail.
This really resonates, especially the part about brand guidelines existing but not actually being used.
What you’re describing feels less like a writing problem and more like a systems problem. AI just exposed it faster. Before, inconsistencies came from different people. Now they come from a single tool that defaults to “average”.
The tone vs rules split is interesting. Most attempts I’ve seen focus too much on tone (“sound friendly”) and ignore enforceable mechanics like casing, contractions, formatting. That’s usually where things break.
Curious how strict your system is in practice. Do teams lean more toward hard enforcement (rejecting output that doesn’t comply) or soft guidance? Feels like adoption might depend on how intrusive it is in real workflows.
100% a systems problem. Great writing at scale has to be engineered. I minored in literature at uni, so the rules and mechanics side was always interesting to me. When AI took off and I learned exactly why LLMs write the way they do, it clicked why those mechanical constraints matter so much.
On enforcement, what we're seeing is that most teams lean toward soft guidance. The rules are always applied, but nobody's rejecting output outright. It's more like guardrails than gates. Teams use the AI preview to tweak their config until the output feels right, then trust the system to hold that standard going forward. The friction is low enough that people actually use it, which is the whole point. A strict enforcement layer that people bypass is actually worse than soft guidance that's always on.
"The most statistically likely next word" is exactly why brand voice gets averaged out — the model is literally optimized to sound like everyone. The MCP integration is the right architectural decision. Uploading a style guide every session is the reason nothing sticks. Having rules present by default changes the workflow entirely. Curious how it handles voice consistency across different content types — the brand voice for a help article and a marketing email should feel related but not identical.
Hey Danilo. Yeah, you're spot on. You don't have one brand voice, you have several. A support article and a marketing email should feel like the same brand but not sound identical. That's why we built Brivvy around multiple voices per workspace, each with its own tone and rules, all delivered through MCP so the right one loads based on what you're writing. Works at the workspace level or per user.
We had brand guidelines too, but AI made everything sound correct instead of us. Making voice enforceable (not just documented) is the real unlock — curious to try this.
Throughout the three startups that I've been involved in as well as working in a large enterprise, you can't imagine how much time we spent reviewing content just to ensure it was on-brand. I'd love to hear what you think of Brivvy and if it helps solve your problem.
Brand consistency in AI-generated content is one of those problems that seems small until you're managing multiple products. I'm building a portfolio of AI SaaS tools right now and keeping a consistent voice across 5 different product sites has been one of the unexpected challenges. The approach of building brand voice infrastructure rather than just prompt engineering hacks makes a lot of sense. Prompts drift over time, models get updated, and suddenly your content sounds different. How are you handling the feedback loop when a user says "this doesn't sound like us" — does that train the model or update the brand rules?
Hey Jarvis, yeah you're hitting the exact scaling problem we're built for. With multiple products, different contexts, that's where prompt engineering falls apart. Our rules live in the MCP so they don't drift when models update.
When a user says, "this doesn't sound like us", they can adjust tone and rules directly, it's a quick but manual process. We're building out examples and directives next, so you can bake in real reference snippets and specific instructions like "always lead with the benefit" or "don't open cold." That'll close the feedback loop a lot faster. Of course, we only just moved out of closed beta, some I'm now doing further discovery with larger user groups to make enhancements.
Really interesting approach to the AI writing problem. I've been building something similar in the AI trust/verification space — the challenge of knowing when AI output is actually reliable is huge. Good to see more people tackling this from different angles.
Trust in AI output has been a huge problem for us at Uxcel and caused us so many headaches. We built AI previews directly into Brivvy so you can see how the output changes in real time as you adjust tone, rules, formatting, all of it. Makes it way easier to trust the config before it ever hits production.
Ran into this building with agents. Told the system no corporate tone, three days later it was generating meeting recaps that read like press releases. The drift is real and it's slow enough you don't notice until it's already everywhere.
Yeah, keeping agents on-brand is like herding cats. The drift is slow enough that by the time you notice, it's already everywhere. That's exactly why we built Brivvy as workspace-level infrastructure, not just a prompt you paste in. The rules persist across every session so agents can't just wander off.
the slow drift is what gets you - by the time you notice, it's been weeks. I reload context at startup for each agent but it's still manual. does Brivvy handle agents that spawn sub-agents? that's where drift usually starts
Right now Brivvy works at the MCP client level, so any agent that connects gets the voice rules automatically. Sub-agent enforcement is something we're actively exploring. In the meantime, you can also call Brivvy's MCP after content has been generated to verify everything is on par with your guidelines, kind of like a post-generation check. So even if a sub-agent drifts, you have a way to catch it before anything goes out.
The "statistical average" framing really nails it. We ran into a version of this problem from the content/SEO side — we audited our own blog posts and found that the ones where AI had helped with drafting all had the same tell: "studies show" without naming a study, confident-sounding stats with no source, and that particular brand of fluent-but-empty phrasing that LLMs default to. Eight instances of "studies show" across nine posts, zero actual studies named.
What's interesting is this isn't just a brand problem anymore — it's becoming a discoverability problem. AI search engines (ChatGPT, Perplexity) are now citing sources in their answers, and the research is showing they heavily prefer content with specific data points and named sources over the generic "here are five tips" format that most AI writing produces. So the statistical average that LLMs default to is also the format that AI search engines are least likely to cite. Distinctiveness is becoming a ranking factor, not just a brand asset.
The MCP integration angle is smart for the same reason Josh mentioned — making rules ambient instead of re-prompted. Curious whether you've seen teams use the mechanical rules (heading case, contraction frequency) as a proxy for detecting when AI-generated content slipped through review without the voice layer applied.
The "studies show" tell is a perfect example. We built a templates feature in Brivvy for exactly this, you can create markdown-based structural playbooks for any repeatable content type. So you can define that when you cite research, you actually name the source, and specify whether references go inline as links or collected at the end. It's a small thing that completely changes whether the output reads as credible or just fluent.
The discoverability angle is really interesting too. If AI search engines are prioritizing specific, sourced content over generic listicles, then the statistical average isn't just a brand problem, it's a visibility problem. Distinctiveness becomes a ranking signal. That tracks with everything we're seeing.
On the mechanical rules question, yes. Things like heading case, contraction frequency, list formatting, those are all configurable in Brivvy at the voice level. And you're right that they double as a detection layer. If content slips through without the voice applied, those inconsistencies are the first thing you'd catch in review. We plan to introduce a Brand Guardian feature in the future that helps enforce guidelines for both existing and new content alike.
"Every LLM predicts the most statistically likely next word, which is the linguistic equivalent of always ordering the most popular dish on the menu." That's the line. Brand voice is distinctive by definition, it's the thing the average of all language would never produce on its own. Most people fight that with prompts. You built a constraint system.
That's exactly it. The reality is most content is going to be AI-generated from here on out. The question isn't whether brands will use AI to write, it's whether that writing will sound like them or like everyone else. We built Brivvy so you can lay down the constraints, the tone, the rules, the guardrails, and then let AI do its thing within that system. You get your voice back instead of settling for the average.
The menu analogy holds. Most brands using AI are ordering the specials, safe, predictable, and what everyone else is having. You're building the kitchen. The constraints aren't limitations. They're the recipe. Without them, the output is just a consensus. With them, it's yours. That's not AI writing. That's AI cooking from your cookbook.
This resonates a lot.
I’ve been working on an AI that tries to learn how a specific person thinks instead of generating generic responses — and this “off-brand” feeling is exactly what I’m trying to solve.
Interesting project, would love to check it out.
And yeah, the "off-brand" problem was real long before AI. Anytime you had a team of writers, things drifted. AI just made it exponentially worse because now every tool is generating content, not just people.
This comment was deleted 2 months ago
Really appreciate the thoughtful questions here, Josh. The re-prompting pain is exactly what MCP solves, your voice rules are just there, every session, no pasting required.
On conflicts, rules win. If you set "no contractions" as a rule but have casual tone cranked up, the system follows the rule and instructs the AI to find other ways to hit that casual feel, things like shorter sentences, simpler vocab, direct phrasing. So you get the tone without breaking the constraint.
The mechanical parameters came directly from our experience at Uxcel. We have a ton of educational content with multiple writers, so building a comprehensive style system was fundamental for us. You're right that most brands probably care deeply about five rules, but here's the thing, with AI writing to the average, the other 25 are what create variety and distinctiveness. Without them, everyone's output converges to the same default. Those "didn't know I needed it" rules are what keep your content from blending in with everyone else's.
This comment was deleted 2 months ago
I appreciate you Josh. If you ever try Brivvy, please reach out and let me know any feedback you might have. I've sent Bordair to our dev team to explore further. Good luck and look forward to crossing paths again.
This comment was deleted 2 months ago