A good message can fail if the tone is off. A mediocre message can succeed if the tone matches the audience.
After 30 years in marketing, that’s the one truth that holds across every industry, every channel, every format. So when I built Paxelo, I knew features weren’t enough. The machine needed a voice. And over two days, I taught it how to find one.
Why Tone Is the Most Overlooked Variable in AI Content
Most people building AI content tools focus on output quality: is the writing grammatically correct, is it factually accurate, does it cover the right topics? Those are necessary. They’re not sufficient.
Tone is the variable that determines whether content that is technically correct actually works. A technically correct piece of content in the wrong voice will fail with its intended audience. A slightly imperfect piece in exactly the right voice will convert.
Marketers who have been doing this for decades know this intuitively. The problem is that AI systems don’t. Left to their own defaults, they produce content in a generic middle-register that offends no one and resonates with no one. Competent. Invisible.
The Brand Voice feature was built to solve that problem.
The Brand Voice Feature: A Text Input, a Toggle, and 30 Years of Logic
The implementation is deliberately simple: a text input field where the user describes their brand voice, and a toggle that activates it on every piece of content the system produces.
The simplicity is intentional. Brand voice documentation is often buried in a 40-page brand guide that no one reads. Distilling it into a text field that informs every content output forces clarity: what actually defines how this brand sounds? Not the list of adjectives in the style guide. The real, operational answer.
To explain the concept to users who haven’t worked in brand marketing, I used the press release framework. A press release has a specific, recognizable voice: authoritative, factual, structured, third-person. Any marketing professional who has written one knows exactly what that voice sounds like. Using it as a reference point bridges the gap between an abstract concept (brand voice) and a concrete operational instruction the system can follow.
The Final Polish Bug
The trickier problem was what I call the Final Polish Bug: the tendency of AI-generated content to add filler at the end.
You’ve seen it. The conclusion that summarizes everything you just read. The closing sentence that tells you the content was “comprehensive.” The paragraph that explains why the topic matters, after spending 800 words demonstrating exactly why it matters. It’s the AI equivalent of a nervous presenter who keeps talking after they’ve made their point.
The fix was system-level instructions that explicitly prohibit the behavior: no summary conclusions, no restating the headline, no meta-commentary on the content itself. The content ends when the content is done. Not when the AI decides it has demonstrated sufficient thoroughness.
Those instructions, once added, produced a measurable shift in output quality. Agency-quality writing isn’t just about what goes in. It’s about what gets cut.
The Agency Subscription Package
Day 7 also introduced the agency subscription tier. The approach to designing it: let the AI design its own best offer.
I described the customer. A marketing agency producing content for multiple clients. And asked the system what package structure would deliver the most value for that use case. The output informed the final packaging: a subscription designed around the agency workflow rather than around the features available to build.
The result is a package that makes sense to the buyer because it was designed from the buyer’s perspective. That’s not a novel insight in marketing. It is a novel use of AI in product development.
The Moment the Product Was Ready
Day 7 is where Paxelo stopped being a side hustle experiment and became a product ready to meet customers.
Not because every feature was built. Because the core promise. Content that sounds like your brand, structured for your audience, produced in minutes instead of hours. Was fully functional and demonstrably working.
The machine had a voice. It knew how to use it. And it knew when to stop talking.
Watch Day 7: YouTube
Paxelo Build Series Day 8: The Day My Side Hustle Stopped Feeling Like an Experiment