I built a tool that tells you how your brand actually sounds versus how you think it does.
5 min readShrey Mehra
Real work, in build. Not live yet; the link goes here when it is.
Role
Solo. Product definition, the build itself, and the positioning.
Brief
Self-initiated. It came out of a pattern seen across client work: brands with strong guidelines that keep producing content which misses them.
Constraint
A solo build, so the reasoning had to come from a general model rather than a trained classifier.
State
In build. Not live, and not linked until it is.
Problem
The most discussed element of brand strategy, and the least measurable
Companies run workshops, write guidelines, produce a document that lives in a shared drive, and then the content, or three different agencies, produces work that sounds like several different people. The gap between the stated voice and the actual voice is real, persistent and expensive. Nobody has a fast diagnostic for it.
What does your brand actually sound like, right now, in this piece of content? Not what the guidelines say it should sound like. What it does.
The build
React, Claude API, Vercel
The choice of Claude specifically was deliberate. This is not a sentiment tool. Positive, negative, neutral is a blunt instrument for a nuanced problem. Voice characterisation requires reasoning across register, rhythm, specificity level, formality, warmth and authority simultaneously. A custom-trained classifier for a solo build would not do that.
The build was motivated by a pattern observed across client work: brands with strong guidelines keep producing content that misses them. Not because the team doesn't understand the guidelines, because there is no feedback loop between production and intent.
Key decisions
Three calls, and what each one cost
A tool is a series of refusals. These are the three that shaped what it does and does not do.
Decision 01
A reasoning model, not a sentiment classifier
Options
Sentiment analysis · a custom-trained classifier · the Claude API
Why
This is not a sentiment problem. Positive, negative, neutral is a blunt instrument here. Voice characterisation means reasoning across register, rhythm, specificity, formality, warmth and authority at once, and a custom-trained classifier for a solo build would not do that.
Trade-off
Every run costs money and takes time, which a local classifier would not. The output also has to be held to a fixed shape, or two runs cannot be compared to each other.
Decision 02
A gap, not a score
Options
A single voice score · a gap map between stated and actual
Why
Stated voice and actual voice are placed side by side, and the space between them is the brief for the edit. A number tells a team they have a problem. The gap tells them where it starts.
Trade-off
A score travels on its own and gets screenshotted into a deck. A gap map has to be read. It trades shareability for the thing that is actually useful in the room.
Decision 03
The brand has to say what it is aiming for
Options
Infer the intended voice from the content · require guidelines, reference examples or stated principles as input
Why
A tool that guesses the intent and then measures against its own guess is measuring nothing. The diagnostic only means something if the target came from the brand.
Trade-off
It will not work for a brand that has never written its voice down, which is a real share of them. The friction is the price of the answer meaning anything.
What it does
Makes the gap visible
Input a piece of content. Input how the brand says it should sound: guidelines, reference examples, stated principles. Voice’D returns a gap analysis across voice dimensions: what's landing, what's drifting, and where the intervention starts.
The output is visual. The brand's stated voice and the content's actual voice are placed side by side. The space between them is the brief for the edit.
The output, at the size it is readIllustration. Two fictional brands, so nothing here claims to describe a real one. Not a live run.
What the brand says it sounds like
Plain-spoken and warm. We explain money the way a friend who happens to know would.
The content it sent
“Our platform leverages a comprehensive suite of financial instruments designed to optimise outcomes across a diverse range of customer circumstances and objectives.”
Registerconversational / institutional
Rhythmshort, varied / long, even
Specificityconcrete / abstract
Formalityplain / formal
Warmthwarm / cool
Authoritypeer / expert
StatedActualThe gap is the brief
Where the edit startsFormality and specificity are the widest gaps, and they are the same problem twice. The sentence names no instrument, no outcome and no customer. Rewrite it around one of each and the other four dimensions move on their own.
What the brand says it sounds like
Playful and direct. We talk about skin like adults, and we are not precious about it.
The content it sent
“A clinically validated formulation containing 2% encapsulated salicylic acid, indicated for the management of comedonal congestion in combination-oily skin types.”
Registerconversational / institutional
Rhythmshort, varied / long, even
Specificityconcrete / abstract
Formalityplain / formal
Warmthwarm / cool
Authoritypeer / expert
StatedActualThe gap is the brief
Where the edit startsOnly formality and authority are genuinely off. Specificity runs ahead of the guidelines: the copy is more concrete than the brand asked for, which is the guidelines being wrong rather than the writing. That is the more useful finding and the one a workshop never produces.
This page shows the shape of the answer, not an answer. A captured run and the live link follow when it deploys.
Learning
Most brands don't know what their voice sounds like
They know what they want it to sound like. The gap between aspiration and reality is a product, and it is also the most honest brief a brand strategist can work from.
The secondary insight: a tool that makes the gap visible changes the conversation from “this doesn't feel right” to “here is specifically what is off and here is where it starts.” That shift from intuition to evidence makes client relationships faster and more productive.
Once it is live, this will be the only page on this portfolio where a visitor can see something working, not just read about it. The rest of the work here is thinking made visible. This is thinking made functional.
