Lizzie: the agent that turned tone of voice into a spec

Integrate the apps and services you already use to streamline workflows.

In 1904, a woman named Lizzie Magie patented a board game about land, rent and the way money concentrates when nobody is watching. Decades later, the world would know that game as Monopoly, mostly without knowing her name. When Magie was founded, she was part of the story from day one. The company name itself was the tribute. We wanted to go further and give her a visible place in the product, and we kept failing to find one. There was no screen, no feature, no moment in a WhatsApp conversation about paying a bill where a reference to a game designer from the early 1900s would land. So the reference stayed where it started: in the name, and nowhere else.

Context

Magie started as a conversational bank on WhatsApp. AI agents handled payments, transfers and balances for Magie's own users, in Magie's own voice. Then Magie went B2B. Each client organization now deploys its own branded agent, running on Magie's technology and handling real financial operations for that client's end users, on behalf of that client's brand. A telecom, a retailer and a fintech can all run agents on the same platform. I am the only designer on the team and part of the founding group, working day to day with Product, Engineering and the people who tune what the agents say.

The problem

Every agent sounded the same. It did not matter whether the client was a telecom or a travel company: the agent that answered their users spoke in the same neutral register, because there was no structured way to define how a given brand should communicate and no way to configure it once defined. Tone was handled by editing prompts by hand, one adjustment at a time, by whoever had the context.

Three things made this urgent. Clients expect their agent to sound like them, and a generic chatbot wearing their logo is a broken promise. The handoff between teams had no artifact: design made tone decisions, engineering implemented them, and the translation between the two lived in chat threads and documents that nobody could reassemble later. And with several clients going live in the same period, hand-crafting a voice per organization was already unsustainable before the pipeline had really filled.

My role and the decisions that shaped it

I identified the problem, framed it, designed the tool and built it. The obvious fix was more process: a tone template for onboarding, a checklist, a review step. The decision that mattered was to treat voice as a design and AI problem and build an agent whose job is to turn brand material into configuration. That is where the reference we had never managed to place finally found its job. Lizzie came back as Magie's Voice Architect.

Three choices defined the work:

Accept brand material in whatever shape it exists.
Requiring a formal brand guide would have excluded most real clients. Lizzie takes guidelines, social media posts, support message samples, screenshots, website copy, pasted text. Anything that shows how the brand talks. It also refuses what does not belong: hand it something that carries no brand information and it explains why, and what would work instead.
Keep humans in the review seat.
Everything Lizzie infers from the material arrives pre-filled and visibly tagged as inferred, so the team knows what came from the AI and what came from a person. The main dimensions of tone are set through quick selectors, with open fields for the parts that resist a scale, like how the brand would behave as a guest at a dinner party. When the material is ambiguous or contradictory, Lizzie flags it instead of silently picking a side. The trade-off was an extra step in the flow. I took it, because a voice profile that ships with hidden assumptions is worse than one that takes two more minutes.
Design the output for the people who have to implement it.
The result of Lizzie's work is a complete Brand Voice Profile, exported as a single structured file that engineering can use directly. This was the decision engineering shaped most, because they are the ones consuming it, and it is what turned the output from readable into implementable. I built the first version in five days, on my own, without pulling engineering off their roadmap, and it shipped with feedback loops from Engineering and GTM.

My role and the decisions that shaped it

Three choices defined the work:

Accept brand material in whatever shape it exists.
Requiring a formal brand guide would have excluded most real clients. Lizzie takes guidelines, social media posts, support message samples, screenshots, website copy, pasted text. Anything that shows how the brand talks. It also refuses what does not belong: hand it something that carries no brand information and it explains why, and what would work instead.
Keep humans in the review seat.
Everything Lizzie infers from the material arrives pre-filled and visibly tagged as inferred, so the team knows what came from the AI and what came from a person. The main dimensions of tone are set through quick selectors, with open fields for the parts that resist a scale, like how the brand would behave as a guest at a dinner party. When the material is ambiguous or contradictory, Lizzie flags it instead of silently picking a side. The trade-off was an extra step in the flow. I took it, because a voice profile that ships with hidden assumptions is worse than one that takes two more minutes.
Design the output for the people who have to implement it.
The result of Lizzie's work is a complete Brand Voice Profile, exported as a single structured file that engineering can use directly. This was the decision engineering shaped most, because they are the ones consuming it, and it is what turned the output from readable into implementable. I built the first version in five days, on my own, without pulling engineering off their roadmap, and it shipped with feedback loops from Engineering and GTM.

What Lizzie does

Lizzie works in three steps:

Import: the team uploads or pastes the client's brand material, and Lizzie reads all of it in one pass, extracting who the brand is, who it talks to, how it sounds, what it says and what it never says.

Review: the team adjusts what was inferred and fills what was missing.

Generate: Lizzie produces the profile, grounded in what the agent can actually do for that client, so every example it writes is a conversation that could really happen.

The profile has four parts. A short voice summary. A set of scored tone axes covering formality, warmth, brevity, humor, emoji, how errors are handled, greeting pattern, persona, vocabulary and cultural register. Example interactions mapped to real operations, from the first greeting to a transfer, a bill payment, a balance check and the moments when something goes wrong. And the engineering handoff.

The engineering handoff

This is the piece that fixes the broken handoff, and the reason Lizzie exists as a tool rather than a document. For each organization, the profile delivers everything engineering needs to configure the agent: a complete system prompt written in the brand's voice, concrete formatting rules for WhatsApp, response templates for each way a request can end, specific messages for each operation, templates for the error scenarios, and the configuration values the agent needs. An engineer opens one file and has the whole voice. No interpretation, no back-and-forth, no digging through old threads.

Results and impact

The first version of Lizzie shipped in five days, with zero engineering cycles consumed. Output generated by Lizzie is already running in production for a client, and every organization on the platform can now have its voice translated the same way, from the same kind of raw material, into the same structure.

Produced

faster voice configuration: from 25 hours of manual prompt work to 40 minutes per client.

Reduced

-0%

review rounds with engineering to get a client's tone right: from four or five to one.

Processed

Brand materials from B2B clients in less than 5 months of go-live.

Beyond the numbers, the project set a precedent for how design operates at Magie: a brand consistency problem that would normally take a committee and a quarter got solved by a designer, in a week, with code. And the name we could not fit anywhere for years is now the thing that makes sure every brand on the platform sounds like itself.