A client came to us with an unusual request. They were building a mobile word game and needed a team to generate creative content for it. Not translation. Not cultural adaptation. Original content, from scratch. In English. With humor.
We’re a localization company. We’ve spent 20+ years making products sound natural in any language. But this was different — and there was no playbook. We love that kind of thing.
Three people picked it up:
Dmitriy, our prompt engineer, who knows how to talk to LLMs in their own language.
Nadya, a localization PM, who keeps everything together: the client, the linguist, deadlines, expectation
And a linguist — a native speaker with a feel for what lands with an American audience and what doesn’t.
What the game actually needed
The player sees a chain of word pairs where the last word of one pair becomes the first word of the next. Each pair comes with a witty definition. Every pair is unique; no repeats. We needed two hundred of these chains, five pairs each.
Here’s one to give you the idea:
The thing everyone talks about, but the snacks are more exciting → Main Event
The person who arranges things, but secretly enjoys chaos → Event Planner
The place where you write down all the things you never do → Planner Book
A gathering of people who mostly just drink wine and talk about not reading → Book Club
The only sandwich that’s more layered than your personality → Club Sandwich
Each word pair is real and meaningful. The definitions are funny. And the whole chain connects logically: one word hooks into the next. Now multiply that by two hundred, with no pair repeated anywhere.
A fun challenge, for sure, but too much for one person to juggle: uniqueness, humor, logical links, cultural context for Americans. AI alone wouldn’t cut it either — a model can generate forever, but it has no taste. It can’t tell what’s funny from what’s flat.
The most interesting part was building a setup where people and machine each did what they’re best at.
First attempt: “just prompt it”
The logic seemed simple. Take an LLM, feed it the client’s requirements, generate the chains, hand them to the linguist for review — done.
In practice, it was messier. The model didn’t understand “no duplicates” — to it, “Book Club” and “Book Club!” are different things. Without strong examples upfront, it produced dull, template-like definitions. And when we crammed all the requirements into a single prompt, the model got lost and output something average.
The first iteration flopped. But that flop gave us a clear picture of where exactly things broke and what to do next.
How we rebuilt it
Dmitriy tried several approaches and found the key insight: one model can’t build a logical chain and write funny definitions at the same time. So he split the task between two AI agents. The first one assembles the chain skeleton — pure logic. The second writes a humorous definition for each pair.
Claude turned out to be funnier than GPT, and the linguist picked Claude’s versions again and again. Extended thinking mode made the difference: without it, the model produced surface-level definitions. With it, the model actually worked through unexpected angles.
Dmitriy also wrote scripts for automated uniqueness checks (because you can’t trust an LLM on that: a period instead of an exclamation mark, and the model thinks it’s a different phrase).
Meanwhile, Nadya ran her part: passing results to the linguist, collecting feedback, sending deliveries to the client. The linguist filtered everything: kept some, cut others. Anything that didn’t pass went back for regeneration, this time with feedback baked in. Each iteration got sharper.
How it came together
Stepping back, the whole thing was easier and faster as a team than it would’ve been for anyone working alone. Dmitriy built the pipeline and talked to the machine. Nadya managed client communication and coordinated the linguist. The linguist shaped the creative quality. Everyone handled their piece — and it clicked.
Two to three iterations, about a week and a half from start to final sign-off. The task felt like a challenge worth solving, not a grind. The process stayed manageable, the output got better with each round, and nobody ended up burned out.
The result
“We think that your team has done a really good job with these, so we are happy to give final sign off. Many thanks for your help!”
Getting that email felt good. Not because of the praise, but because we’d taken on something we’d never done before, figured it out, and it worked. A bit of nerves, a bit of excitement — and it came together. The client was happy, and we came out stronger.
But the real result isn’t two hundred finished chains. It’s the pipeline. A system that can generate game content at any scale. The client didn’t just get a one-time content delivery — they got something they can build on. Add new languages, generate hundreds more sharp, playful chains that keep players coming back.
Why we’re telling this story
Tasks like this will keep coming — not just for us, but everywhere. Where there used to be a clear process, there’s now a blank space: “we’ve never done this before.” That’s not a bug. That’s the new normal.
The future belongs to teams that adapt quickly and reshape their process around the task, not the other way around. We didn’t know how to create content for a word game. There was no template, no reference, no one who could say “here’s the right way.” Just a client request, a deadline, and three people who decided to figure it out.
And they did. Dmitriy chose the model and figured out how to split the task. Nadya built the process with the client on her own. The linguist didn’t just proofread — he shaped the taste of the entire output. Nobody waited for instructions.
We think that’s how it should work. Trust people for real, not just in theory. Not approving every step. Not layering process on top of process. Hand over the task and step back. Professionals will find the way. And if they don’t know how yet, they’ll learn fast.
Next time there’ll be a different unfamiliar task. We’ll figure that one out too.
If you’ve read this far and thought “that’s my kind of team” — take a look at our Talents Pool. We don’t always have open positions, but we’re always glad to meet great people. Tell us about yourself, and maybe soon you’ll be the one working on projects like this, side by side with people who are just as into it as you are.
QACAT
AI Video Studio
Localize.io