Act I: Where We Started
The “figuring out what this AI thing can actually do” phase
Every good story starts with curiosity and a little naivety. Early in the year, we were still poking at the edges of what was possible. Testing models. Exploring workflows. Asking, “Wait… can it do that?”
This phase was about learning the surface area of new tools and models before we could trust them.
AI Ticket Whisperer & Figma Ticket Writer
The Ones That Slotted Into Real Work
These were some of the first experiments that didn’t feel like just a side project. They fit directly into how we already work. Starting with design documentation in Figma, then evolving into AI-assisted ticket creation inside our actual tooling, this line of work showed us what happens when AI is designed into a workflow instead of bolted on afterward.
O.B.S.C.O.R.E.
First MCP Experiment
This was one of our first real experiments with model-context protocols, tools, and orchestration. We used it to pull in music metadata, wire external sources into an LLM, and force the model to reason across real data instead of hallucinating its way through an answer. It started as a music project, but it taught us how powerful tool-driven workflows could be. A pattern we’d come back to again and again.
Model Showdown
The Right Tool for the Job Moment
A side-by-side look at different AI models. Less “which one is best” and more “which one is best for what.” It turns out the answer depends on the question. A theme we’d keep returning to all year.
Act II: Where It Got Weird
The “we can’t believe we actually shipped this” phase
At some point, the questions stayed serious, but the experiments got stranger. Getting weird turned out to be a useful way to think differently. Call it creative exploration.
When a question didn’t fit neatly into a demo or a dashboard, we let the form loosen up. These were the projects that made us laugh, question our sanity, and occasionally made our spouses look at us funny. But that’s where a lot of the real learning happened.
Coach! 2.0
The One That Actually Made Our Spouses Laugh
The original Coach gave you Friday Night Lights–style motivation for your to-do list. Version 2.0 turned it up to eleven. Clear eyes, full hearts, can’t lose. Underneath the chaos, it was a test of how new tools like AI voices and video can change the emotional resonance of a product. How far you can push personality before a tool becomes distracting instead of motivating.
Nugget Tycoon
Most Ridiculous Use of LLM Memory
We wanted to understand how LLMs behave when they’re allowed to remember over time. Not just chat history, but layered memory. What to store, what to summarize, what to forget, and how decisions change when context actually persists.
So we took it one step further and, of course, built a chicken nugget empire simulator 🦖 🍗 🏦. Absurd on the surface. Surprisingly useful for testing long-running memory, narrative state, and agent behavior over time.
SSMoL (A Small Solar Model of Language)
Most Ambitious (“Matt, are you okay?”)
It started as an offhand comment while looking at a computer running a local LLM. “Could you hook a meter up to it… or power it via solar?” One of those questions you ask out loud without expecting an answer.
So we tried to answer it. We built a tiny computer, ran a small language model on-device, and powered it with solar. Completely unnecessary. Incredibly informative. It forced us to think about constraints, energy, latency, and what “local” actually means.
Act III: Where It Actually Started to Work
The “oh wait, this is genuinely useful” phase
And then there were the experiments that stopped feeling like experiments. The ones that we actually started to use. The ones that quietly shaped how we approach client work. The ones that proved this wasn’t just play.
Victory Roast
The Only One That Actually Made Money
We didn’t make a million dollars. But we made more than one dollar. In the world of side experiments, that counts.
Launch Protocol
The One That Changed the Conversation
A playful, side-scrolling way to explore scope and tradeoffs. Instead of explaining our process, people experience it. It became a way to explain the way we work, from prototype through MVP.
SoleMate 3D
The Show, Don’t Tell One
Sometimes the fastest way to explain what you can build is to just build it. When someone asked how we might approach a 3D project, we didn’t explain it. We showed them.
Mocktopus
You Never Have to Spend a Day Clone-Stamping Again
A mockup generator that actually respects the image you put into it. And this one got some attention when we shared it. Designers reached out asking, “Can I use this and never have to spend a day clone-stamping again?” Turns out a lot of people have been waiting for that.
Chief
The “Oh, This Might Actually Be a Real Product” One
An AI-powered BI assistant for agency life. Ask questions in plain English. Get real answers. Bookmark what matters to your dynamic dashboard. It became one of the most valuable things we built all year. Maybe we’ll never have to create a complicated spreadsheet again.
Coming Attractions: What’s Next in 2026
🖥️ More of what we do.
In 2025, We Built a Thing focused heavily on experiments. A way to explore totally new ideas. And a way to share work faster than we can usually show in production client work.
In 2026, we hope to share more of how Township actually operates day-to-day. The systems we design, the tradeoffs we navigate, and the real problems we help teams solve. That includes more than web apps and AI experiments. We also build connected devices, medical tech, and systems that exist in the physical world. Some of the harder problems we solve. The moments where software meets reality and has to hold up.
💪 New formats. New muscles.
Not everything we learn shows up as a prototype. Some of the most valuable insights are about how to think, how to design, and how to build in this new era. Over the past year, a few themes have started to take shape. New ways of developing software. More opinionated workflows. Tools that are smaller, more personal, and more specific to the teams using them. We’ve also been collecting lessons from the past year and shaping them into something more durable.
👥 More voices from the team.
You’ve mostly heard from me (Matt). You’ve heard some from Caleb too. But we have a team that’s deep in this work, learning fast, adapting workflows as tools change daily, and forming opinions worth sharing. In 2026, we want this newsletter to feel a little more like a chorus. Different perspectives from the team, different lessons learned, same spirit of curiosity.
