The Roadtrip

Boredom is optional nowadays. We are always pursuing some source of entertainment, and at the slightest sign that dopamine will cease, we promptly get our phones to keep it flowing.

That's very unfortunate, as sometimes you have to stay alone with your thoughts to make sense of things, to build your own understanding, and also to have great ideas, or at least interesting ones.

In 2025 I had to drive my wife to perform a surgery in another state. It was a neighboring state, but as Brazil is huge, it was enough for me to drive 8 hours on each leg. When we were returning, with my unusually quiet wife beside me, recovering from surgery, I was left alone with my thoughts as the road passed by.

At the time, and really constantly since the current wave of AI started, my thoughts revolved around the FOMO of not building something and the dread of being useless if true AGI comes. So many build ideas have crossed my mind that I've lost track of them, but they always get rejected by my market research department. A simple google search is enough to find ten different implementations of any idea that I have. Sometimes it feels like everything that could be built is already being built.

Then I started to think about some frustrations that I have. What personal problems could I really solve using AI?

My adventures as a DM

I had a small group that played D&D once in a while. In the Brazilian countryside we didn't have much access to D&D content, as one can imagine, so we did our best to gather material online and play when we could. During the pandemic I was the DM for a few sessions, but preparation was always hard for me. With a lot going on in my life, finding a few hours a week to prep the next session felt almost impossible, and I always felt like I wasn't covering everything. I'm not great at improvisation, so I tried to prepare in advance for the most likely ways the story could evolve. It was overwhelming.

The table died slowly. First I started postponing sessions because prep wasn't ready. Then they started postponing because life got in the way. The group chat that used to blow up after sessions turned into "next week?" with no replies. The campaign just sat there, frozen mid-quest.

First Iteration

Back to my head during the roadtrip. I realized AI could help with my DM struggles. Maybe we could resurrect the table with an AI DM. Of course it's not that simple. As anyone who's played AI Dungeon or similar knows, AI alone is not a great DM: hallucination, consistency errors, failure to follow rules, insufficient context, the usual suspects.

We need something way more robust than a simple AI with prompt tuning. The good news is that D&D and other RPGs have deterministic rules. So maybe we can offload rule application from the AI entirely. AI is also notoriously bad at random number generation, but that's easily solved with a pseudo-random function.

But what about the sheer volume of information a rich world contains? How can we load all of that into an AI without overwhelming it?

Well, when you interact with the world, you don't interact with everything at once. You're always at a specific place, with specific NPCs, items, monsters, and so on. Other elements might be relevant, but usually only if they're close in some dimension to where you are.

Fortunately, computer science provides just the right tool to model that: graphs.

If we model the whole world as a graph, we can load only the nodes relevant to a given turn and let deterministic rules handle the rest. The LLM ends up with a much smaller job.

How Cursor + Sonnet 3.5 Failed Miserably

Once I finally got home from the roadtrip, I started turning this vision into code. I grabbed some D&D SRD elements and started loading them into a graph. That was the intent, at least. Everyone was talking about vibe coding at the time, so I figured I'd give it a try. I wouldn't have time to implement it manually anyway.

I fired up Cursor with Sonnet 3.5, the cutting edge at the time. It failed miserably.

After a few weeks, I couldn't even get a reliable pipeline for SRD data extraction working, let alone build the whole system. It kept getting stuck in loops, flip-flopping between approaches. It just didn't work. Hands-on, it failed to stay consistent with my guidelines. Hands-off, it got overwhelmed and stalled.

Maybe a skill issue, but it is what it is.

After a few weeks of fruitless effort, I gave up. The idea stayed in the back of my head. Maybe it was just a matter of time until the tech caught up.

What Changed in 2026

Now in 2026, I had some time to play with the newest tech. Opus 4.5 generated a lot of buzz and I was finally able to test it. On my first try, it solved in two hours what Sonnet 3.5 never could. In less than a week, I had a working first version of what I'd envisioned.

After that, Claude Code hooked me. There's something genuinely addictive about building fast and seeing things work. The rate of progress slows as the codebase grows, of course, but I still feel like I'm shipping what a full team would, without the communication overhead, and hopefully without burying myself in technical debt.

What I'm building now

After a year of the idea settling and crystallizing, it has a much clearer shape and a new name.

Tapestria isn't just a graph-driven RPG with a deterministic engine. It's being built to sustain a shared world. All players will interact with the same NPCs, the same locations, the same world. The lore will evolve on its own through players' interactions and decisions. The main story drivers will be quests, which can take many forms. No two will be the same, and each will push the world forward differently.

Several different concepts converged into this vision, and I guess they only did because I was left alone with my own thoughts long enough for them to wire themselves up in the background.

I'll be writing here every week as I build Tapestria in the open: the world, the engine, and the first players willing to seed it.