For the past few weeks, I've been running an AI assistant on a Mac mini in my apartment. It manages my study tracking, helps me write code, drafts blog posts, keeps tabs on my projects — basically acts as a second brain while I'm on clerkship rotations and have zero spare cognitive bandwidth.
I started with Telegram. One chat thread, one bot, everything in one place. It was simple. It was also, I eventually realized, kind of stupid.
Not the model — the model was fine. The interface was the problem.
The single-thread trap
Here's what a single Telegram thread looks like after a few weeks of actual use: a message about a UWorld question, then a request to check on a coding project, then something personal, then back to studying, then "hey can you look at this error log," then a question about dinner.
The AI sees all of this as one continuous conversation. It has no idea if you're context-switching or continuing a thought. It doesn't know if "fix that bug" refers to the thing you mentioned three messages ago or thirty. Every topic bleeds into every other topic, and the AI's context window becomes a junk drawer.
I kept thinking the fix was a better model with a longer context window. It wasn't.
Rooms, not threads
I moved everything to Discord. Not because Discord is inherently better chat software — it's not, it's kind of a mess actually — but because it has one thing Telegram doesn't: channels.
Now my setup looks like this:
- #uworld — study tracking. 1,700+ questions logged. The AI knows that when I'm in this channel, I'm studying. It doesn't try to talk about my side projects.
- #blog — writing. This post was drafted in that channel.
- #chartlens — a coding project. Bug reports, feature ideas, architecture discussions. All contained.
- #vicky — personal. Long-distance girlfriend. There's a lot to coordinate.
- #chat — general stuff that doesn't fit anywhere else.
Same AI. Same model. Same Mac mini humming on my desk. But the behavior changed completely, because the context changed.
When I ask a question in #uworld, the AI doesn't waste tokens wondering if I'm talking about a React component. When I'm in #chartlens debugging an API endpoint, it doesn't suggest I review my Anki cards. The channel is the context. The AI knows where it is and acts accordingly.
Threads are even better
Discord channels give you topic isolation. Discord threads give you task isolation.
Need to debug a specific issue? Spawn a thread. The AI gets a fresh context window scoped to that one problem. Work on it for twenty minutes, close it, go back to the channel. Meanwhile, another thread is running in parallel on a different task with its own isolated context.
This is the thing nobody talks about: parallel work without context bleed. In a single chat thread, you can't work on two things at once. In Discord, I regularly have three or four threads going — one drafting a blog post, one tracking down a bug, one planning a feature — and none of them interfere with each other.
The interface is the intelligence
There's a popular debate right now about which AI model is best. GPT-4 or Claude or Gemini, benchmarks and vibes and leaderboard rankings. I get it. Models matter.
But I've learned something running my own assistant for months: the gap between a well-structured interface and a poorly-structured one is bigger than the gap between models.
A brilliant model in a single chat thread will eventually drown in its own context. A decent model with clean channel separation and isolated threads will outperform it on real tasks, because it's always working with relevant context instead of sifting through noise.
A single chat box is a constraint disguised as simplicity. It feels frictionless — just type and go. But that frictionlessness means you never organize your thoughts, and neither does your AI. You're both wading through the same mess.
The boring lesson
People spend hours debating model selection. They should spend ten minutes thinking about how they structure the conversation.
Give your AI rooms. Give it walls between topics. Give it the ability to focus on one thing at a time without the cognitive equivalent of someone tapping its shoulder every thirty seconds with an unrelated question.
The model is the engine. The interface is the steering wheel. You can have the most powerful engine in the world, but if you can't steer, you're just making noise.
I didn't make my AI smarter. I just stopped making it guess what I was talking about.