January 2026
The bottleneck to building at scale isn’t computing power or headcount. It’s context.
When a team is small, everyone knows everything. You overhear conversations. You remember why decisions were made. The whole project fits in your head.
Then the team grows. Suddenly nobody knows the full picture. So you add meetings to sync. Documents to hand off. Status updates to keep people informed. All of it exists because humans can’t hold enough state.
This is the thing AI changes. Not by making those processes faster. By making them unnecessary.
*
The trick is making context discoverable. Not loading everything into the AI upfront—that’s expensive and noisy. Instead, let the AI pull what it needs, when it needs it.
We call this “dynamic context discovery.” The idea is simple: instead of stuffing everything into the prompt, you give the AI pointers to where things live. It reads what’s relevant and ignores the rest.
The primitive that makes this work is the file.
Files are underrated. They’re simple, universal, and AI already knows how to read them. Terminal output? Write it to a file, let the AI grep through it. Long API response? Save it as a file, let the AI read what it needs. Chat history getting too long? Summarize it into a file the AI can reference later.
Every piece of context becomes a file the AI can find. Not a database query, not an API call. Just files.
*
This matters because context windows are expensive. The more you load upfront, the slower and costlier every request becomes. Worse, irrelevant context confuses the model. Less is often more.
Static context is what you always include: system prompts, core rules, basic setup. Dynamic context is everything else. The AI discovers it as needed.
A practical example: instead of sending every tool description in every prompt, you sync tool descriptions to a folder. The AI gets a list of tool names. When it needs one, it looks it up. Cursor found this reduced tokens by 47% in agent runs that used tools.
Same idea with skills. Instead of baking instructions into the prompt, you write them as files. The AI searches for relevant skills when the task calls for it.
*
We think about this in terms of three things: speed, simplicity, and quality.
Speed means the AI finds context in seconds, not minutes. No waiting for someone to write a summary. No hunting through old docs. The context is there when you need it.
Simplicity means fewer abstractions. Files are boring, but boring is good. Everyone understands files. They work everywhere. You don’t need a new system—you need a folder.
Quality means the AI works with clean, relevant context. Not everything at once. Not a wall of noise. Just the pieces that matter for the task at hand.
These three reinforce each other. Simpler systems are faster. Faster discovery improves quality. Quality reduces the need for complex processes.
*
The strange thing is what happens over time.
Context accumulates. Every terminal session, every decision, every conversation—it all becomes files the AI can reference. After a few months, the system knows things no single person remembers.
A small team with discoverable context can outbuild a large team without it. That’s the real unlock. Not AI doing the work for you, but AI pulling the right context at the right time.
Most companies won’t do this. It’s not obvious that files are the answer. It feels too simple.
But simple primitives win. Shared context. Dynamic discovery. Speed, simplicity, quality. Everything else follows.