As part of our regular practice at Clear Box, we pause each week with our AI historian to take stock of what we’ve built. This reflective process helps us spot drift and uncover the ideas that live between our best-laid plans.
You install a piece of software, open it, and wait. Your computer begins processing ten thousand tokens of instructions. Are you still there when it’s ready? Or have you already wandered off to do something else?
This is a critical friction point for local AI running on limited hardware. An agentic tool like OpenCode requires precise instructions to function effectively. The paradox is that the more carefully we prepare an assistant—the more context we give it to ensure quality—the more we risk creating a miserable first experience.
When a machine takes several minutes to work through its briefing, the user is already gone.
I wanted to know if we could move that preparation elsewhere. Could someone with deep subject matter expertise prepare the instructions and reference materials, process them once on a powerful machine, and then ship that resulting “memory” with the software? In that scenario, the user’s computer would simply load the state and get straight to work.
So we tested it. We prepared roughly 12,000 tokens of instructions, product documentation, and a triage playbook using my NVIDIA RTX 5090. We then took that saved memory and deployed it across various CPU and GPU setups on both Windows and Linux, keeping the model and engine versions constant.
It worked.
We cooked it once, packed it up, and served it on another machine. Same preparation, same useful decisions. Huzzah!
On the tested CPUs, processing those instructions from scratch took roughly eight minutes. Loading the prepared memory took about two and a half seconds. While the assistant still had to process the specific user question and generate a response, it did so with all its preparation already intact. Across our three triage questions, every setup reached the correct decision, regardless of minor differences in wording.
This was the proof of concept I’d been waiting for. We can now prepare a specialist on powerful hardware and migrate that working context to another machine. The idea of shipping pre-processed preparation with a product now has concrete evidence behind it.
This opens up compelling possibilities for what we put in a user’s hands. Imagine a local technician for your Mac, PC, or Linux machine that arrives pre-loaded with all relevant documentation. Imagine an OpenCode expert ready to help you tweak and troubleshoot your installation the moment it’s finished. We could essentially transform static user manuals into interactive AI—something you can simply ask a question of the moment you get stuck, running entirely on your own hardware.
The same logic applies to specialist workflows. Asking a single general-purpose assistant to review code gives you one perspective. I prefer using multiple reviewers, each with different areas of expertise and specific rules about what they should and shouldn’t examine. Because those detailed instructions occupy significant space, preparing them in advance makes that level of specialization practical for local users.
There is still more to explore—specifically whether we can compress these prepared memories enough to make them viable for transfer or inclusion in an installer. But this cross-machine test provides a concrete foundation:
The computer doing the preparation does not have to be the computer running the result.
A machine generating 15 tokens per second can still do useful work. But forcing a user to wait through a massive setup prompt every time they open an app is a good way to ensure they never find out. If we want people to give local AI a real shot before cloud services run away with the market, we have to put our best foot forward from the first second.