Instructions / Local models Back to the garage →

001Field guide

Run your own AI.
On your own damn computer.

No API keys. No terminal wizardry. No 47-tab YouTube rabbit hole. You copy my prompts, a smart AI does the heavy lifting (and asks before it touches anything), and you end up with a model that runs on your machine, offline, for free, forever.

yes, even if you think "terminal" is an airport thing

  1. 01

    Borrow a big brain

    One month of Claude or ChatGPT. It's the contractor. It does the setup so you don't have to.

  2. 02

    Paste my prompts

    Step by step. Each one has safety rules baked in, so the AI explains first and asks before it runs anything.

  3. 03

    Fire the contractor

    Cancel the subscription if you want. Your local model doesn't care. It lives on your computer now.

01

Step one

Get yourself a helper

First things first. You need an AI that can actually do stuff on your computer, not just chat about it. Claude and ChatGPT both have desktop apps that can run commands, with your permission. So:

Do you pay for Claude or ChatGPT?

02

Step two

What's under the hood?

Before we download anything, we need to know what your computer can handle without crying. This step is look, don't touch. The AI only reads info about your machine. It doesn't install or change a thing.

The one rule

The whole AI setup (the model, the engine running it, and its memory of your chat) has to fit in 80% of your RAM. The chat memory (called "context") gets 50% of the model's size. Leaves breathing room so your computer doesn't turn into a space heater.

Prompt 02 · Hardware check 
[SAFETY RULES: Kevin's security checks go here]

You're helping me set up an AI model that runs locally on this computer. This step is READ-ONLY. Look, don't touch.

1. Find out my operating system, CPU or chip, total RAM, graphics card and its memory (VRAM) if I have one, and free disk space.
2. Only use built-in, read-only commands. Don't install anything, don't change any settings, and never use sudo or admin.
3. Before each command, tell me in plain English what it does, then wait for my OK.
4. When you're done, give me this summary and nothing else:

HARDWARE REPORT
- OS:
- Chip / CPU:
- Total RAM (GB):
- Graphics card + VRAM (GB), or "shared memory" on Apple chips:
- Free disk space (GB):
- Memory budget (80% of RAM, or of VRAM on a PC with a graphics card):
- Max model size, (budget - 1 GB) / 1.5:

Or just do the math right here

How much RAM have you got? (Mac: Apple menu → About This Mac. Windows: Settings → System → About.)

AI budget (80%)
12.8 GB
Engine
1 GB
Max model
7.9 GB
Context (50%)
3.9 GB

Rough guide, not gospel. On a PC with a separate graphics card, use its VRAM instead of RAM if you want things fast. Close Chrome's 80 tabs while it runs. You know who you are.

✓

Checkpoint. You've got a Hardware Report with your max model size. Keep that chat open, because the next step builds on it.

03+

Still cooking

The rest is in the garage

I'm still writing these up. Check back soon, or follow along on Instagram while I break stuff.