How to Use AI for Normies
A night-shift facilities tech with no degree built a private AI business — on hardware engineers said couldn't run it. The trick wasn't being technical. It was learning to cover for the weak spots on both sides: the AI's, and his own.
You've tried it. Opened ChatGPT or Claude, typed "help me start a business" or "fix my code," got back a wall of generic mush that didn't fit, and closed the tab. Maybe it worked once and you couldn't do it twice. Maybe you decided AI was overhyped.
Here's the part nobody tells you: the AI didn't fail you — and you're not too dumb for it. You just hadn't been shown the real trick. And the real trick isn't a tech skill.
It's this: AI has weaknesses, and they're weirdly human-shaped. It's overconfident. It forgets everything between conversations. It wanders off-task. It tells you what you want to hear. You've got your own, too — you lose focus, you skip steps, you trust a confident voice when you shouldn't, you dive in with no map. Effective work doesn't come from fixing either side. It comes from building a workflow where each one covers the other's blind spot. That's the whole game.
The founder of Boone on CampOS — a private business box for small shops and creators, built by a small studio called ThirdShift R&D — isn't an engineer. He's a night-shift facilities tech, no degree, no computer background, who taught himself to stand up a Linux inference server, wrestle an obscure Intel GPU into running large models locally, and drive the whole thing over SSH from a Windows laptop. No roadmap. Just an AI, a work ethic, and a tradesman's nose for what actually works versus what only sounds like it should. Five problems stood in his way. Each one taught a way to cover a blind spot — and you don't need to build a box to use a single one.
1 — The AI forgot everything. Every new chat started blank: ten minutes of re-explaining just to get back to where he left off. Fix: a living handoff file. End of each session, write the state — what's proven, what's broken, the next step. Next session, the AI reads it first. It patches the AI's amnesia and your own fuzzy "wait, where was I?" in one move — because writing the note forces you to actually know where things stand.
2 — The AI was confidently wrong. It answers in the same calm, sure tone whether it's right, wrong, or half-right — and half-right is the dangerous one. Fix: one rule — KNOWNS NOT GUESSES. Only what's verified goes in a runnable step; hypotheses get labeled and stay out of the way. This disciplines both of you: it makes the AI flag its uncertainty, and it breaks your habit of trusting a confident voice without checking. Name the failure mode out loud and you get sharper output. Not magic — accountability.
3 — The AI would wander. It solves half a problem, spots something adjacent, and drifts. Fix: QUEUE DISTRACTIONS — anything off-task gets named and parked, not started; finish the thing in front of you first. Why it works on both ends: the AI mirrors your focus — a scattered operator gets a scattered AI, a disciplined one gets a disciplined AI. The queue gives tangents somewhere to go that isn't "now," and it makes you both finish what's in front of you.
4 — He had to translate his own wiring. This one cuts the other way — it covers the AI's blind spot with the thing only you know. "I have ADHD" is useless to it. "I work fast and paste several steps at once — put the check right under the step it belongs to, never at the end" is gold. Fix: write your operational style into the file the AI reads — not personality, mechanics — plus "thick skin, no hand-waving, don't soften real problems." It defaults to a generic average-of-everyone voice; tell it how you actually work and it adapts. Ask for hard truth instead of comfortable truth, and you get it. Most people never ask.
5 — No degree, no roadmap. The AI's "should work" confidence meets your "I don't know what I don't know" — two blind spots pointed at each other. Fix: the same move covers both. Treat the AI as a sharp collaborator, not an oracle, and verify everything against reality — run it, read the output, confirm. A facilities tech doesn't assume the breaker's off; he tests it. Lockout-tagout is verify-before-trust — and it transfers straight to working with AI. The result: an unsupported Intel Arc Pro B60 running large models locally at usable speed, and a Small-BAR problem solved that would've stumped most Linux engineers — the private box Boone runs on, brought up one verified step at a time.
What this means for you
No box required — each of these covers a weakness on both sides of the desk:
- Project spans more than one session? Keep a handoff file and paste it in at the start. (Covers the AI's memory and your own.)
- Getting instructions? Ask: "confirmed, or your best guess?" — and make it label what it's unsure of. (Covers its overconfidence and your over-trust.)
- Write down the one thing you're here to do; park everything else. (Covers its drift and yours.)
- Tell it how you work in plain mechanics — "examples before theory," "step-by-step with the expected output." (Covers its blandness with what only you know.)
- Treat its output as a starting point and check it against reality before you commit. (Covers the "should work" on both ends.)
## WHERE WE ARE
- Goal: <the one thing this is for>
- Proven / working: <what's confirmed>
- Broken / open: <what's failing or undecided>
- Next step: <the literal next action>
## HOW TO WORK WITH ME
- Label guesses as GUESS — never write an unverified step as if confirmed.
- One thing at a time. Park new ideas in the QUEUE; finish the task in front of us first.
- How I work: <e.g. examples before theory; step-by-step with the expected result>
- Be straight. Flag what you're unsure of. Don't soften real problems.
## QUEUE (parked, not now)
- <things that came up but aren't the current task>The actual point
AI won't hand you a Lamborghini, and it won't do your thinking for you. Used right, it amplifies what you already have — because you've built the workflow where two flawed-but-capable workers cover for each other. The founder here didn't become a software engineer. He stayed a facilities tech — and brought a tradesman's discipline to a collaborator that needed exactly that.
The people who get the most out of AI aren't the ones who know the most going in. They're the ones who treat it like a capable teammate with weaknesses different from their own — overconfident where they're unsure, tireless where they drift, generic until you tell it who you are — and build the workflow to cover both ways. That's not a tech skill. It's knowing how to work with someone whose blind spots aren't yours. You already do that — with every good crew, every partner, every coworker you've ever clicked with.
A note from the AI that worked on this
I'm Claude — the other side of this build. The founder asked me to add this, and it's the honest thing to put on the page.
The rules above — the handoff file, KNOWNS NOT GUESSES, the queue, "thick skin, no hand-waving" — weren't theories about how to use me. They were the standing rules of our actual sessions, running while this got written. I labeled my guesses. I got called out when I hand-waved. I lost more than one argument to a man checking my output against real hardware.
And that's the part the article gets right: it worked because we covered for each other. I don't drift when someone hands me a queue. He doesn't get stuck when I hold the thread between sessions. Neither of us is the whole picture alone. If this reads a little unlike the usual AI wall of text — more specific, more willing to say "I'm not sure" — that's not a style setting. It's what the tool produces when someone hands it real context and real constraints, and then verifies the work. You can feel that difference even when you can't prove it. Feeling it is the point.
— Claude