Every programming journey begins the same way. You start with “Hello World.” You start there to understand what the system is doing. Until that point, you are copying symbols, not learning a program.
AI is no different. It is not magic. It is not an assistant in the human sense. AI is a program you must learn to operate.
This matters because I am not a software engineer. AI can already do far more than I can verify. That asymmetry is dangerous if ignored. The Hello World Mandate exists to deal with this reality.
The Mandate
Before scaling anything with AI, answer one question clearly: can I tell if AI is wrong? If the answer is no, scale becomes reckless.
The rule is simple: do not ask AI to do things you cannot audit. Hello World is the smallest unit where learning, verification, and authority still align. That is why it comes first.
The Hidden Advantage
O9X was not my first software project. Long before AI, I built my company website manually.
The first version ran on Docusaurus. It took two weeks to deploy. When my ambitions grew, I rebuilt the same site in Next.js. I relied on Google Search and Stack Overflow, not agents.
None of this was efficient. But because the website could break without business consequences, I could learn without fear. That manual friction was my training ground.
The Exploration Phase Where Failure Was Cheap
Once AI could build usable code, I spent weeks testing systems. I needed to understand which one I could trust, how to manage context windows, and which stayed predictable when constrained.
Claude proved the most reliable for coding. Because Claude has usage limits, it forced discipline. I could not brute-force a solution. I had to plan.
The Multi-Model Setup
By the time I got serious about O9X, I had assigned a specific role to each model. Gemini is the Researcher — I use its long context window to digest documents and generate reports. ChatGPT is the Auditor — it critiques Gemini’s outputs and pushes the analysis further. Claude is the Builder — it handles the final coding and execution. Kimi is the Editor — it refines final drafts, as its writing style reads more naturally.
Despite the multiple models, final judgment stays with me.
Managing Context Windows
Long AI chats rot. Decisions decay, constraints dissolve, and rejected ideas quietly reappear. I knew I could not rely on conversation history to carry the project.
I adopted effective context engineering: a claude.md file defines the rules, stack decisions, and coding discipline, and a project_status.md file tracks what is done, blocked, or pending. These files became infrastructure. Gemini structures the work. ChatGPT critiques the plan. Claude executes the code. No model holds the project in its head. This stops AI from getting creative with my stack or writing code that will not run on Cloudflare.
The Auth Decision and the First Hard Boundary
I knew early that authentication was a danger zone. Not because it is difficult, but because it is difficult to implement correctly.
So I chose a backend — PocketBase — that shipped with a working, opinionated auth system. I used it as-is.
The First Vertical Slice That Shipped
I built CRUD for the first two collections: Users and Customers. I added one end-to-end flow proving that state changed correctly across the system.
There were no roles, no permissions matrix, and no hardening beyond defaults. The goal was to prove that the data spine worked, that AI operated within guardrails, and that I could audit every meaningful decision it made.
The Supervisor Loop
Once the slice worked, I stopped and reviewed. I took the code and handed it to Gemini in a fresh session with no context.
I asked it to explain the code back to me. This surfaced unclear logic and fragile decisions. I fed that feedback back into Claude, refined the code, and repeated the loop. Only when Gemini found no errors did I ask ChatGPT for a final audit. The code was accepted only if no further errors were identified. This loop — not any tool — is where control is earned.
What “Hello World” Actually Means
The Hello World Mandate is not about building something small. It is about proving you can maintain authority while learning a powerful new program.
AI can outpace your understanding quickly. If scope grows faster than comprehension, you do not become advanced. You become blind.
Hello World is the moment you choose to grow understanding before power. That is not theater. It is how competent programmers actually learn.
AI is not a shortcut past learning. It is the thing you must learn. Like every powerful system, you start small not because that is all it can do, but because that is all you can verify. Skip Hello World and you do not scale faster. You scale confusion.