Whenever I watch people trying to get better results out of AI, the conversation almost always starts in the same place: the prompt. Write it better. Add more instructions. Include more context. Switch models and try again.

For a long time, I did the same. When the output was wrong, I blamed the model. When a new model made the same mistakes, I blamed my prompts. When better prompts didn’t cure it, I blamed the context. Every fix helped a little, and none of them held.

The failures kept coming back wearing different clothes. A run would start well, then halfway through the rules would fade. The style would drift. The AI would invent things that don’t exist in my world, and fill gaps confidently instead of asking what belongs there.

What bothered me wasn’t that it failed. It was that the failures always looked strangely familiar, no matter what I was building. At some point I stopped asking how to write a better prompt and started asking why the same collapse kept repeating.

It finally clicked when I noticed what I was actually asking the model to do. Every prompt was really two jobs pretending to be one. The first job: understand an unfamiliar world. The second job: solve a task inside that world.

Those sound similar, but they’re not. Imagine asking someone to renovate a house they’ve never seen, using materials they’ve never heard of, and expecting perfect decisions on the first day. Even a skilled professional would walk the rooms first. We skip that step every day and call it prompting.

I didn’t arrive at this as a philosophy. Reality kept forcing me into it, one project after another.

The first time was a mythology encyclopedia I built. The goal sounded simple: generate encyclopedia articles backed by real sources, connected through consistent taxonomies. I expected the AI to summarize existing knowledge. Instead it confidently filled gaps that should never have been filled. It invented facts, and then it invented the sources for them.

It wasn’t malicious, and it wasn’t stupid. It simply had no stable world to work inside. Mythology looks soft from the outside, but an encyclopedia is an unforgiving format. Every creature touches other creatures, cultures, symbols, and eras. Invent one detail and the contradictions spread through the whole web.

So before any article was written, I had to build the world itself. A single creature became a defined entity with its cultures, abilities, weaknesses, and relationships. Sources went into a library, each with its own level of trust. Only inside that structure did the writing stop drifting into fiction.

A year later I hit the same wall again, building AI tools for design. Every designer knows the moment: you look at a generated screen and something is off. Not because the layout is terrible. Because the AI is designing for a product that doesn’t exist. Plausible buttons that aren’t your buttons. Colors that feel right and match nothing you own.

So I stopped asking it to design and first taught it the design system. What its real components are and what role each one plays. Which color tokens exist. Which text styles are actually in use. All of it mapped before any prompt was written.

The difference was measurable. In production runs with roughly 190 component families available, every element the AI placed came from the real system, and every color came from a real token. The numbers matter less than what they showed me: when the model understands the world first, it stops spending its effort inventing one.

Words, mythological creatures, pixels. The material changed and the pattern didn’t. I suspect you could substitute your own domain and recognize it immediately.

That’s why I’ve stopped thinking of the prompt as the beginning of an AI workflow. Prompts matter, models matter, instructions matter, but they all come after something else. I’ve started calling that something else Step Zero: before asking AI to create, define the world it is allowed to create inside.

It’s worth being precise about what that means, because it’s easy to hear “give the model more context” and think you’re already doing it.

Pasting background into a prompt is not Step Zero. That’s advice shouted in the hallway, rewritten slightly differently every time, half remembered by the next conversation. A world is different. It’s built once, outside any single prompt. It’s structured, so the model can navigate it instead of skimming it. And it’s stable, so today’s run and next month’s run stand on the same ground.

Step Zero is also not retrieval for its own sake, and not memory. Fetching facts on demand helps the model answer questions. It doesn’t tell the model what is allowed to exist. The heart of Step Zero is boundaries: this is the vocabulary, these are the real objects, these are the trusted sources, and nothing outside them counts.

In practice, I’ve found the work comes down to four questions, asked before writing any prompt:

• What does the model need to know about my world that it cannot possibly guess?
• What are the real objects in that world: the components, the sources, the terms, the rules?
• What is the model allowed to invent, and what must it never invent?
• Where does this world live, so I build it once instead of retyping it into every prompt?

If you’re not sure whether this is your problem, the symptoms are consistent. The AI invents entities that don’t exist in your system. Two runs of the same request produce wildly different worlds. You keep pasting the same background paragraph into every conversation. Results are good while a chat is fresh and decay as it grows. None of those are prompt problems. They’re world problems, and a sharper prompt won’t fix them.

And starting costs less than people assume. You don’t need agents, pipelines, or any special infrastructure to try Step Zero. Pick one AI task you run repeatedly. Write one structured document that defines its world: the real objects, the terms, the rules, the sources you trust, and the things the model keeps inventing that it must not.

Then make every run begin from that document instead of from a blank prompt. It’s a modest step. In my experience it changes the output more than any prompting technique you’ll learn the same week, and it tells you quickly whether your problem was ever the prompt at all.

I don’t know whether this idea applies everywhere. Maybe it doesn’t. But I’ve now met it in enough unrelated corners of my own work that I can’t treat it as coincidence anymore. The more I build, the less interested I become in teaching AI to improvise, and the more interested in preparing the ground it stands on.

Near the end of building TrueUI, my current product, I realized I was never really trying to make prompts smarter. I was trying to make the starting point smarter. That shift changed how I think about AI products more than any prompt I’ve ever written.

There’s a second half to this idea that I’ve deliberately left out. Building the world is only the beginning. A world the model can ignore is only a suggestion, and getting a model to respect a world is a different problem than describing one. That’s the next thing I want to write about.

Until then, I’m curious. When you look at your own AI workflow, where does it actually begin?