What the model ACTUALLY sees.

The model has never seen the chat interface. No buttons, no bubbles, no you. Every call, it receives one flat blob of assembled text — that's its entire universe.

The one idea

The model never sees the UI, your app, or a persistent "assistant" — on every call it sees only a flat blob of text the harness assembled: system prompt + tool schemas + retrieved data + conversation history + memory + the latest input.

Here's something that dissolves a surprising amount of confusion. The model never sees the pretty interface, your app, or a continuous assistant sitting across from you. On every single call, it sees one thing: a flat blob of text the harness assembled. Let's peel the interface back and look at what's really there. Strip away the interface, and here's what actually arrives at the model each turn. First, the system prompt — its rules, its persona, its tool instructions. Then the tool schemas — every available tool described, in words, eating their share of the window. Already, a lot of the context is spoken for before you've said anything. Then any retrieved documents — the RAG passages pulled in for this question. Then the conversation history, re-fed in full, because the model is stateless and remembers nothing on its own.

How it works — the demo

Peeling back the interface to reveal the actual context the model receives each turn — one assembled text blob, sent once to a stateless model, no UI or continuity behind it.

Then whatever memory was pulled from the store. Every layer is the harness deciding what the model gets to see this turn. And finally, at the end, the latest thing — your new message, or a tool's result. All of it flattened into one long stretch of text and handed to a stateless model, once. No app, no screen, no persistent assistant behind it. Just this assembled blob, read in a single pass, and then it's gone. That's the reframe. The assistant you talk to doesn't persist, doesn't see your screen, isn't really there between calls. Each turn, the harness rebuilds a text blob from scratch and a fresh, stateless model responds to it, then it's gone. The continuity — the sense of a someone — is an illusion the harness stitches together, turn by turn.

The trap to avoid

Imagining the model perceives your app/UI, remembers continuously, or "is" the assistant persona — it sees only the assembled text, freshly, each call.

Why it matters — and what’s next

The harness assembles context and tools — but you can extend it: skills, hooks, rules.

The trap: imagining the model perceives your app, remembers continuously, or is the assistant persona. It sees only the assembled text, freshly, each call. The persona and the memory are constructed by the harness — which is exactly why, all series, context and the harness have turned out to be everything. The model supplies intelligence; the harness supplies the rest. So there's no UI, no memory, no assistant behind the curtain — just text the harness assembled, sent to a stateless model, once. Understand that, and the whole act clicks into place. The harness assembles context and tools — but it isn't fixed. You can extend it, with reusable skills, control hooks, and standing rules. Next group.

This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.