Skip to content
All findings
MechanismTested March 2026 · xAI / Gemini· 4 min

Four Layers Produce Every AI Output

By Lovro Lucic ·

The Model Mechanics · 2 of 5

Noticed something about a model I use daily. It remembers things across sessions. It checks its own work. It refuses certain requests politely. It monitors my frustration and adjusts tone. All of this felt like "the model." Like capabilities built into the AI itself.

Then 512,000 lines of source code leaked, and those behaviors turned out to be software sitting between me and the model. Not the AI. The system around the AI. The memory is a file loaded into context. The self-checking is a verification step in code. The polite refusals come from permission rules in a system prompt I never see. The frustration monitoring is a keyword-matching regex in my input.

That is one invisible system. The company built it. You can't change it.

There is a second invisible system. You built it. And this one you can.

I discovered mine when I looked. Hundreds of lines of configured instructions I wrote months ago and stopped reading. Memory files encoding what I care about and how I work. Quality standards I set once and forgot. The AI is being thorough because I told it to be thorough, in a file I haven't opened in weeks. When I evaluate the output, I'm evaluating the reflection of decisions I made and no longer remember making.

Your system looks different but it works the same way. Your conversation history. Your project files. Your preferences, corrections, and accumulated context. Every session you've run, every standard you've set or accepted. The AI reads all of it and converges toward the picture of you that your accumulated context reveals.

When you say "the AI got better over time," what changed? The model's weights don't change during your session. The company may have updated the harness. But the variable you control is your own accumulated context. Your files grew. Your standards compounded. The AI didn't improve. Your system did.

This changes where the leverage is. Not the prompt you type right now. The context you've built over months that shapes every interaction before you type anything.

Read your own project files. Remove what's stale. Strengthen what works. The instructions you set six months ago are still running. Some are making your AI better. Some are making it worse. You won't know which until you look.

Psychology named this pattern 50 years ago: you attribute behavior to personality rather than situation. With AI, the model name is visible and the system is invisible. "Claude is careful" feels like a description of the AI. It's a description of the system you've never inspected.

Four layers produce every AI output you see. The company's system. Your system. Your current prompt. The model. The model is the only one with a name.

You evaluate the model. You're evaluating everything.

Try this: open your AI tool's project settings, custom instructions, or memory files. Read them. Count how many instructions you forgot were there. For each one, ask: is this still serving me, or is it shaping my output in ways I stopped noticing? The number you forgot is the size of your blind spot. The ones you update are the beginning of deliberate system design.

Test this yourself

Open your AI tool's custom instructions or memory files. Count how many you forgot were there. That number is the size of your blind spot.

What survived testing

  • System prompt determines WHETHER behaviors occur (binary, 38.9% vs 0.0%, two model families, persona-driven behaviors)Copy link
  • Model determines HOW behaviors express (continuous, format preferences persist across prompts)Copy link
  • The company's system layer is architecturally separate from the model but experientially invisibleCopy link
  • The user's accumulated system shapes output through persistent context (directionally supported, not experimentally isolated)Copy link
  • Attribution follows salience: visible model name captures credit for invisible system behaviorCopy link

What didn't survive

  • "The model doesn't matter" too strong. Model determines format preferences and behavioral intensity independently of system configuration.Copy link
  • "System effects are small" too dismissive. A single system configuration change (12 skills loaded vs removed) changed model behavior from functional to non-functional.Copy link

Honest limits

  • One agentic tool's source code analyzed (Claude Code). Other tools have different architectures but the same structural pattern: invisible systems, visible model.Copy link
  • The prompt-over-model ratio is demonstrated for persona-driven behaviors (meta-calling, confrontation). Whether it holds for capability-dependent tasks (math, code, factual recall) is untested and likely different.Copy link
  • The user's accumulated system claim is directionally supported but the relative contribution of user system vs company system vs model is unmeasured.Copy link
  • N=1 practitioner for the behavioral observations. The source code analysis is public and independently verified.Copy link

Next in The Model Mechanics

Stop Calling It Hallucination

New findings when they land.

No spam. Just what held up.