Think Like an LLM

Before you can direct an AI agent well, you need to understand what it actually is — and what it isn't.

What an LLM is

A Large Language Model is a next-token predictor. Given everything it has seen so far — your message, the conversation history, any files context loaded at startup — it predicts the most likely next word. Then the next. Then the next.

There's no reasoning engine behind the scenes. No memory between sessions. No understanding of your project unless you put it there.

Context is everything

When people say "the AI got it wrong", what they usually mean is "the AI didn't have the right information". The model's quality is largely fixed. What changes is what it can see.

Everything the model sees when it generates a response is called the context window. This includes your current message, the conversation history, files you've opened or attached, any system instructions your tool has loaded, and outputs from tools it has already called.

The quality of the output is directly proportional to the quality of the context.

Why prompt tricks plateau

Rephrasing a prompt only gets you so far. "Be more careful" or "think step by step" help a little — but they don't change what the model knows about your codebase, your standards, or your customer's requirements.

Context is the lever. Prompt tricks are friction reduction. You can only reduce friction so far.

What this means for AL development

When you ask an agent to write AL code:

• If it can't see the base application source, it will guess at table structures and event publishers
• If it doesn't know your coding standards, it will invent its own
• If it doesn't know the customer brief, it will solve the wrong problem

The rest of this workshop is about giving the agent the right context — systematically, repeatably, without you pasting things by hand every session.

That's the whole game.