You’ve heard the complaints. AI just makes stuff up, it lies, it hallucinates. It’s confidently wrong about everything, and you can’t trust it.
Those people aren’t crazy. They’ve been burned. They asked ChatGPT something reasonable and got back confident garbage. Wrong facts, generic advice, the kind of stuff that sounds smart until you actually try to use it and realize it’s just word soup with a confidence problem.
So they wrote it off. Overhyped, useless, a toy for tech bros who think slapping “AI-powered” on something makes it worth ten billion dollars.
Fair. But they’re missing something.
Every bad answer has the same root cause. The AI didn’t have what it needed to help. Not because it’s stupid, not because it’s lying, but because nobody showed it enough. You gave it a sentence and expected a miracle. The AI gave you the average of everything it’s ever seen, which, if you’ve spent any time on the internet, explains a lot.
That’s not an insult. That’s the fix.
The Mental Model Problem
Most people treat LLMs like they know everything. Like they’ve got access to your project, your history, your constraints. Omniscient helpers just waiting for the right question.
They don’t. Not even close.
LLMs don’t “know” things the way you do. They recognize patterns from training data. Ask about the capital of France and you’ll get the right answer, because that association has been hammered in millions of times. But ask about your specific project, your audience, your weird edge case? The AI has never seen any of that. It’s flying completely blind, and it’s too polite to tell you.
The only way to show it your world is through the context window.
What That Actually Means
The context window is everything the AI can see when it generates a response. Your current message, the conversation history, any system prompt running in the background, any files you’ve attached. Add it all up, and that’s the box the AI is working inside. Nothing else exists.
No memory of last week’s conversation. No awareness of your codebase unless you showed it. No knowledge of the Slack thread that prompted your question, or the five browser tabs you have open, or the meeting you just walked out of. Every conversation starts completely fresh. A savant with amnesia.
The AI has seen billions of patterns during training. But it’s never seen you until you show up in that window.
Why You Get Bad Answers
When you get a bad answer, the AI isn’t broken. It’s doing exactly what it’s supposed to do, generating the most plausible response based on what you showed it.
The output reflects the input. Always.
Vague input gets you generic output. Specific input gets you specific output. The AI isn’t lying to you, it’s working with what you gave it. And if you gave it almost nothing, you got almost nothing back. Garbage in, garbage out.
If you’re mad at the output, you’re mad at a mirror.
See For Yourself
Open a new chat and type:
“Help me write a function.”
You’ll get something useless. A generic function that does something vague, probably adds two numbers or prints hello world. The AI doesn’t know what language you’re using, what problem you’re solving, or what your codebase looks like. So it gives you the average of every “write a function” request it’s ever seen. Which is basically a tutorial example from 2019.
Now try this:
“I’m working in TypeScript. I need a function that takes an array of user objects and returns only the users who haven’t logged in for 30+ days. Each user has a lastLogin timestamp. I want to handle the case where lastLogin might be null.”
Same AI. Same moment. Completely different output. One is a tutorial. One is a solution.
The only difference is what was in the context window.
The One Thing You Control
You can’t control what patterns the AI learned during training. You can’t control the model’s weights, its architecture, or whatever the company decided to fine-tune for. That’s all locked in.
But you control what goes into that window. That’s the only thing that matters.
The magic prompts, the temperature settings, the model version, the secret techniques some guy on Twitter is selling for $49? None of that matters if the window is empty.
That’s the whole game. Fill the window or get garbage back.
Next: How to build a context file that gives the AI what it’s missing about you.