The Basics

Three ideas that, once you understand them, explain almost every bad experience you've had with AI.

You don't need to be technical to use AI well. But you do need three simple mental pictures in your head. If you skip this page, the rest won't make much sense. Go slow here.

1. Tokens

An AI model doesn't read words the way you do. It reads little pieces called tokens. A token is usually a short chunk, about three quarters of an average English word. The word "cat" is one token. The word "running" might be two.

The way to think about tokens is as fuel. Every question you type and every answer the AI gives back uses up some fuel. The bigger the question, the more tokens it costs. The longer the answer, the more tokens it costs. If you get close to running out, the AI starts acting strange: forgetting earlier parts of the conversation, repeating itself, or cutting off mid-sentence.

There's one more thing tokens explain, and it surprises people. Because the AI reads chunks, not individual letters, it's bad at anything letter-by-letter. Ask it "how many R's are in strawberry?" and it may get it wrong. Ask it to count the words in a paragraph and it may miscount. This isn't a sign it's broken. It simply never sees the letters the way you do.

2. Context windows

A context window is how much the AI can hold in its head at one time during a single conversation. Everything you've typed, every answer it's given, every file you've attached, it all lives inside this window.

When the window fills up, the AI doesn't crash. It quietly starts forgetting the oldest parts of the conversation to make room for new stuff. This is why, after a long chat, the AI seems to ignore something you told it at the beginning. It didn't ignore you. It forgot.

Different tools have different-size windows, and the exact numbers keep changing, so treat these as rough and "as of writing":

• ChatGPT varies by plan, smaller on the free tier, larger once you pay. A paid chat can hold roughly a long report's worth of text at once.
• Claude can hold something like 150,000 words in one conversation, about the length of a long book.
• Gemini tends to hold the most, on the order of several books' worth.

Don't memorize the figures. The point is only this: the window is big, but it is not infinite, and every tool has one.

Two habits come straight out of this whiteboard picture, and they'll save you a lot of frustration:

• Clear the chalkboard when you switch tasks. If you've spent twenty minutes with the AI planning a launch and now you want to write a birthday message, start a fresh chat. Leftover clutter from the old task quietly drags down the new answer. A clean whiteboard gives a cleaner reply.
• Put the facts in front of it, don't ask it to remember. If you want the AI to summarize a document, don't ask "what did that article say?" and hope it recalls. Paste the actual text in, or upload the file, so it's sitting right there on the whiteboard. Accuracy jumps the moment the real words are in front of it instead of being dredged up from a hazy memory.

3. The model has no memory: the whole chat is re-sent every turn

This is the fact that makes the context window finally click, and almost nobody is told it.

The model does not sit there remembering your conversation between messages. It remembers nothing. Each time you hit enter, the tool quietly gathers up the entire conversation so far, your first message, its first reply, your second message, its second reply, all the way down to what you just typed, and hands that whole pile to the model as if for the first time. The model reads all of it, writes one reply, and then forgets it ever happened. Next turn, the pile is gathered up and handed over again, now one exchange longer.

Nothing carries over between turns except the transcript itself. The model is like someone with no short-term memory who is handed the full notepad before every single answer. It reads the whole notepad, speaks, and forgets. The only reason the conversation feels continuous is that the notepad keeps growing and keeps being re-read.

Three things you've probably noticed all fall out of this one fact:

• Long chats get slower and more expensive. Every turn re-sends everything said so far, so a chat that's gone on for an hour is handing the model a small book on each reply. More to re-read means more time and more cost.
• The "forgetting" isn't the model ignoring you. When a chat gets long enough that the pile won't fit in the context window, the oldest messages get trimmed off the top before the model ever sees them. It didn't skim past your early instruction; that part of the notepad was cut to make the pile fit.
• Starting a fresh chat is a real reset, not a superstition. A new chat is an empty notepad. Nothing from the old conversation comes with it, because nothing ever lived anywhere but in that transcript. This is exactly why a clean chat so often fixes a conversation that had gone strange.

4. Grounding

This is the most important idea on the page.

Grounding means pinning the AI to specific, real information you give it, instead of letting it guess based on the whole internet.

Consider two ways of asking the same thing:

"Write me a welcome email for new customers."

That's ungrounded. The AI will write some generic welcome email using patterns it has seen a million times. The result will feel bland, corporate, and unlike you.

"Here is my brand voice doc, my last five welcome emails, and my list of services. Write a new welcome email in the same tone, mentioning my free intro call."

That's grounded. The AI is pinned to your real materials. The answer will sound like you, not like a stock template.

Almost every complaint about AI ("it's so generic," "it doesn't sound like me," "it just makes things up") is really a grounding problem. The AI wasn't told enough. It was guessing.

5. What it knows, and what it's only guessing

Here's a mental picture that ties the whole page together. Think of the AI as a lossy zip file of the internet. It read an enormous slice of the web and squeezed it down into patterns, the vibes and statistics of how words tend to go together. It did not keep exact copies of anything.

That one fact predicts what it's good and bad at:

• Things discussed endlessly online, it nails. How to structure a polite email, how a booking confirmation usually reads, what a friendly Instagram caption sounds like: there are millions of examples in the squeeze, so the patterns are strong.
• Rare, niche, or private facts are hazy. The specifics of your project, a figure from your own spreadsheet, a rule that applies only in your town: there was little or nothing about these in the pile, so it half-remembers or invents. This is exactly why you ground it with your real files instead of trusting its memory. Your own numbers have to be handed to it, not guessed at.

6. Lean on its tools

The AI has two weak spots you already met on this page: it can't do exact letter-and-number work reliably, and its built-in knowledge is frozen at the moment it was trained, so it doesn't automatically know today's news, prices, or events. Both have a simple fix, and it's the same move each time: tell it to use a tool.

• For anything with numbers, counting, sorting, math, say "use code." It writes a tiny program that computes the exact answer instead of guessing. This is the reliable way to get it to count, add, or crunch a spreadsheet.
• For anything recent or local, today's news, current prices, a business's hours, say "search the web." It goes and looks instead of leaning on stale, half-remembered training. When facts matter and they might have changed, always nudge it to look them up.

Why this matters

If you take one thing from this page, take this:

The rest of this guide is about fixing the setup. You'll learn how to load your real information into both Claude and Gemini once, so you never have to re-explain yourself, and how to ask questions in a way that gets grounded, specific, useful answers.