Last verified: Sept 2, 2026
This is the foundation the rest of the site is built on. It takes about ten minutes.
If you’ve been using these tools for a while, skip to “The part that catches out experienced users.” That one is worth your time regardless of how long you’ve been at this. The glossary at the bottom is reference, not homework.
One idea, and everything else follows
ChatGPT, Claude, and Gemini are all the same kind of thing underneath: a system that has read an enormous amount of text and become extremely good at producing the next sensible piece of it.
That’s the trick. Not a database being queried. Not a person thinking. A pattern machine, operating at a scale that makes the output feel like something else.
Here’s the shorthand I keep coming back to. The tools have changed many times since I started; this has not needed to: it’s a brilliant, widely-read stranger who started five seconds ago, has no idea what your life looks like, and will never tell you when it’s guessing.
Three things follow from that, and they’re most of what matters.
1. It has no context, so context is your main lever
The stranger doesn’t know your kitchen, your budget, your calendar, or that Thursdays are impossible. Everything it doesn’t know, it fills in with the statistical average of everyone who ever wrote about your situation. That’s why generic questions get generic answers.
This is also why prompt tricks are oversold. “Act as an expert” and “think step by step” help at the edges. Telling it that you have forty minutes, three kids, and a nut allergy in the house changes the answer completely.
More context, better answer. That’s the whole of prompt engineering, and everything else is decoration.
2. It won’t flag its own uncertainty
A wrong answer is produced by exactly the same process as a right one, so it arrives in exactly the same confident tone. There’s no internal signal that gets weaker when it’s on shaky ground. This is what the industry calls hallucination, which is a gentler word than it deserves.
Where it happens most: names, dates, statistics, citations, legal and medical specifics, and details about products that changed recently.
The practical rule isn’t “verify everything,” which nobody does. It’s this: it’s strongest where you can evaluate the output yourself, and weakest where you’d have to take its word for it. You can tell immediately whether a rewritten paragraph reads better. You cannot tell whether a cited study exists.
3. It starts over every time, unless you build it somewhere to stand
Each new chat begins from nothing. That’s why re-explaining your household every session feels like work, because it is.
The fix is a saved workspace: Projects in ChatGPT and Claude, Notebooks in Gemini. Your standing instructions, your files, and your related chats in one place, loaded automatically. It’s the difference between briefing a new sitter every time and having one who’s been coming for a year.
If you do one thing after reading this, do that one. Here’s the ten-minute version.
The part that catches out experienced users
Here’s what a year of daily use won’t teach you: the settings multiply, and the opt-outs aren’t connected.
Most people who care about this read a privacy guide at some point, turned off the training toggle, and filed it as handled. That was reasonable when there was one toggle.
As of this year, ChatGPT alone has three separate controls governing what happens to your conversations, sitting under the same menu, two of them on sub-pages you have to click into. Turning off training does nothing to the other two. All three default to on. Google restructured theirs twice in a year and renamed the products while they were at it.
You do get told, technically. An email arrives saying the terms have been updated, linking to several thousand words of legal language with no indication of which sentence changed, whether it affects you, or whether there’s now a switch you’d want to flip. That’s notice in the legal sense and not in any useful sense.
So a setting you configured once and stopped thinking about becomes a setting you no longer understand, and you never got a moment where it was obvious that had happened.
Closing that gap is a large part of what this site is for. When something changes that actually affects how your information is handled or what you can do on a free account, I write up what changed, in plain terms, and what if anything you should do about it. The running log is here.
That’s the argument for checking twice a year rather than trusting your past self, and it’s the reason the privacy article on this site carries a “last verified” date. What to check, and what each control actually does.
Pick one tool and stop shopping
The three main tools are far more alike than different for the things this site covers, and the differences shift every few months, which makes comparison shopping an excellent way to feel productive without starting.
Use whichever you already have an account for. Give it a month. Thirty days with one tool teaches you more than thirty minutes with three.
What I use, since you’ll wonder
I started out an OpenAI devotee and ran almost everything through ChatGPT. These days most of my work happens in Claude, and I go back to ChatGPT for things that need more creative range or better visual output. I move to newer versions as they ship. So the specific tools behind anything I write here have changed many times and will change again.
I also pay for my accounts, and I’d rather say that plainly than have you find out later. A lot of people recommending AI tools are quietly working with a version you can’t get to.
That said, nearly everything I write about works on a free account. When it doesn’t, I’ll tell you which part needs a paid plan and whether I think it’s worth the money.
And the reason I use more than one tool is not that you should. It’s that after enough years of this I know where the seams are, which is knowledge you only get by using one thing long enough to notice what it’s bad at. That’s the same thing I’m suggesting you do first.
The words, if you want them
Reference, not homework. Skip if these are already familiar.
Prompt. What you type. “Prompt engineering” means describing your situation clearly.
Context. Everything the tool knows about your situation when it answers.
Model. The specific version doing the work. GPT-5, Claude Opus, Gemini 3. They change often and the names are not meaningful.
Chat or thread. One conversation. Start a new one when the subject changes.
Project or Notebook. A saved workspace holding instructions, files, and related chats.
Memory. Separate from Projects. The tool retaining facts about you across unrelated conversations. Usually on by default, and worth knowing where the switch is.
Hallucination. It made something up and stated it confidently.
Token. How these systems chop text into pieces. You will never need this. It’s here so it doesn’t unsettle you when someone says it.
Then go do something small
Not a life overhaul. One irritating, low-stakes task this week.
The reason home is a better place to learn this than work is that the feedback is immediate and nothing breaks if it goes badly. You’ll know within a day whether the dinner suggestions were any good.
That’s the method. Everything else on this site is a variation on it.
Subscribe and hit reply to tell me your getting-started story, or send me a note if you’d rather.
