A guide

Which AI app should you use? Pick one and stay with it.

Five names come up in every conversation about AI: Claude, ChatGPT, Gemini, Perplexity, NotebookLM. They are not five versions of the same thing, but they are not five separate tools to learn either. This guide sorts them into three groups, makes one recommendation, and walks through setting it up so it starts working for you specifically.

The whole path, in one view

1Know the mapThree groups
2Pick oneAnd commit
3Teach it about youOnce
4Use it dailyTwo weeks
5Add specialistsWhen needed

No accounts to compare, no feature tables. One decision, made once, and a setup that takes an evening.

Know the map

The apps sort into three groups, and the groups matter more than the brand names.

General assistants: Claude, ChatGPT, Gemini. The same kind of tool. You talk to it, it drafts, explains, summarizes, plans, and works through problems with you. The differences between them are real but smaller than the marketing suggests, and they trade places with every release.
Search: Perplexity. Built to look things up. It runs the search, reads the results, and gives you an answer with sources. Closer to Google than to the group above.
Your documents: NotebookLM. You feed it your own files and it answers only from those. That constraint is the feature. For study materials built from your own PDFs, it is the best of the group, because it will not pull in references you did not give it.

They are not just chat windows anymore

Claude and ChatGPT both have modes that work on your actual computer: organizing files, drafting documents in place, building small tools for you. I have used Claude's version, called Cowork, to build things I had no business building with zero coding background, and most of the guides on this site run on it. This is where the leverage is, and it is the part nobody tells beginners about.

Pick one general assistant

The instinct is to try all three and keep rotating. I would not. Over time the app you use accumulates your context: memory of past conversations, your preferences, the way you like things done, the projects you keep coming back to. None of that transfers when you move. Six months in, the value is not the model, it is that it knows you. Rotating between three apps means none of them ever does.

Which one? I use Claude. The honest reasons: the writing quality suits me, and Cowork is the mode I lean on for everything beyond chat. The guides on this site are built on it, so if you follow them, Claude is the path of least friction. But ChatGPT does everything Claude does, and going that route is just as reasonable, especially if the people around you already use it and can help when you get stuck.

One more case points elsewhere. If your work life already runs on Google, Gmail, Docs, Calendar, then Gemini is built into all of it and that convenience is real. The wrong answer is not any of the three. The wrong answer is all of them at once.

What it costs

Each of the three has a free tier, and it is enough to test with. The paid tiers run about $20 a month, and the desktop modes from the last step generally need one.
Upgrade when the app is doing daily work for you. But paying is not a privacy setting, the consumer plans handle your data the same either way. The setting that matters is in the next step.

Teach it about you

Out of the box, every answer is written for a generic stranger. Ten minutes of setup fixes that, and it is the single highest-return step in this guide.

Turn off training on your conversations. On the consumer plans of all three, free and paid, your chats are used to train future models by default. Each app has one setting to switch it off: in Claude it is under Privacy, in ChatGPT under Data Controls, the toggle called "Improve the model for everyone," in Gemini it is the Apps Activity setting. Flip it before you type anything personal. Opting out applies going forward, it does not pull back what you already sent.
Fill in your preferences. Every app has a settings section for this, usually called personalization, preferences, or custom instructions. Write a short paragraph about who you are and how you want answers.
Turn memory on. Look for a memory setting and enable it. This is what lets the app carry context from one conversation to the next instead of starting from zero each time.
Make a project per recurring area. Most of these apps let you group chats into projects or spaces. One for work, one for a course you are taking, one for the house renovation. Context stays where it belongs.

A starting point for the preferences paragraph. Change the bracketed parts:

Preferences

I am a [your role]. Most of my questions will be about [the two or three areas you will actually use this for].

When you answer: be direct, skip the preamble. If a topic is in my field, assume I know the basics and go deeper. If I ask for writing, keep it plain, no hype.

Do not agree with me by default. If my reasoning is weak or my plan has a problem, say so.

When facts matter, tell me how confident you are and what you would verify. Never invent sources, citations, or numbers. "I do not know" is a good answer.

The last two paragraphs are there for the two failure modes worth knowing on day one. These apps agree with you by default, and they state wrong things with full confidence, including references that do not exist. Writing the instruction up front does not eliminate either, but it helps, and it sets the habit of asking "how sure are you" when the answer matters.

If you work in healthcare, one rule before anything else

Patient information does not go into a consumer AI app. No names, no record numbers, no dates of birth, no detail that could identify someone, even in a quick question. De-identification is a defined standard, not just removing a name, and your institution's policy governs what you can do. When in doubt, describe the general scenario instead of the actual patient. A full guide on this is coming.

Use it daily for two weeks

Picking the app is not the skill. Reaching for it is. For two weeks, bring it real tasks from your actual day, not test questions to see if it is smart.

A starting list, roughly in order of payoff:

A draft you have been putting off: an email, a recommendation letter, a difficult reply
A document you need to understand: a contract, a policy, a dense paper. Upload it and ask questions
A plan: a talk outline, a trip, a week of meals around what is in the fridge
Something you would normally ask a colleague at their desk: a quick sanity check, a second read of your reasoning

Correct it as you go. Tell it when an answer is too long, too padded, or wrong, and with memory on, the corrections stick. The test at the end of two weeks is simple: when a question comes up, is this the first place you take it? If yes, the setup worked.

Add the specialists when the need shows up

The two apps from the other groups are worth having, but not on day one. Add each when its trigger appears.

NotebookLM, when you are studying from your own materials. Board review, a set of guidelines, lecture PDFs. Load them in and it builds summaries, study guides, and answers grounded only in what you gave it.
Perplexity, when you would otherwise google. Current information, sources you need to cite, anything where "who says so" matters as much as the answer.

One at a time

Add a tool when a real task calls for it, not because a list said to. Two apps you use beat five you signed up for.

Why one beats three

Every month someone publishes a new ranking of these apps, and the order changes every time. The ranking that matters is different: the app that knows your work, your preferences, and your last three months of context will out-help a slightly smarter model that is meeting you for the first time. That compounds, and it only compounds in one place.

So pick one, set it up properly, and give it two honest weeks. The comparison shopping can wait, and I think by then you will not want to do it.

More guides as I build them. They go out through RadBrief.

Get the next guide →