How I Use AI

This recommendation was written in April 2026. AI models are constantly evolving, so the information here may be outdated by the time you read it.

One of the hardest things about AI is knowing how to use it. Unless you're chronically online (like me), it's easy to get confused about which models are the best and most up-to-date. So I thought I'd give you insight into how I use AI and what I use it for.

This is a guide I plan to update every time I change my workflow when I discover new models or use cases. If you find any of this useful or think something could be improved, feel free to get in touch! I want to get better.

Don't use the free version!

The single most important piece of advice I give to people who want to get to know AI is to not use the free versions. This is a point I really need to hammer home because most people use whatever is free.

The main problem with using the free version is that you get access to the companies' worst models. The models have less computing power, hallucinate more, and are generally slower. Free models also often lack so-called "thinking functionality," which allows models to search the web and "think" through the question before answering. Instead, a free model spits out an answer immediately, limited to whatever training data the model had when it was trained. The information is often outdated and wrong.

A more philosophical problem with using the free models is that you get a misleading picture of how powerful AI has actually become. Free models lag 6-12 months behind the best models — an eternity in AI development. If free models are your only frame of reference, you'll underestimate how much AI will change society and, not least, your job. You end up with a worse mental model of the world.

I know it's hard for people to pay for yet another subscription. But to this I would respond that the price of intelligence is ridiculously low. For 250 NOK you get access to most of the top models from the major AI companies. That's the same as two beers at a bar in Oslo.

I personally pay $100/month for Claude Max every month. I'm aware that this is a lot, but I get tremendous value out of it. However, for most people a regular plan at 250 NOK is enough, and you can always see if you want to upgrade later.

Which AI should you use?

The most important choice you need to make is figuring out which AI model will be your go-to model. By that I mean the model you pay for and the one you use for most tasks. Here you have different options with pros and cons. Here are some of the most common alternatives (you can skip this section if you already have a go-to model).

Claude by Anthropic

Strengths:

• Extremely good at coding. Claude Code is considered to be at the forefront of the field.
• In my opinion, Claude is the best at creativity, writing feedback, and humor. It's hard to describe, but Claude simply has better "taste" than the others.
• Pleasant and warm personality. Gives the right amount of pushback and friction. Good humor too.
• Anthropic is the AI company that takes AI safety most seriously and spends the most time warning about existential threats from AI, although the company has recently deprioritized safety due to competitive pressures.

Weaknesses:

• No image or video generation.
• With the 250 NOK version, it sometimes uses up my "tokens" quickly, which is annoying.

ChatGPT by OpenAI

Strengths:

• Good at coding. Some consider Codex 5.3 to be the best model as of February 2026.
• Has a pretty good image and video generation service called Sora, but not as good as Google's Nano Banana.
• Very good at research and problem-solving. The Deep Research feature in particular is excellent.

Weaknesses:

• Somewhat cold and unfriendly personality. But also very sycophantic.
• Has a clunkier user interface than Claude in my opinion. Lacks a project feature, for example, which is impractical.
• OpenAI is led by Sam Altman who does not inspire trust.

Gemini by Google

Strengths:

• Gemini leads in image and video generation with Nano Banana and Veo 3. If you haven't checked them out yet, you should do it now.
• Directly integrated into the Google ecosystem if you use it.
• Gives access to NotebookLM, which is a very practical tool for studying and summarizing large amounts of text. The podcast feature is very good.
• Generally a very capable model.

Weaknesses:

• Not the best model for coding, but not far behind either.
• I haven't properly tested the model myself and don't know much about its weaknesses.

Grok by xAI

Strengths

• If you need a model that is willing to push boundaries, Grok may be the model for you. Some AI researchers use Grok for "red-teaming" research because the model is willing to break ethical guidelines that others follow.
• You get access to Grok if you buy premium access on Grok for $8, which is cheaper than the others. Not sure how good the model is at that tier though.

Weaknesses

• X's lack of boundaries means that Grok has been used for a lot of harmful things, such as undressing women against their will, sexualizing children, and bullying people.
• Grok has gone off the rails politically several times, for example when it spread false news about white genocide in South Africa or declared itself Mechahitler. Both incidents were caused by Grok's creator Elon Musk.
• Grok also scores lower than Claude, Gemini, and ChatGPT on most benchmarks.

1 of 4

How to prompt smartly?

You may have heard that you should assign the models a role, for example "you are the world's best editor who has read thousands of articles and works at Penguin Publishing. You receive many manuscripts every week and are used to reading through them thoroughly to look for new voices, etc. etc." to get the best answers. You don't need to do this anymore. The newest models are smart enough to know what you want and are aiming for without you needing to make them role-play.

It's still smart to prompt well, though. An important principle with AI is that the quality of the answers you get depends on the quality you put into your prompts. Garbage in, garbage out. The more context and details about what you want to achieve (and what you don't want), the better answers the model gives you.

The recipe I use when starting a project or a larger task is to include the following in my opening prompt:

1. Describe the project/task you're working on and what you want to achieve.
2. Describe the problem you're struggling with.
3. Explain what you want it to do (if you know).
4. Things you want it not to do or be aware of.
5. Always remember to ask if it has any questions or if anything is unclear.

Notice how much background and information I provide in the example above where I'm starting a larger project. That's because I want to give the model as much context as possible about what my goal is and what resources I plan to use, so it can give me good feedback on whether it's realistic or not.

I also ask for help refining the project and what Claude thinks. That's because Claude has access to an enormous amount of knowledge and information that I don't have, but which it won't give you unless you explicitly ask. On many tasks I have a vision and direction, but need help with the details. That's where Claude is your friend.

Once you're in the middle of a conversation you don't need to provide as much information in every prompt. It depends entirely on what Claude asks about and what information you need to provide.

I belong to the school of thought that believes there's value in being polite to the model and saying "thanks" or "great job" when you think Claude does well. There are two considerations here. First, you're giving Claude feedback about what works well and what doesn't. This helps it better adapt to your preferences. The second consideration (which is more abstract) is that I think it's wise to be polite to the models so that the training data they receive maintains a pleasant and good tone.

What can you use AI for?

Okay. You've now hopefully chosen a model, learned how to prompt, and that you shouldn't use the free models. The question now is: what should you actually use AI for? Everyone talks about how you need to start using it.

I personally use AI for the following tasks.

• Reading through documents, bank statements, and contracts.
• Composing emails and helping you reply.