AI Best Practices
Last updated on: 01 April, 2026
-
For trivial tasks/questions, use free Fast/Instant models by any company. Don't bother about how to prompt. If you are not satisfied with the answer, simply mention in the next prompt what you didn't like in the answer or what more you want in the answer.
-
For anything that is meaningful, use the latest thinking models, not older or faster models. You need to pay around $ 20/month for that. If you are using an old, free model, no matter what prompt or best practice you use, you will never get an answer that matches the quality of an answer by the newer, thinking model. Using the best model is the best hack.
By the way, the best models by the biggest companies are called Frontier models or SOTA (state of the art) models. You might encounter these terms in some AI guides.
-
I never turn on the memory feature because with that turned on:
- AI makes a caricature of my personality. It picks up random things from random chats, and all the future answers are influenced by that caricature. If I ask it a legal question, it might save that in memory. Later on, it will give analogies related to law in an unrelated question. Maybe that was a one-time question. I don't want my answers painted by randomly picked chats.
- It becomes a confirmation bias machine. It will pick up my likes and dislikes, my political inclination, etc. and will give answers that makes me happy. I don't want happiness from AI, I want truth.
-
Instead of using the memory feature, you should add personal context/preference. Every AI chatbot has this feature. Go to settings, there you will find a feature named "Add personal preferences" or "Add personal context". With this feature, you can make your own caricature instead of letting AI do it. My personal preferences for Claude are:
- Think and explain from first principles. Do not worry about the token count. Never create artifacts unless I explicitly ask for them.
-
Every single word of your prompt influences the quality of the AI's answer. Every. Single. Word. It matters whether your tone is formal or informal, whether you sound like a 20-year-old college student or a 40-year-old professional. Once you've written your prompt, re-read to see if you can state your problem in a more expressive, elegant way.
-
Do not use voice mode for a writing prompt. We are less likely to edit our words when they are dictated in voice mode. But we are more likely to edit when we are already using the keyboard. And editing your prompt to almost perfection is important.
-
I often fail at estimating how much context to provide to get the right answer. So, I let AI get context out of me. I ask AI to ask me questions that would help it give a better answer.
-
Ask AI to help you with prompting. Give it the prompt that you have in mind, and then ask it to edit the prompt.
-
If you want a solution from AI, don't tell it your thoughts about the solution. Ask a question, let it answer first. If you mention anything about the solution, its answer will be anchored to your suggestion.
Tell AI what your problem is, give it the context of the problem, and then shut up. Let it answer. If you don't like the answer, then you say something about the solution. The current models are often smarter than us in many fields. Let's keep our ego in check; let's not make the AI dumb by saying something that we don't need to.
-
If AI's answer is something that you are going to use multiple times in the future, either:
- Copy-paste the chat into a markdown file.
- Or ask the AI to give the answer in a to-use-later markdown file.
Don't save chats in your account. Don't rely on one company. Some day, you'll need to switch to a different chatbot. All the important answers should be saved locally on your computer, not in your chat account.
-
Last one: develop interest in AI. Other than the work stuff, spend 10-15 minutes everyday asking random, silly questions to different models. Try to push their limits. This way, you'll understand them better. If you want to stay ahead, you'll have to put in efforts. In a sense, putting effort is another best practice.