The CORRECT way to use ChatGPT (in 2026)

ChapT has rolled out over 1. 1 updates per week this year. It's hard enough keeping track of all these changes, let alone to figure out which new features are worth our time. So, in this video, I'll cut through all the meaningless jargon like reasoning models are best for complex reasoning. Thank you. And instead, I'll just show you using real world examples when to use what feature. Let's get started. First and most

important, choosing the right model. I'll make this super simple. Most of us should default to the latest O number reasoning model with the cleanest name. Put another way, just like with romantic partners, pick the one without the extra baggage at the end. I'm blurring the model numbers on purpose because they change too quickly to matter. The key is knowing when you need a basic chat model and reasoning model. In simple terms, reasoning models perform best when the query is important or hard and you're willing to wait a little bit for a better answer. Whereas basic chat models, as denoted by GPT- number, are great for tasks that are low stakes and you want a fast response. Example number one, which fruits have the most fiber? A chat model will do just fine here, and we don't really care if the answer is off by a gram or two. On the other hand, for something like act as a nutritionist and create a vegetarian breakfast with at least 15 g of fiber and 20 g of protein, we need a reasoning model to think through our requirements. Example number two, who is the guy who said uh success is never final or something like that. Again, the chat model correctly tells me it's from Winston Churchill and that it's part of a longer quote. In contrast, I would ask a reasoning model to act as a British historian and tell me why Winston Churchill was ousted even after winning a world war since the answer will contain a lot of nuance. Example number three, for straightforward emails where we know exactly what to say, chat models can easily turn our bullet points into something coherent and professional. But if you're dealing with a 20 message disaster where someone's upset, you're going to want to upload that entire email thread as a PDF and ask a reasoning model to write a fake, I mean diplomatic reply. This last example disproves the claim that chat models excel at everyday tasks like emails. The model you choose should depend on the complexity of the task and not type of task. Closing the loop, most of us should default to reasoning models since worst case we wait a bit longer for a better answer. Three pro tips for prompting reasoning models. First, use delimiters to separate information. For example, put your instructions under a task section and the content you want acted on under a document section. This helps a reasoning model differentiate between what you wanted to do versus what it should analyze. Second, there's no need to include think step by step. This phrase helps chat models but actually hurts the performance of reasoning models. They already think step by step without being told. Third, examples are optional. I know this is counterintuitive, but reasoning models work great with zero shot, aka zero example prompting. Add examples only if you're getting wrong results. Now, one of the most common questions I receive from you all is, "Jeff, do you have a girlfriend? " Nope, that's a lie. Uh, no one cares. But a lot of you do ask, "Do you have an AI course I can take? " And

although I don't, this video sponsor, Corsera, does. Regular viewers know I've taken many of Corsera's courses, including AI Essentials, and I'm about to start their prompting essentials specialization, which they've created in partnership with Google. Corsera's annual subscription gives you access to over 10,000 learning programs from not just Google, but also IBM, Microsoft, and amazing professors from top universities. So whether you want to finally understand AI, get better at data analysis, or learn project management, you can learn at your own pace on your own schedule. Now that we're halfway throughout the year, which is kind of crazy, it's the perfect time for a summer reset on your learning goals. So, if you haven't made as much progress as you hoped in H1, no worries, because Corsera has got you with a special summer offer. For a limited time, they're offering 40% off their annual subscription, but only through July 21st. Click the link in the description to get started. Thank you, Corsera, for sponsoring this video.

Moving on to Chat GBT's web search feature. The biggest trap people fall into here is forgetting Google search still exists for now anyways and is more effective at certain types of queries. For example, chatbt with search enabled will give me Nvidia stock price no problem. But opening up a new tab and typing NVDA stock price is probably faster and gives us better peace of mind. So when exactly should we use chat search? The rule of thumb is if you need a single fact, go with Google. fact with a quick explainer, use chat search. Building on top of the previous example, instead of just asking for the stock price, I'd input when was Nvidia's latest earnings call, did the stock price go up or down, and why. And as you can see, not only do I get an interactive chart with the stock price, but I also receive all this context. Example two, weather forecast in Zurich should go to Google search. Whereas for weather forecast in Zurich from December 1st to 7th, what clothes should I bring? Chache search will return a much more meaningful answer. Example three with a twist. Google does great with latest vaccination rates who WHO, but if I wanted global vaccination rates over the past 5 years in table format, chatbt search gives me exactly what I need. And I can also ask follow-up questions like, wait, why are there multiple columns and not just one number per year? Two pro tips for chatbt search. First, it does a fairly decent job for quick fact checks. And I'll leave the prompt I use in the description below. Second, just like in Notion, you can use the forward slash command to togg on search from the chat box. Speaking of, if you're looking for a reliable system to organize your life and work, check out my notion course where I walk you through how to build your own command center from scratch. I spent the last 5 years perfecting this workflow. So, if you want to avoid the trial and error, click the link below to join 5,000 other students. Coming back

to one of the most powerful features in ChachiBT. In a nutshell, deep research is an agent that wanders off for 10 to 20 minutes, reads through dozens, if not hundreds of links, and produces a detailed report on a given topic. Diving right into an example. Instead of opening multiple tabs and manually going through Nvidia's AMD's and Intel's earnings reports, we can toggle on deep research and ask ChachiBT to analyze and compare the AI chip road maps for these three companies based on their latest earnings calls. Or how about a personal finance example? Instead of going through the websites of different banks, we can ask ChachiBT to analyze the top five high yield savings accounts available in the US, their hidden fees, and create a projection for someone saving $1,000 a month. We can even take this a step further by connecting sources we have access to like Google Drive so that deep research can draw from our private/proprietary information. In a business context, this unlocks extremely powerful use cases like generate a report on how our company's Q4 performance compares to our competitors. Cite both internal data and external industry reports. Pro tip: Deep research works best with detailed and comprehensive prompts, but since nobody got time to write that from scratch, we have two options. First, simply enter your research topic in this custom GBT from this awesome Reddit user to receive an extremely detailed deep research prompt template. You can then make edits as needed. Or if that's still too much work for you, you can upload a PDF of OpenAI's deep research helps article into a Chat JBT project and use individual chats to generate prompts for your research topic. As of today, I get better results from Chat GBT's deep research than Google Gemini. But with how fast the underlying models are evolving, I suggest testing both with the same prompt to see which one handles your specific task better. Moving on to

the canvas feature. The rule here is simple. Toggle this on if you know you're going to edit and build upon ChatBT's response more than once. For example, when preparing for my performance review, I would enable Canvas, upload my company's performance rubric, and ask CHBT to draft an initial outline, knowing that I need to make further changes. After the outline pops up in the canvas/standalone window, I would first go through and delete unnecessary sections, input relevant achievements, and follow the unspoken yet universal corporate rule to inflate numbers for better rating. Next, I make inline edits by asking Chachib to rephrase specific sentences and paragraphs. And once I've listed all my achievements, I can even instruct Chachib to generate an executive summary given all this context. By the way, canvas is extremely powerful for coding, but since I'm not a programmer by trade, I mainly use this for copyrightiting tasks. Three pro tips for the canvas feature. First, we can use the back and forward buttons up here to jump between versions. Second, we can use the built-in shortcuts within the canvas window to make changes to either the entire document or highlighted snippets. I especially like the suggest edits feature where ChachiBT will highlight different parts of our writing and recommend specific changes. Third, we can download the final output in markdown format. And after uploading the file onto Google Drive, we have a perfectly formatted text document in

Google Docs. As a bonus for those of you still watching, here are three of my favorite commands for texttoext models. First, use the word elaborate to add more detail. For example, elaborate on these three bullet points. Second, use the word critique to spot problems early. For example, I'm arguing for more headcount based on this data. Critique my approach before my presentation. Third command, use the word rewrite to improve previous content. For example, rewrite the second paragraph using a friendly tone of voice. If you found this helpful, you probably want to know when to use Google Gemini versus ChachiBT. See you on the next video. In the meantime, have a great one.