Cómo usar ChatGPT Work: guía práctica desde cero

0:00Until now, most people used chat like this: they’d write a question, wait for an answer, and copy the result. But what if, instead of asking for an answer, you gave it an entire folder and explained the result you need? And you let it research, analyze the files, and review its own work. That is what we are going to test today with ChatGPT Work. I’m going to provide data from a fictional business that is selling more and more, but is left with less money after covering direct costs, and I’ll ask it to figure out what is happening and turn all this information into an executive report and a

0:34presentation that we can actually review and use. ChatGPT Work is not designed just to answer you; it is designed to complete a job. But this doesn’t mean we can give it a vague command, walk away, and blindly trust whatever result it gives us. The real difference appears when we know how to delegate well, provide the right

0:54context, and review what it has done. That’s why, in this video, we’re going to understand what changes with Work, how to start a task from scratch, and what you should review before using the result in your work. And if you don’t know me yet, I’m Pedro, the artificial intelligence instructor at Hashtag Capacitaciones. On this channel , we teach you how to use the main AI tools in a practical way so you can work better, stand out professionally, and achieve incredible results. If you are interested in learning about artificial intelligence and its impact on work, give me a like and subscribe to the channel. And since we’re going

1:28to work on a practical exercise today, you’ll find the complete class guide in a link in the description, with all the materials I’ll be using so you can download them and follow the tutorial step by step. There you also have our WhatsApp group to receive notifications about new videos and more information about our impressive program. Now, let’s get started. Well , the simplest way to understand chat is to compare it to a normal conversation. When you use chat, you usually want an answer, an explanation, some ideas, or even to generate a document. When you use Work, the most important thing isn’t the answer within the chat, but the result that

2:05should be finished at the end. That result could be a report, a spreadsheet , a presentation, research, or a set of files. Work can review materials, use authorized tools, create documents, and continue working through several steps while we track its progress. The difference isn't that it writes more; it’s that it can execute an entire process. Imagine you need to diagnose a business. In chat, you could upload a spreadsheet, ask for an analysis, and receive an explanation. In Work, you can hand over an entire folder containing all customer reviews, meeting notes, goals, reference files, documents, whatever you need. You can hand them over and ask it to check the

2:49files, calculate metrics, cross-reference comments, create a report, prepare a presentation, and verify everything it generated. In other words, we're no longer just thinking about what the best prompt is. We are designing delegation: what result we want, what materials it can work with, what limits it must respect, and how we will know if the result is well done. Before starting, there is an important decision to make. In the desktop application, Work can work locally when it needs to access files or applications on your computer. If the cloud option is available, you can also use it for tasks that must continue even if you close the app, or

3:26that you want to resume from the web or mobile, for example. These are not two names for exactly the same thing. A cloud task does not magically enter the folders on your computer. It requires you to upload files, use a compatible project, or authorize a connected application. And a local task should only receive access to the specific folder it truly needs. If you don't understand what I'm talking about, don't worry, because in today's example , I'm going to show you how to access Work locally, as I want to work with a demo folder saved on the computer. All

4:00data will be fictitious or anonymized. If you are going to repeat this exercise with your company's information, first check what data you can share and what permissions you are granting. Now that we know where the work will be executed, we can prepare a task that allows us to truly see the difference. Alright, so this is your first time accessing Work. If you have never heard of this tool, don't worry, I want to show you exactly how you can access it. Alright, I am logged into the ChatGPT homepage. I should mention that this is a feature that will only be available in paid plans, okay? But that said, you can download the ChatGPT

4:39application to your computer. It's very easy. If you are logged in, all you have to do is go to settings. Click here, access help, and then click on download apps. Alright, I'll click on it. It will take me to this page where I can download the ChatGPT application by clicking on this button. Very easy. Once you download and install it—the process is quite common and standard— you log into your account, and the interface will be quite similar to what we are already used to. We have the main ChatGPT conversation interface here; if it’s not appearing for you, you can click here to access your conversations. From up here, you can

5:33also access Codex. So, it’s a fairly simple interface, and to use the Work functionality, all you have to do is click this button. It might be set to chat mode for you, but if you want to use these Work features, you need to click this button, and here we are. You’ll see that the interface changes slightly from here, right? Well, now we have a request approval feature, we have project selection, and we have plugins. I don’t want you to worry too much about that for now, but to start working with Work—which is basically this ChatGPT agent capable of taking action, making decisions, and selecting the tools it will work with

6:18—you need to be working within a project, preferably connected to a folder. So for this example, I’ll be connecting to an example folder I created on my computer, as this allows me to control ChatGPT's actions a bit better. Keep in mind that the moment I give it access to this folder, it will start working within that environment and be able to read all the files in it , so be very careful when granting a folder permission to ChatGPT if it contains any sensitive files or content . Alright. So, what I’m going to do now is select the project folder. And there we go, this is the folder we are

6:59going to use. We have a workbook with sales, costs, and customer reviews, some meeting notes, business objectives , and a reference for the format we want for the report. So, for example, if I want to use this spreadsheet we have here, the diagnostic data for this fictional company called Nova Casa. So here we have this spreadsheet with sales, opinions, a dictionary of terms, and so on. Not all spreadsheets are like this, of course, but what I want ChatGPT to do is read this information and combine it with the data in the other files, which I’ll show you now, to extract some very interesting insights. So, there you have it, this

7:41is our spreadsheet, and our files are in this format here—they are MDs, or Markdowns, and I open them with Notion, which is an app I love, but you can use any you like. Very good. So, for example, here we have these meeting notes from a supposed meeting I held with the sales, operations, and finance teams, and so on. So, uh, pretty standard meeting material, like typical meeting minutes. We also have something like a summary of business objectives with goals for April, criteria for recommendations, or perhaps a reference report, a standard report like the ones we’re used to using in our company, the mandatory structure, and the

8:24writing rules. In other words, this would basically be like a large prompt that the artificial intelligence will use as a basis to produce what we’re going to generate in this example. So we have the database, we have the meeting notes, we have this reference report, and some business objectives. Keep in mind that these are four distinct files, and ChatGPT, by being in this folder I’m about to give it access to, will be able to read them all and will surely take them into account when producing anything from them. Okay? Very well. So, let's go back to ChatGPT. Now it's time to select the project. So I’m going to

9:03select the project. You’ll see that I already have some projects, but since this is a new project, I must select " new project." Let's see, and here it gives me some name suggestions, right? "Trip to Copenhagen," but what we want is to analyze this company called " Nueva Casa." So, it would be something like "Analysis Nueva Casa," for example , something like that. And here you see

9:27that I can now add folders in ChatGPT. So I’ll click it and it will open this Windows menu to select the folder. All right, so I already know which folder I’m going to use. It’s called "Diagnostic Input." And you’ll see that the files don’t appear here because what I really want is to select the folder. So, once I’ve reached the folder I wanted, I click on "Select Folder." And from this moment on, this will be the folder that ChatGPT will work with. Very good. I’ll click " Create Project" and now we’re ready to start working. Notice something important. I haven’t created a perfect file that already contains the

10:10answers. The data is spread out because that is precisely the type of work we want to delegate. What I have done is remove sensitive information and give the files clear names. Artificial intelligence can help us organize our work, but it shouldn't have to guess what each document contains. In case you have many documents, ChatGPT will be able to open them all and interpret them. So keep in mind that the more files ChatGPT has to interpret to achieve the goal you initially asked for, the more of your session quota will be consumed. And now comes the decisive part. A good prompt for work has five elements. A result, the

10:49materials to use, the process to follow , the boundaries it cannot cross, and the criteria by which we will review the result. And this is the prompt I am going to use. I don't need to give a very detailed description of the steps or anything like that, but I do need to make it clear what finishing the job well means. Right, and here I have this prompt. You'll see that it's a fairly simple prompt; there's nothing very complex here, but it is a sequence of tasks. Let's see, so this would be the prompt. Work with the files in this folder and prepare an executive diagnostic for the fictional business

11:23Nova Casa. Before creating or modifying files, review all available material, list the sources found, present a brief plan, stop, and wait for my approval. So it is a way to control the result a bit better and ensure that ChatGPT read everything. To be honest, this won't always be necessary, but if the work is very important, very relevant, or very sensitive, I would recommend ensuring

11:52that ChatGPT is considering all files. But okay, then after I give it approval , analyze the data from January to April and calculate only indicators that can be traced back to the source files. Compare the results with the business goals that will be in that document we saw earlier. Contrast the quantitative data with the opinions and meeting notes. Clearly distinguish between facts, calculations, interpretations, and missing information. Identify the three findings with the greatest impact on decisions, even if they aren't the most frequent topics. Create a folder called chat outputs. Inside this folder, create executivereportnovacasa.md following the reference we gave it earlier. Also create a six-slide presentation called

12:41executivepresentationnovacasa.pptx, which is the format, with this sequence . Situation, status of goals, three findings, evidence, recommendations, and next steps. Do not invent figures or fill in missing data. Use "no data" when appropriate and point out any calculation you cannot verify. Do not confuse contribution margin with net profit; that is, just an additional recommendation. Upon completion, verify that the figures match between the report, the presentation, and the source files. Summarize what you verified, what you could not verify, and what requires human review. Let's see, what we are asking for is for you to analyze the information, give us the most important findings, create this report, put together a presentation,

13:28provide a summary, and verify that everything matches up. What do you think? Very interesting, right? And very practical. So, what will I do now? I am going to copy this entire prompt, and I don't need to upload files or anything like that, since they will already be in the folder. All right, let's go back to ChatGPT. Now I will

13:48paste the prompt we just looked at. Don't worry if it's not numbered like it was before; ChatGPT will understand it all the same. So, I always recommend using the o1 model with high reasoning. If there are many files and they are very large, then...I would also recommend using the very high setting, but I don't want it to consume so much usage for this session, so I think the high setting will be enough. And of course, always try to use the best model, in this case, o1. So, I will be using o1 with high effort, and now I will send the prompt. So, you will see that it opens a new menu here on the

14:28right with the results. So, the created files will appear here. And now it starts processing our request, our prompt. What happens? This can take about 3, 4, or 5 minutes. If you ask for something very complex, it can take up to 30 minutes. It has happened to me ; it can even take longer. This depends a lot on the complexity of the task you are asking for. So, the simpler it is, or the fewer the files, the faster it will finish. Or the opposite, in the case that you ask for something very complex. Okay? So, now I will leave it working. I can go get a coffee, I can

15:05go to the restroom or something like that, I can talk to my team colleagues, but at this moment ChatGPT is working for me. Very well. And after 2 minutes and 22 seconds, we have our result ready. And look, we aren't just asking for an answer; we asked it to deliver a complete result. It started by reviewing the files, showing what it thought and what it did to reach the result. And here I want to verify four things: that it found all the sources, that it distinguishes quantitative data from qualitative opinions, that it understands what the two deliverables are, and that it includes a final verification. If something is missing,

15:43this is the time to correct it. We can tell it, for example, before continuing , to clearly separate confirmed results from interpretations, and so on. This way we could avoid arriving at something that isn't a false correlation. Once the plan reflects what we need, we approve the execution. Alright, so let's see what work did. Here we can see it found the right sources. So, the diagnostic data, that's fine. Uh, the cost sales, reviews, dictionaries—I mean, each of

16:13the tabs that were in the file, right? We also have the MD files we saw before , meeting notes, business objectives, report reference. Very well. So, it was able to find the main files. Then it gives me this plan: let's see, recalculate indicators from January to April, compare January to April variance and target for each objective, contrast the results. Explicitly separate facts, calculations, and interpretations, create the files I asked for, reconcile a unique list of figures between Excel, and provide a

16:48final summary. Very good, very good. Fantastic. So, now we just need to approve the plan, right? So I'll say approved, and it will start working on the files. So notice it took about 2 minutes and 30 seconds to read the files, interpret them, and now it will probably take a bit longer to complete this task. There's no need to watch every step in real-time. We can follow its progress, answer if any questions arise, and review important actions before authorizing them. If it needs to access the web or any additional resources, it will ask us first, and we

17:28will be the ones to grant permission. So, we are always in control of what it is doing. This point is fundamental; the AI can perform the activities we ask for, but we remain responsible for the decisions, permissions, and how we use the results. Very good. So, look how interesting. It is processing the information. You see the timer keeps running, and it will start the analysis and such. It read the files, gives me some info, like the calculation structure is complete for the five requested objectives, right? This first sentence here tells me that since we didn't provide a visual template for the report, right? We gave it the

18:11structure and such, but we didn't actually give it a visual template. The presentation guidelines ask it to check available templates before designing the file. So, it's going to open a selector with options suitable for an executive diagnostic. So, what happens? From this moment on, it needs approval. Why? I need to approve ChatGPT accessing this tool to be able to create a presentation. And here is the request, right? The request, an executive financial presentation in Spanish for "Nova Casa," a fictional business, six slides, and so on and so forth. So, look how interesting. I can always allow, reject, or allow once, or

18:55always allow through this arrow here. Okay. So, I can allow it only for this conversation and such. I want to allow it only once. Let's see if it will ask me more times. So I click allow once, you see it asks me to select what the visual style of the presentation should be. So I can even select from some templates it has. I really like this simple dark mode one, this one we have here. So I'm going to select it, click continue, and it will keep working as normal. So, we'll let it work once more and come back to see the result. Very well. And after 20 minutes and 19

19:41seconds, look, what a long time it took to complete, our report is ready. And look how interesting, here we have the summary we asked for, right? It told me it's completed. The diagnosis concludes that Nova Casa met only one of five objectives. Revenue grew by 32.6%, but the contribution fell by 52.6%. The Starter Kit concentrated the deterioration and the Pro Plan sustained the result. Very well. So, here we have the executive report. If I

20:14click on this, it will open the report. Here we have it. It is in a markdown format, okay? But I could ask it to generate it in Word, okay? I just ask it to do it in Markdown so I can read it, edit it, and later, as it consumes more usage, ask it to generate it

20:32directly in Word, I prefer it that way. Okay. Right. So, we have the executive diagnosis, the executive response, the status of the objectives, the definitions, and formulas. Look, the whole report will be here in this little file that ChatGPT created for me , even with recommendations for the next 30 days. Additionally, it also generated for me—and look that from the moment I click, the screen shifts to the left, this new screen appears here, which is where I can read the documents, and we also have this folders section. If we go to the original folder, you will see that it is using exactly this format in the

21:18folder. Since ChatGPT Work works directly with our files, because it has access to a specific folder, it can create new files right inside there. And here it shows me where these files it generated for me are. So it generated two files which are these that we are going to see, and they will be in the folder. Okay. Alright. Then we also have the executive presentation . I click here, it opens up on the right again, and here we can see the presentation it generated for me. Let's see, let me look at the five. Here we have it, the six slides. And look, we have the objectives, we have the

21:56priority findings. Honestly, it's not the prettiest thing I've ever seen, but it does have charts, it has the summary of what's in the analysis, so I can read all the information, do a check, in addition to the check ChatGPT already did for me. Okay, let's see, so I can read here in the chat on the left what ChatGPT did. So, it verified the formulas, verified that all the figures match, that the report has seven sections, eh, the presentation has six slides, just as we asked, eh, the six slides were visually checked without detecting any overlapping or cropping; it didn't verify the net profit, inventory, unique customers, or

22:38representativeness. So, maybe it needs more information, okay? But with the information that was already there, this would be the result. What requires human review? Pausing or modifying the campaign, approving or minimum margin, evaluating the supplier, in other words , the next steps that require a human review, well, it won't give me recommendations that could affect my business. Very well, and if we go to the folder, you'll see that we now have this "salidas chatpt work" file, which are exactly these files it created for me. So, I double-click here, select PowerPoint, and it will open the

23:17presentation we just saw in PowerPoint. And look, here we have it. Spectacular, right? It's not the prettiest, I think, but it's not bad. So, we have the two documents we asked for, the verification, to see if the values match, and I would also recommend an additional verification just to ensure everything is correct. Okay? Very well. We go back to ChatGPT and here we can ask for adjustments, reduce text, change the order of the slides, ask it to incorporate other charts. So we can continue working with the information here. And asking for more changes depending on what we need. And look, here appears one of the most

23:55interesting advantages of using ChatGPT Work. We stop correcting isolated paragraphs and start directing the entire workflow. The quality of the delegation will determine the quality of the work, but the manual labor, the hard work, ChatGPT does for us. Our example ends with a report and a presentation, but Work can go beyond an isolated execution. If this diagnostic is repeated every month, we can document the process as a reusable skill. What files does it need? What calculations must it perform? What format must it follow? And what must it verify before delivering? Skills serve to preserve instructions, resources, and criteria that we want to apply consistently and continuously. We also

24:40have plugins that can add connected tools like Google Drive, your Notion, or your Asana, working with these tools by communicating directly with ChatGPT and what you are creating. And when the process works manually, we can consider if it deserves to become a scheduled task. For example, every Friday, review the available information and prepare an executive update. To create it, you don't need to learn a technical command . It is enough to write what it should do, when it should run, and where we want to receive the result. After the diagnostic, we could write something like this. Repeat this task every Sunday at 4 PM. I would hit send, and

25:29from this moment on, it will understand that this works as a repetitive task, an action it must always execute on this day, time, or schedule we asked for. Since this task depends on a local folder, the computer must be on, the application running, and the project available when the time comes. And there is another sensible rule. First test the flow manually, review its first executions, and only then automate it. So, let's see, here it asks me for this permission to be able to create this uh this automatic automation. So, I click allow once, and it will be creating this repetitive

26:12flow, this repetitive task. Very well. And here we have it. This task is already configured so that every Sunday at 4:00 in the afternoon, as long as my computer is on, of course, it will execute this task, okay? I can even click to open and access this configured task. So I can click open settings. If I want to check this within the appropriate menu, I can click these three dots; I can delete it , run it now, or pause this task. So all the tasks I have created will be in this menu on the left. I can access it from here and select scheduled. Then you will see that all the tasks I have

27:00scheduled in my ChatGPT appear. So, does ChatGPT Work replace everything we did before, right? It can greatly reduce the mechanical work of searching , comparing, organizing, and producing files, but it does not automatically know the context we never provide. Nor does it turn a weak source into a reliable one. If the data is incomplete

27:21, the report will be limited as well. If we grant excessive permissions, we increase the risk unnecessarily. And if no one checks the numbers, a visually flawless presentation can still contain an error. That is why, before delegating, ask yourself four questions . What is the exact result? What sources can you use? What are you not allowed to do? How am I going to verify the work? If you can answer them, Work has a much better basis to respond to you. If you cannot answer them, you probably still need to talk, investigate, or clarify the problem

27:53before delegating the execution. ChatGPT is not interesting because it writes longer documents. It is interesting because it forces us to change the unit of work. We stop asking for fragments and start defining complete results. The new skill is not writing the most complicated prompt, it is learning to delegate context, limits , and quality criteria. Let's take a quick tour of what we learned. First, we saw that chat is better for conversing and that work is designed to complete a result. Then we learned when it is convenient to work locally or in the cloud. We built a prompt with five elements: result, materials, processes, limits, and review criteria. We saw how

28:35to check the plan before approving it, how to review the traceability of the numbers, and how to improve deliverables without starting again from scratch. And finally, we understood that a proven process can become a skill or a scheduled task. And since this is a practical tutorial, you don't need to reconstruct it by watching the video frame by frame. In the link in the description, you will find the complete class guide and all the materials used so you can download them and repeat the exercise. There you also have our WhatsApp group and more information about the full and impressive program. And now I want to know your opinion. What complete work

29:11would you like to delegate to ChatGPT Work? Not an isolated question, but a result that currently forces you to open several files and use several tools. Tell me in the comments why we can use some of those ideas in upcoming tutorials. ChatGPT can execute more steps for us, but deciding which result is worth it, providing good sources, and checking quality remains our job. I hope you liked this video. Give me a like and subscribe to our channel. See

29:42you in the next video.