Cómo usar ChatGPT Work: guía completa para principiantes (2026)

[0:00] You can now give Chad GPT a job in the cloud. Close your laptop, walk away, and come back to a finished report, a spreadsheet, or a website. And by the end of this video, you're going to know how to use ChatGpt Work, how to use codeex to edit your files, work across your apps, and run recurring jobs. That's right. This is the complete guide I am issuing now for everyone as to how to use these tools. You'll know which features to use, which settings matter, where I'd start. You don't need to know
[0:30] how to code to do this stuff. I want everyone to have this kind of a guide because if you don't have it, you're going to be behind. And if you do have it, and if you use these tools, this is an in-depth guide. You're going to learn something new. I'm Nate B. Jones. I use these systems across my work, including jobs that have nothing to do with software development. And we're going to move all the way through a complete guide, every aspect of working codecs. We're going to move through practical tips you can use as you go, like editing a doc without rewriting everything,
[1:01] getting useful data out of photos, talking to a task while it works, setting up work that runs again and again and again. And we're going to start with how to choose your experience. Now, chat is the natural place most of us got to know chat GPT. So, you can chat for a question, an explanation, a quick back and forth. You want to choose chat GPT work when you want an assignment carried through to something you can use. So if you want to produce a file or an analysis or an
[1:31] update in an application. Codeex has particular strengths in software work, but it also handles research documents and work across apps. And if you're comfortable there, it's going to work just like work for you. In fact, OpenAI's own guidance says that if you're already using codecs for non-coding work, you can just go ahead and stay there. work offers the same core capabilities as codeex in an experience designed around everyday work, not coding. That doesn't mean every button is identical. It just it's about choosing the place that can reach your material and do the job you have in
[2:03] mind in a way that makes you comfortable. Now, Astra is a model you can select inside those different experiences. It isn't the name of another app you need to install. And there's a useful availability detail here. Plus includes Astra in work and codecs as it rolls out. While ordinary chat access will differ. So you want to check the relevant picker and update the desktop app before concluding that something is missing. The recorded examples I'm going to show throughout this video indicate what these tools did
[2:33] on those particular runs. They aren't new benchmarks that I would set for a particular model like Astra. They're meant to show more evergreen capabilities so you can get to know this tool as it evolves. Because one thing we know about AI is it's going to keep going. All right, let's tackle that first big capability I teased at the top of the video. How can I keep AI working when I close my laptop? First, check where the task is running before you leave it with a long job. A task on your computer and a task in the cloud have
[3:04] different requirements, even if the conversation looks really similar in the app. So, in the desktop app, look for the work locally control when cloud is available. That's the option for work you want to continue independently of this computer. It can use sources and tools available in that cloud environment. A supported cloud task can keep going when you close the app, turn the computer off, walk away for the night. The cloud agent is just going to keep ticking away on that task. Now, you
[3:34] want to choose local when the job needs a folder, an application, or another capability that is unique to this machine. Local work can run in the background and local schedules are possible. So you can schedule something to be available and schedule something to retry at a particular time. But here's the critical thing. Your computer needs to remain on and available with the app running if you're going to do local work. And where the setting is available, you're going to want to prevent sleep while running a big job to help keep your long local job from
[4:05] stopping just because the machine went to sleep. Now, here's an easy mistake you want to avoid. You might choose cloud because you want to walk away and leave, but you might accidentally refer to a file that sits on your local desktop along the way. The cloud task does not inherit your local files. So, you will need to attach any material or put it into a connected location the task can reach, and you should confirm your agent can reach the files before you walk away. The reverse is worth knowing, too. A local task can use a
[4:37] connected online service, but local doesn't mean that everything has to stay inside the folder. The question is which environment can complete the entire job that you gave it with the access that you've provided. Now, I would make this choice before attaching material and setting up recurring work. So, for a task that needs to run, say tomorrow morning while my laptop is off, I would choose cloud and check the sources. for work inside a folder on the Mac, I would choose local and keep the machine
[5:08] available. And by the way, changing the model does not solve an environment problem. If the task cannot reach the file, a fancier model is still going to need a reasonable path to the file. It's just the way it is. Okay, let's get to prompting. Can I just talk through what I need? Yes. And there are two different features to pay attention to here. Dictation turns your speech into text that you can inspect before you send. Whereas chat GPT voice is a live conversation with more limited tool
[5:38] access. You can start work, you can ask how it's going, and you can change direction without returning to the keyboard. In both cases, in a supported desktop task, you can choose start voice chat to go straight into a voice chat. And for a new conversation, you can use start new voice chat. If the option isn't showing inside an existing task, you should check the app and host updates as well as your account's availability. You can still use dictation when you simply want to get a prompt down quickly. In fact, I sometimes do that by
[6:10] just clicking the voice memo in chat GPT worker codecs and sometimes I use an external tool like whisper flow. They both functionally work the same way. What I'd use voice for is the context that comes out more naturally when I talk. This is for someone who already knows the background. They're going to care about this part. Keep the recommendation short, but leave the evidence available. This is going to tell you much more than make a professional document. And the model will interpret your verbal context in ways that you don't even realize. You're
[6:41] giving the model so much to work with. And you can just keep talking while work keeps going. I have had 20 minute conversations that are all one-sided with AI where I just tell it for 20 minutes what I want. So ask where you need a decision made. Ask where you need to narrow down a request or explain why an answer isn't useful yet and like modify what you are seeing on the screen with your voice. Voice can also help coordinate separate tasks and bring progress back into the conversation. You're not limited to reading one task's
[7:13] response aloud. You can have a lot more flexibility. For a complicated brief, I would ask for short written feedback from the model. And I would want to catch the wrong audience or a mistaken permission really quickly before it shapes the whole job. And you can just ask for that directly. It doesn't require restaining every detail. You can just confirm that the outcome looks good. Confirm the constraint you set is met and then say, "Okay, this is what should happen next." On supported Max, screen context is another useful option
[7:44] under settings during the conversation. It means you can ask it to look at the window you're discussing and that can make this part right here feel very useful instead of just as an ambiguous reference cuz it can literally see what you are looking at when you say this part right here. Choose that window deliberately. The captured context can include accessible text outside the visible scroll area. So don't assume that only the words currently visible on the screen on that page are included. There can be more than that. If you use
[8:14] voice often, there's a voice chat hotkey setting you can set and configure. Make it easy to start the conversation. The value is being able to give the missing context where it occurs to you while you're still looking at the work. So, you don't want to make it hard to start chatting. Let's get to fixing docs. Can AI fix a document I already have but do so surgically? Absolutely. You don't need to start a new doc every time. With the right connection, work can actually edit inside Google Docs, inside Word where it already lives. It can also
[8:45] create and revise doc files on other programs, too. The useful instruction is often a much smaller and more precise phrase than rewrite this. Don't be generic. Try saying what specific parts that you want to modify must change and what parts must stay. You could say something like, "Keep the headings, preserve the links, change the explanation in this section, add comments below the relevant paragraphs." If you're only going to ask for generic improve the doc, you're leaving the AI
[9:15] model to make a lot of the editorial decisions that you should be making. In one of the recorded jobs I'm showing, I asked for a copy of an article with six short red review notes. the original stayed untouched and the copy kept its formatting, its links, uh, all of the payw wall markers, everything that made the doc the dock. And that's worth paying attention to. The model can make a precise change without rebuilding the dock, as long as you can say, even with your voice, this is how I would change it. A copy is a good choice for an
[9:46] important first attempt. Once you're comfortable with the way editing works, you can absolutely authorize an edit to your working file. But I always say, don't modify till you know what you're doing. So if you're new, name the exact file, provide the link, make sure that you're familiar with the changes before you make changes to something you're working with all the time. Now, if the AI asks you to paste a document you've already connected, ask it whether it can read the connected version. There may be an access problem that's worth
[10:16] resolving. You should not have to create a second quickly outdated copy just because the task hasn't figured out where to find the source yet. And not all AI models are super forthcoming about this. Recent chat GPT models are better. For a new doc, you want to give it a real example of the format if you have one. So say who's going to read it and what they need to do afterward. A person approving a decision needs a very differently structured doc from someone learning the process for the first time.
[10:46] And even that much context spoken verbally can be enough to give AI a sense of where to go next and how to format the doc for you. When the result comes back, absolutely check the actual file. Check what changed. Glance at the pieces you asked to see if it preserved them. For a formatted doc, make sure you're inspecting the way the pages render and make sure that text is correct. And particularly check for table overruns. Again, newer chat GPT models are better at this. One direction I use a lot is to be specific about the
[11:18] level of your change. Use specific words. A review, a small edit, and a complete rewrite are very different assignments to the AI. Tell it which one you want, and you can get very useful help without giving up the parts you've already decided. All right. What about spreadsheets? Can I trust AI with a super messy spreadsheet? I had, by the way, a 29 tab spreadsheet years and years ago as a marketer. That's what I always think of. work can absolutely turn a messy workbook into something you can filter, sort, and actually use. But
[11:49] before you ask it to clean the file, make sure you tell it what clean means. Changing the appearance and changing which records exist and how formulas work are very different Excel jobs. Look at duplicates. In the supplier workbook, the rule was exact agreement across every original field. That matters because supplier names were inconsistent as well. So if you standardize the names first, two records that were different in the source may suddenly look identical when you didn't mean them to.
[12:19] So think about your order of operations when you start to work with a large spreadsheet. The recorded run I'm showing identified exact duplicates before standardizing names, which is the right way to do it. It removed 236 duplicate rows out of 1516. It left 1280 records and it preserved all seven real grant records in the test set. You want to define the duplicate rule before the agent changes fields that it's then going to have to compare. Of course, ask for a cleaned copy and a short
[12:49] explanation of what was removed. If you're making these kinds of changes, you should be able to reconcile the starting total, the removed amount, and the final total. And if a record disappears, you should be able to find the rule that caused that record to go somewhere. Also, just ask for a workbook that you can work in. I know that sounds obvious, but filters, frozen headings, proper data values, and consistent currency formatting are not only useful changes. They're changes that you would expect and you can ask for with these new models. A date that only looks like
[13:20] a date can still sort incorrectly. A date that is formatted as a date won't. A displayed currency symbol doesn't prove the underlying value has been converted to currency. Make sure it has been. If formulas are involved, say that they need to remain formulas when you expect the workbook to be editable. Don't just freeze the and hardcode the values. If formulas are involved, you want to say that you need formulas to stay formulas when you expect the workbook to be editable. And then you
[13:51] want to test that by changing an input and seeing whether the dependent result, the output, changes as well. A sheet full of plausible static numbers can look finished while being much less useful than what you originally intended. Have the task reopen the exported file when it's done and check it rather than relying on the version it held while working. That catches a different class of problems. What was actually saved may not quite behave like the internal draft that chat has been
[14:22] working on. I would use this on a file where I understand the records well enough to define all of those rules and then the model can do a lot of the inspection and cleanup and then you can supply the meaning of a real transaction and the changes you're willing to make and that can help you to apply those edits successfully. So I guess what I'm saying is just as I had to learn what the 29 tabs on my marketing spreadsheet actually meant, you should probably know what your spreadsheet means if you're going to edit it. All right, next tip.
[14:53] What can you pull out of a bunch of photos or how do you work with images in chat? GPT work and codecs. Work can read a set of photos and turn the information into a table that you can sort, check, and use elsewhere. A receipt example is really useful in a lot of work situations. You can begin with 30ome photographs as I'm showing and end up with 30 traceable rows in an expense workbook just like that. Each row can keep a source file name. so you can see where it came from. And that little
[15:24] detail is one of the best instructions that you can add to a good photo extraction job. Keep a way to get back to the source. You can just say that verbally and that will remind it to include that column. And you want to in these situations tell chat which fields you need before it starts. So for receipts you would say I want a merchant column, I want a date column, a total, a currency, a category, and anything that needs review. uh for a different photo collection, you could choose the fields that make the result useful to you.
[15:54] Don't ask AI to just invent a complete database because it can make it useful. And this is a little gotcha that I found working with a specific receipt example. Two receipts may not provide a taxable amount. I certainly have those on my receipts. So those cells may stay blank and get flagged. zero might be used in those situations where the source explicitly supports it, but a blank and a zero in terms of tax are very
[16:24] different answers. And so you want to be specific about your edge cases when you're classifying images for things like finance. To help with these kinds of problems, you want to ask the model to keep the difference visible. So saying if the source doesn't say leave it blank and mark it for review, is a much better instruction to AI than expecting every salt to look finished. You can also ask for an explanation of ambiguous entries so you know where your attention would help one way or another. And check completeness before obsessing over formatting. That's a general rule
[16:55] for working with any AI. Did every image produce a row? Did any image appear twice? If the folder contains unrelated files, did the task distinguish them from the material that it was supposed to process for this job? Those are the kinds of questions you should be asking regardless of the task. You can also apply this to other sets of visual records, but you want to test the fields on examples first. Another good general tip, test on a few examples first. A layout that looks really obvious to you
[17:25] can still contain ambiguity that AI has covered up. Like a subtotal may be beside a total, an issue date beside a due date, or a handwritten correction may exist, and AI may not pull those out till you try a few different ones. I like this kind of image assignment because the source and the output are really easy to compare. You can open up the rose image, you can make the decision, you can tell the task how to handle that case next time, and it can get smarter and smarter. And ultimately, you get a reusable improvement from a specific correction. And that's how you
[17:55] get something like pretty much automated receipt and expense filing. Let's jump over to the research side. Is chat GPT work going to tell me when research disagrees with my hypothesis or is it just going to flatter me? One of the most useful things you can ask chat GPT work is to preserve disagreement. And that's true of research or any other task. A confident summary is easy for you to read, but it can hide the uncertainty that you really need to make good decisions. So, I would start by
[18:25] saying which kinds of sources really matter. If you're comparing product capabilities, you want to be using current product documentation. If you're looking for customer experience, you want to be including your customers, right? That makes sense. If you're researching the market, you want to be asking what are you actually measuring when you look at this market research. Those are different research jobs, but you need to ask for the evidence that's going to challenge you in each case. Also, I would ask for publication dates
[18:55] and the date the underlying event happened. So, if you're trying to get research done and understand when something happened, a recently updated page can be deceptive. It can still describe old information. And so, in that situation, you want to be just as we talked about with the spreadsheet, someone who's familiar with the material and able to distinguish between, oh yeah, this is really recent information. It's well sourced. I get multiple sources across different sites and oh,
[19:26] this looks like a retelling of an old story that isn't well sourced. It doesn't matter that it's a new date. I'm not going to trust it. And you can actually make this specific. You can give that instruction up front. If sources disagree, show me the disagreement. Explain whether they're using different definitions and be explicit about your assumptions. Sometimes the apparent conflict is something that you can correct with your perspective. like a market size definition or a marketing geography or a time period that you're evaluating.
[19:57] Sometimes it's going to remain unresolved. Either way, you want to know. Now, if you're tackling a big research job, you want to ask for a compact evidence table before a report. So, the claim, the source, the date, the uncertainty, all have to be reviewable. You want to be able to catch a category or unsupported inference before it spreads through the rest of a really big, and I'm talking 50, 60, 70 page piece of research. Then you want to ask the model to make a recommendation. Keeping uncertainty visible doesn't mean
[20:27] refusing to help you decide. In in the recorded market work that I'm showing here, the proposed customer segment, the price, the initial product direction are all labeled as hypotheses, but the next step is to test them with real customers, not to treat the spreadsheet as some sort of claim that you've done your due diligence and the market really exists. And so in this situation, you want to be comfortable enough with the data that you can push on it. If the conclusion feels too broad, ask what
[20:57] would change it? Which missing piece of evidence matters most? What's the cheapest useful next test? What turns, oh, we're going to do more research into actually no, this is a job that helps me develop conviction and I can move forward. And don't confuse a long source list with conviction. I would rather have a few relevant sources attached to the actual claims than dozens of links that make the output look really thorough, but use the research to reduce the uncertainty that I care about and make me feel falsely confident. Now, can
[21:28] that same research become a deck and a report? It definitely can. You can ask for several related outputs in the same task. And that's useful when you need to agree about the same evidence to produce useful outputs across multiple output vectors. It also saves you from repeatedly explaining the assignment in separate chats, which was something we had to do in earlier AI models. The meeting notetaker research in this example produced a workbook, a written report and a 14 slide deck. The useful
[22:00] feature here is that there are shared source materials driving all of that. The workbook can hold details. The report can develop reasoning for that and the presentation gives you a way to share that decision. To do that successfully, you need to give each output a different job. Don't simply ask it to squeeze the report into slides. Instead, tell it what the audience already knows, what they need to decide, and which supporting details belong in the notes or in an appendix. For a deck,
[22:32] you want to provide a real template if you need one followed. So, ask for editable charts. Ask for elements where you expect to make changes. An attractive picture of a chart, as you know, is a lot less convenient when someone asks you to update the numbers. So, ask for what you need. Here's a quick way to review an argument when you're working with multi outputs like this. If you're reading through the slides, read just the slide titles. Do they tell you where the presentation is going? If every title AI gives you is a
[23:03] topic label, ask for titles that state the point of the slide. Ask that they tell a story and then inspect the slides themselves for legibility and crowding after that. Now, sources can go in speaker notes. So the main slide stays readable. And in the recorded deck, the notes can carry source links. So you don't have to lose that information as you go. In fact, the person presenting can reach the evidence in the middle of the presentation without ever jumping off of the slide. One thing that I would make explicit is how corrections should
[23:34] be made. So if you say I've changed this assumption and you're working with multiple outputs like I'm describing, you should then specify I want you to update the workbook, the report, the deck, and then confirm that they agree. Being in one task across all of that work helps the model coordinate that. It does not though mean a correction will automatically update every file unless you are asking for it to be done that way. Save the finished outputs somewhere you can find and feel free to edit them
[24:04] yourself to polish. If the work was assembled in a temporary cloud workspace, I would not leave your only usable copy buried there. Pick a folder you can get to and get the get the files there where you're comfortable with it. Now, can AI use websites I already use? The answer is yes. And it's really cool. Computer use has gotten a lot better in chat GPT work. The browser, I find, is super useful when the answer depends on getting the information off the website because it changes frequently or when the task needs to interact with the site
[24:34] because of functionality. Uh, and you can let AI search and compare while keeping a way to inspect the pages that it used so you can see the history of what it did. Local browser work can use supported browser tools on your computer. And cloud work, as we talked about at the top of the video, will have its own browser session. Do not assume that being signed in on your laptop on one browser means the cloud browser or another local browser is signed into. That's not how it works. If a login is required, you'll want to use the
[25:04] supported takeover step to securely enter your password and a security code yourself and then let the task continue. You should not need to put credentials into the conversation and you should not do that ever. The Japan travel example I'm showing here shows how I would go through that whole process. The recorded search found a connecting round trip around $770, a non-stop Tokyo round trip at 848, and an itinerary flying into Tokyo and home from Osaka for $855
[25:37] bucks. Now, that third option cost only $7 more than the non-stop round trip, but avoided roughly $90 of backtracking. And so whether that's useful to me depends on how I read the route, how I want to plan my vacation, not just the pure lowest airfare. And those were prices from that particular search. And that's going to vary, as we all know, depending on where we search from, what currency and geo we're searching in, and what day we search in. So in that situation, we want to give the browser
[26:07] the constraints that affect the overall experience. We want to talk about dates, number of travelers, location, delivery area, or whatever determines whether a result is actually usable. A hotel price for the wrong occupancy is not going to give you a good deal on your vacation. And then when the AI returns options, you want to ask for direct links and you want to ask for details that you can check before making a choice. What's included? What's conditional? Which information did it verify? Which part still needs confirmation? Put your
[26:39] thinking hat on. Does this make sense? You can also narrow a broad search as it runs, which is one of my favorite things to do. If the first results you're seeing print reveal a tradeoff you had not thought about, tell the task to either keep going that way or to change direction. There's no need to let a chat response finish. If you are seeing a bunch of unsuitable options just because the initial prompt was too broad, change it in the middle. For a purchase or a booking, keep that final action clear in
[27:11] the instruction because researching, an itinerary, and buying are very, very different permission sets. I want the browser to do the searching and comparing. I want the final choice presented to me and then I want to make that choice. What about connections, plugins, work in actual apps? Can AI help with that? So, a connected app on chat GBT work and codecs does let the AI work with the service you are to use. And that's how something you give the AI can end up somewhere else in a place you already go like Google Docs or Slack
[27:42] instead of as yet another answer for you to copy and paste. So to get there, you want to open plugins. You want to look for the service that you need and you want to look at what the plug-in provides. Then you can install it and connect the relevant account when prompted. Common ones are email, Slack, Docs, etc. Some connections authenticate during installation. others ask when you first use them. Regardless, after installation, you should start a new task to make sure the capabilities are available. You can name the service in
[28:13] your prompt or use an at mention where supported, but be specific about the account if you have more than one. Like my calendar, for example, may not identify which calendar you mean if multiple calendars are visible, which is possible now in chat GPT work. So in this case, the school calendar job went from an image to entries in Google calendar. The task proposed specific changes and items. I selected the destination of where they're going and I approved timed events that were created
[28:43] and checked in that destination calendar. A dateonly school closure was then kept separate rather than being given an invented time. And that's kind of a practical tip to take away here. You want to decide where your tasks go with these connections and these integrations and you want to preserve what you are saying and not saying along along the way. Right? A calendar entry needs to have the right time zone and you should feel free to specify that. If an event should not invite everybody, then say who it should invite and who it
[29:13] shouldn't invite. Adding an event to your calendar and sending invitations, those are different things. You should say what you want. And before a batch write into any of these integrated tools, ask the AI to check for existing records and then have it read that destination again afterward to see what changed. That helps you catch both duplicate creation and a task that reported success without producing the expected result. An example would be duplicate entries on multiple calendars. You don't want that. Now, connections can differ. A service name in the
[29:44] directory doesn't guarantee that every action is supported or permitted by your account. You want to ask whether it can perform the particular read or write what you need. If it can only read, you may still be needing to do more research or drafting before you can copy and paste because it's only able to read and understand things on the ingest. It's not able to write. As usual, I recommend beginning with a task you understand well. Once that works, figure out how to do it again and how to start expanding
[30:14] how you use that tool. Here's one. How do I stop uploading the same context in a bunch of prompts? Use a project when the work continues over time. That's where related tasks and sources and instructions can stay together so the next assignment doesn't begin with you constructing all of that background. There are a few places that files can live and they solve different problems. For example, you want to attach a file directly when it's only for this task. You want to use library in chat GPT for
[30:46] material you want to reuse where that feature is available. And you want to connect the current source to the chat when the information keeps changing and a flat file won't work. A saved copy of last month's document won't update just because the title still looks right. And so you want to be aware of where the original of that lives and point AI to it where relevant. Now in a chat GPT project, you can look at the sources for the shared material for the project and start the new task from that project to
[31:19] take advantage of all those sources. Right? Project instructions can carry your notes on the audience on terminology on expectations across chats and that's a really useful place for context that applies to the entire project. A local project in the desktop app can connect to folders on your computer and you can open the project menu and choose edit project to inspect or change them any time. Supported local projects can actually have more than one attached folder with one set as primary, but be really deliberate about which
[31:49] folders you make available so you don't overload the context window. The primary folder matters for where new tasks begin and where project guidance is discovered. Adding a second folder makes its files available, but it doesn't mean every instruction file inside that second folder automatically becomes part of the project's instructions. Instead, think of it as more like a book that the AI might pull off the shelf. So in that situation, I would keep the active
[32:21] folder really easy to identify and I would say whether a current draft or workbook needs to go and get a particular file beyond that. And pay attention to your file naming. If you have files named final, final two, and final really, you're going to confuse both people and models. So make sure your files are named appropriately. It's not new advice, but it's really relevant. Then at that point, organize for your convenience. You can pin a project you use often. You can give tasks names that describe the results. You can archive work you finished with.
[32:52] You can pin changes in your sidebar. It doesn't give the model extra context. It just makes it more usable for you. You can also use the same task for continuing corrections when existing context is useful. So start a separate task for independent work and that keeps a small edit from turning into a search through everything that you've ever discussed in that space. Can AI find the thing I was working on? Yes, it can. Computer history can help you find the
[33:22] document you had open before lunch, the page you were comparing, the task you left unfinished before a call. You can remember the work without remembering the file name. And it's one of the things I find most helpful about working with AI. Where available, you can enable computer use and ask about that activity in ordinary language. And AI can look through local files and the record of computer use to help you identify what you were looking at and get you back into the work. And that's a totally
[33:52] different kind of help from searching a folder for a title you can't remember. Although AI can do that, too. There are some specific availability details here. Computer history is a Mac OS feature for supported pro, business, and enterprise accounts. It is off by default, which is wise, and it requires enablement of memories. In a managed workspace, the administrator will need to grant access to that, and you will still need to choose whether to turn it on at that point. Why the safeguards? Because you're enabling AI to see more of your
[34:22] computer. You have to choose that. I think that's right. Look under settings and then computer history to turn it on. You want to review the apps and websites included before enabling it. And then you can use an include only approach if you want to begin with a small set of work applications. So there's ways to say yes in small ways before diving all the way in. The feature uses interaction events and accessible text to create a referenceable history for chat. According to current documentation, it does not include screenshots and it does
[34:53] not include uh microphone or system audio and private browsing is not included either. But typing and accessible text can still contain sensitive information. So choose the included sources accordingly. Now if this is likely to capture communications with other people without their consent, Open AAI recommends that you turn this off unless you have their prior consent. And that's correct. In fact, that's applicable in relevant legal jurisdictions. Now, if you want to go back and look at how this worked in the
[35:23] past, you can actually inspect the timeline of past memories. You can pause collection. You can delete history. Pausing will stop new collection, but of course, it doesn't erase what is already there. If you want earlier material removed, you'll need to use included history controls for that. The other thing worth knowing is this. Some processing happens on OpenAI servers to generate these memories which are then stored locally. So don't equate a local history feature with a promise that all of the processing stays on your Mac.
[35:54] That wouldn't be true. So if you're thinking about enabling this, the question I'd ask myself is this. Is it worth it to me? Do I use it often enough that I think that it is something that I'm willing to make the privacy tradeoff for so that AI is more useful to me as a result? I think everyone's going to have their own response to that. Obviously, if you work at a company, your company's going to have an opinion. But regardless, you should know what's there. You should know how to use it, and you should know how to modify it. Ultimately, history adds another way to
[36:25] recover context and it does not need to be enabled before work becomes useful. As I hope you're seeing from this video, there's so many other ways to use chat GPT work. How do I teach chat GPT work how I work? Well, one, save the decisions you don't want to repeat. Who are you writing for? Which source should win when two files disagree? Where do finished results go in your system? Those are useful instructions that are kind of standing rules and the model cannot reliably infer them from the
[36:56] current prompt because you're probably not going to give the same prompt every time. Project instructions can apply to a continuing area of work. Of course, a skill is a reusable method for a type of job. A plug-in can package skills and connections together. And the connection gives the task an ability or access it didn't have otherwise. Now, if you're just using a skill, that will give your AI the ability to do something in a particular way. A skill is really just a markdown file, like a recipe. In a local
[37:28] project, you may see a file called agents.mmarkdown. It's a plain text place for project guidance. You can ask the assistant to help write it, and you can keep it specific to how the work should be done. So, a I can help write it for you rather than collecting every piece of prompting advice you've ever encountered and dumping it in there. I don't recommend that. In fact, I recently cleaned up my own codeex instructions. One writing workflow had accumulated about 10,000 words of required reading for the model. The revised workflow cut it to,200. I
[38:00] kept all the standards. I kept all the source rules. I kept the authority boundaries while removing repeated procedures and material that every task did not need. Sounds like a win to me, right? That reduced the required reading that the AI had to make before helping me. Now I got significant token savings out of that because that was coming every single conversation. The old instructions repeated themselves. They made routine jobs do unnecessary work and I added to them over time to solve
[38:30] specific problems but it kind of bloated out the model. I encourage you to think about that if you're a regular skill adder. You may need a skill trim. Now, if you're using a skill, you want to start from a job that already work. You can ask AI to capture the method, the example format, the corrections that matter, and the checks that made that result usable. Then try it on a different input. Otherwise, you may have saved a description of one successful file rather than a reusable method you
[39:01] can use as a skill over and over again. I would not turn every prompt into a skill. save the work that you really do repeat and something that several people need or something where your judgment keeps getting lost and you have to prompt for it and then you can revise it as you learn. And when a skill comes from someone else, please, please, please inspect what it includes before using it with any of your computing material or installing it. It may contain code as well as instructions. A
[39:32] convenient method still needs to fit your task and the access you intend to give it. And so don't mistake, oh, I can install a skill for, oh, it works for me and it's secure. Okay, let's jump to the question everybody asks. Which model and how much thinking? You choose the model and the amount of reasoning separately. The model determines which system is doing the work. The reasoning effort determines how much time it can spend working through that task. Look beneath
[40:04] the prompt box for the model and the reasoning controls. Some accounts will show power with faster and smarter choices, plus advanced controls for a specific model or effort. The options will vary by your account and roll out, so use what your own picker offers. As of this recording date, Astra is the strongest option I'd consider for demanding work across several steps and tools, particularly when the job needs careful analysis or visual judgment.
[40:36] Soul remains useful for a lot of substantial work, and Terra is a practical option for familiar everyday assignments. Luna is surprisingly powerful and worth trying for clear, repeatable tasks that don't need strong reasoning or where you've structured the ask with other agents. it can be very coste effective. Those are just starting choices you will have to try and your mileage will vary on particular tasks. I would encourage you to use those as guides. Compare the result from a particular model's work with work you
[41:06] understand. If a routine gets noticeably worse on a less expensive option, the savings may disappear in the time you spend fixing it. So you may want to move up to a slightly fancier option. I would start with whatever your default reasoning level is and I would raise it when you have a good reason. Conflicting evidence, a difficult plan, complicated reconciliation, all of those can justify more thinking. Reformatting a clean paragraph, please don't use a fancy
[41:36] setting. If you have settings like max or ultra, do not treat them as interchangeable names for better. Max can give more room for reasoning. Ultra uses sub agents to work on separate parts of a very complicated assignment. And parallel work can be useful when the pieces can be broken apart and handled independently. And we'll get into that next. But not every piece of works that way. A stronger setting also doesn't resolve missing information. If AI has
[42:06] the wrong source or doesn't understand the decision, you got to fix that first. Asking it to think longer about the wrong file just makes the wrong task much more expensive for you. Now, for a recurring job, you want to spend enough effort to establish a result that you trust. And then you want to test whether a lighter setting can repeat that effort level. Keep the comparison about the finished work and your own review time so that you understand what you're actually spending. You can make the choice to repeat that work with a
[42:37] smaller model once you understand what you're doing with the first run or two. Okay. When should you ask for multiple agents? I would ask for a parallel work when the pieces can truly move independently. So, for example, separate retailer searches are a great use case that's not a coding use case. One search doesn't need another retailer's results before it can begin and run. Uh, in the sofa assignment I'm showing you here, agents can search different stores under
[43:07] the same exact requirements at the same time. They can search for budget, for dimensions, for material, for configuration, for delivery, and for return terms. The results can come back in a very consistent format and the main task can then turn them into a really clean comparison. That is a great example of how you actually farm out a multi- aent task. You want to give each agent a clear piece of work and specify what it should return. If one brings back a product link, another a paragraph of enthusiasm, and another a price
[43:38] without delivery, then you obviously have more work to do. Whenever you can make sure you insist on consistency when you're talking with the lead agent, so the lead agent can insist on that across multiple agents as it farms out the work. Also, be clear with what you expect from the final answer. The main task should be able to reconcile conflicts, eliminate options that don't meet the brief, and should explain any final recommendation to you clearly. Several agents producing answers that aren't the same. It's just going to make things complicated and burn a bunch of
[44:09] tokens. In this case, the short list for sofas has come back with a confirmed delivery and conditional options. And that's a super useful example of something that final comparison needs to preserve. So, for example, a low price on sofas doesn't help if nobody can confirm that the thing will actually ship. You can ask for that kind of distinction directly. You can say something like, "Split out the independent research. Give every part the same criteria. Bring me one combined answer. This is what's important to me." And the available number of agents and
[44:40] their scheduling will depend on the environment you're in. You don't need to manage every single sub agent yourself to enjoy the benefits. Now, I would avoid parallel agents when they are waiting on the same unresolved decision or all trying to edit the same file. That creates coordination work without any benefit at all. So, for example, if you're coding, you obviously want separate workspaces that can keep independent changes apart. This is a work tree approach, but ultimately a master agent will still need to review
[45:11] how all of those pieces fit together. I would ask for more agents when it gets you the combined answer sooner. If you have to reconcile all the answers yourself, give the main task that job before calling it finished, but then use the power of multiple agents to parallelize a lot of that work so you can get the answer quicker. One of the most important questions is whether you're getting something that people can actually use. Because sometimes the useful output is something a person can interact with, but if they need to search and filter and compare and choose
[45:42] and ask whether a site would serve them better than another long document, you're not really getting there. One of the things that helps in this situation is sites. It's actually a feature in chat GBT work. You can turn material from a task into a lightweight website or application. In the reading room example, an article's 29 curated book recommendations become a collection that readers can browse through and search and explore with their own goals and experience in mind. That's a direction
[46:13] that I would kind of lean in. You want to start with what the visitor or customer or end user needs, maybe it's you, and then help figure out how to get there. So, if you want them to be able to filter options, if you want them to keep their original notes nearby so they can read it, uh you want to make sure that you convey that so you're not giving a vague direction like make this a website, but instead giving enough specificity that it's useful for you or whoever is going to be working with it. The same idea applies to any kind of interactive visualization. If someone
[46:44] needs to explore how assumptions change an answer, the controls can be very very useful versus just saying produce a static site, right? Then they can interact and see what works and what doesn't. If they only need one recommendation, well, a flat output could be really appropriate in that situation. So think that through and ask for a preview before deciding how to share it. Try the controls yourself if there are controls. What happens when when you mess with it and something comes up that's an edge case and nothing matches your controls? Uh, can you use
[47:15] this site on your phone? Just do some quick checks to make sure it passes like ordinary common sense usage guidelines. Also, distinguish a working preview from a saved or published site. In the reading room job, the preview was checked and the site received a private deployment and those were separate steps because I wanted to check it first and then deploy privately. So ask the task ask AI which step it's completed and make sure you know where you are and then choose who should have access. Available sharing options depend on your
[47:45] account and your workspace but you want to review the audience before publishing or inviting people especially if the source material was private. If the information needs to stay current, you need to say where updates to that site are coming from. A site does not automatically become a live dashboard just because it looks like one. You may need a connected source and a separate scheduled task to run to refresh that site. A search box and a useful filter might be just enough so you can add the interaction that helps the reader make
[48:15] the choice. Now, how do you give a bigger job to the AI and still keep control? I would say use a goal, which is a specific command when the useful outcome needs several rounds of work. You want to have the task keep making progress, inspect what it's produced, correct problems, all before calling the assignment finished. Goal is a great tool for that. In supported desktop experiences, you can enable that by just entering /goal and describing the result. The progress controls will let you pause and resume and edit and clear
[48:46] the goal when you want to. On the web, you can use work and put the outcome and constraints directly into the prompt and then continue in that task. If you haven't yet decided what the outcome ought to be and realistically sometimes we haven't start with a plan supported desktop experiences offer slashplan you can ask it to identify the decisions ask a few useful questions and propose an approach before making any changes I love this when I'm planning a bigger piece of work it's very very useful the instruction needs a finish it can
[49:18] recognize if it's going to be in a goal form though so if you say work on this research that leaves any kind action open, right? If you say compare these options, identify unresolved evidence, and deliver the agreed report format, that's the kind of goal that gives the AI something it can complete and measure. And then while it's working, steer it. You can say, "Hey, the audience changed. Hey, the source you found there is wrong. Hey, one part matters more than you first thought."
[49:48] Whatever it is, keep the correction in that same task flow to help it use the context it already has. I personally find this tremendously useful. It's one of my favorite features. Where side chat is available, use it for a status question or an explanation while the main task keeps going. That's useful when you want to understand what is happening without turning the request into a completely different assignment. Now, if you're doing coding work in codeex, you want to open the review view to inspect the changes as well as the
[50:19] final message. So, you want to ask for checks that match the behavior being changed. You want to know what the code does and what was tested, not just the files that were edited. So, ask for those functional tests as part of the wrap up from your code request. A goal doesn't grant broader permissions or make a local computer available after it's shut down. So, it's not magic in that regard. It still uses the exact same access and the exact same execution environment. And sometimes it may need a
[50:49] decision from you. It's not that it overrides permissions. I would use goals for an outcome that deserves really sustained work with a clear stopping condition. In fact, that's how I do use them. I love them that way. If it's repetitive work, it's something that keeps going, I would use the scheduling feature. And the two can work together, but they answer different needs. So let's jump to scheduling. Can AI do this for me every Monday, every week, every hour, whatever it is. Yes, it absolutely can. Once a task produces something
[51:19] useful, you can schedule it to repeat. Just ask for the job and the cadence in really plain language and then open scheduled to check that that was saved and to see how previous runs ran. Do not schedule an unclear prompt and just pray that repetition will improve it. That's not how it works. Instead, run the job once, make sure you like what you get, and make sure you make any adjustments that matter, and then ask the schedule to repeat that final version of the
[51:50] work. The weekly briefing example I'm showing here was for a sales exec. The instruction focused on changes that affected prospecting, account research, meeting prep, and follow-up. Each item needed a source. That makes sense, right? A reason to care, and a very practical next step. That's what makes the briefing really useful. And so that's what we're emphasizing. Monday at 8 is just the last part of that. That's just when you need the briefing to arrive. Be precise about your time zone. If you mean 8:00 in the morning Pacific time, say so. Open the saved schedule
[52:22] and check the next run, the timing, and the instructions before you finalize. A message promising to do something later is not enough to establish that the recurring task exists in the rhythm and schedule that you expect. There's also a choice about context here. A schedule inside an existing chat returns to that chat and its ongoing work. A standalone scheduled task starts from the saved prompt for an independent run. Those are
[52:53] different things. So, make sure you're clear about which behavior you actually want. If you're doing continuing work, a continuing investigation, the existing chat may be a useful place for a scheduled task. But if you want a fresh report that should stand on its own every single week, I mean, I just use a standalone run. That may be cleaner, cleaner context window, more predictable results. Either way, keep the sources somewhere the scheduled task can easily reach. Otherwise, you're going to have issues. And just a reminder, check that
[53:23] execution environment again. Cloud schedules can run independently of your laptop. Local schedules need the computer available and the app running. A saved schedule can't use a local file on a machine that isn't there to provide it. That just stands to reason, right? And as I've said before, review the [snorts] first few runs. This is one of those larger principles in AI. Just review stuff. If the briefing is too broad, change the instructions. If nothing meaningful changes that often, reduce the frequency you're wasting tokens. When the work stops being
[53:54] useful, just pause the schedule. Recurring work deserves occasional attention as to whether it's still worth doing. It deserves optimization. Next question. What about watching for something and staying quiet? Does the scheduled run always have to produce tokens? Actually, there's a way to do this. You can tell a monitor agent what would cause an action and then have it watch. So, you can say, "Watch my inbox." But that's super vague. A better message would be, "Look for messages with this subject. Extract the supported
[54:24] dates and notify me only when there's something to do. Now, it has a job and it won't bother you unless something comes up." There are two ways this kind of checking can start. A scheduled check runs because the clock says it's time. An event triggered task runs because a supported event occurs. So on eligible plans, native Gmail, Slack, and GitHub triggers are available on the web and on mobile. They aren't the same feature as a desktop schedule. So I want to be precise about that. So for Gmail, you
[54:54] can narrow the incoming messages by sender or by subject. For Slack, supported triggers cover messages in particular channels with options like author and thread replies. Don't assume every reaction or direct message is going to trigger the same workflow. So, make sure when you're figuring out how to use these things that you choose a start signal that fits the actual task you want to accomplish. One task can have multiple supported event triggers, but it cannot combine those with a timebased schedule, at least not right
[55:25] now. If you're using a periodic desktop inbox check, you need to describe it as a check on a clock to make it work. The recorded school update monitor used that particular periodic approach. It found a matching email, but the body only said hello. There were no exact dates or actions to extract. So, it reported the missing information, handled the message once, and has stayed quiet since. That to me is a much more useful and token efficient behavior than inventing an
[55:55] event because the subject sounded important. Like, you don't want to do that. That just wastes everybody's time. So if you're going to do a event triggered notifier like this, you want to ask the agent to preserve missing information and recognize items that is already handled so it does the work it needs to do without inefficiencies that result from misprocessing of missing information. That's a much more useful behavior than inventing stuff because the email is empty. You want to ask the monitor to not invent stuff just because
[56:28] it sees missing information. So, as you're moving forward with this, as you're building these event triggers, make sure you start by trying a matching example and also trying a non-matching one. Check both so you're sure that you're getting the work you want inside the trigger, but you're not over selecting. Preparing a draft, as an example, is really different from sending a reply. So, keep that in mind when you're testing your triggers. Give your triggers an end condition as well. A monitor for a decision next month should not run forever. Say when it can
[56:58] stop. Say what deserves a notification and say what to do when nothing changes. The quiet periods are a huge part of the value because they save tokens and they don't bother you. All right. How do you hand work to someone else with chat GPT work? You got to decide first whether the other person needs the result or whether they need the conversation or whether they need the method you use to solve the problem. Those are really different things to share and choosing the right one saves you and them a ton of unnecessary work. If you need to see
[57:29] the finished file and share it, just give them the file in a place they can access and take responsibility for it. Make sure it's not slopp. If they need to use the interactive result, you can share the site with the intended audience and you can check the permissions from their side rather than assume a link is enough. So you can make sure they can access it. Now for a codeex task supported desktop sharing creates a readonly snapshot so you can open share choose the available access settings and review the shared view before sending the link. The snapshot
[58:01] doesn't grant access to your computer or project just that view. It also doesn't keep updating as you continue your original task. So you really created a branch later messages and changes are not adding to that existing snapshot. So, you want to share a new result when you need someone to see any later work you pursued against that thread. For personal account snapshots, anyone with the link can open that link up. Workspace sharing has different audience controls, and that's worth checking on
[58:32] before you share a task containing private material and images and file changes. If the other person needs to repeat the work or maybe needs a clear set of instructions or some kind of reusable skill and maybe that's the part that's actually more useful than seeing the snapshot of the conversation, then I would include the intended inputs, the output format, and the decisions the method depends on. And that may take some additional preparation from you. Ultimately, they still need their own authorized
[59:03] sources and connections to fully reproduce what you made. a method that worked on your particular file sets won't magically find those equivalent files in someone else's local machine. Instead, if you're testing a method, you want to ask them to test a small example with all the same files after setting it up to verify it works. And don't confuse importing a snapshot as context with taking over the original task. The recipient can use supported shared material in their own work, but the
[59:33] original task and its access remain separate. They cannot just look at the conversation you shared and magically keep it going in a way that works on your computer, too. In general, I would hand over the smallest thing that is useful to your counterparty. The answer when that's enough, the evidence when they need to actually look at it, and the method when they need to do the job again. And the reason for that is simple. It's easier to share something complete if it's small and manageable. Otherwise, you're going to be forgetting what you shared or sharing too much
[60:04] accidentally or sharing an incomplete amount of information. Ultimately, what makes chat GPT work worth the usage? Look at the actual usage controls for your account before you start a large project. Chat GPT work and codeex both share a single run allowance or single token budget. So changing between the two experiences does not create any fresh tokens or any fresh pool of usage. The amount a task consumes heavily depends on the model. It depends on reasoning. It depends on material process tools and how much work
[60:35] continues. I have seen instances where Astra is more efficient than soul. Even though Astra is the stronger model, a brief edit and a large research package are not comparable just because they began with one prompt. You want to look at the full package of the change when you're measuring your assessments. With the Astra release, Astra is now able to assess the current usage of your allowance and is able to respond to commands like, "Please burn through my usage on this problem until it reaches 25% and then stop." That didn't used to
[61:07] be possible, and that's a nice feature that's been added in the last few days. Now, if you're trying to plan a big run, always start by looking at your scope, and narrowing it down where you can name the sources, the number of options that you're comparing, the output you need. If it's a code project you're planning, look at the components and how you can break them apart. That's just good practice. If you already have enough information to make your decision, then don't ask for another broad search merely to make the result longer. If you already have enough information to go to step two or three or four in your process, then don't ask for a repeated
[61:39] version of step one. And this advice to avoid needless repetition absolutely is true for recurring processes as well. If you're starting a new recurring process, inspect what a useful run is going to consume from a token perspective before repeating it a lot. I've been in a position where I've been burning a 100red million tokens a day on repeated runs that I did not fully intend to repeat every day and I had to clean it up. You don't want to get into that position. The schedule you set for a recurring task can make an inefficient instruction much more damaging. So you
[62:09] want to set the frequency according to when the information will actually change and then act on it. Voice is different. Voice has its own plan dependent allowance while the tasks that you start through voice also consume the work allowance. I know that's complicated. Computer history will use tokens to summarize activity. Features that operate in the background can still use capacity. So you have some hidden shadow capacity you have to plan for. I would also include review time in any assessment I make of how much usage you're going to have. A lighter model
[62:39] may be the right choice for a stable routine, but if it leaves me correcting the same mistakes, the apparent saving is less attractive because it burns more of my time and it frankly may burn more tokens over time, too. It's really about tokens per task, not tokens per conversation. And you can test a lot of this without moving all of your work. You can try a familiar assignment at a lighter reasoning setting and compare the results. You can keep the stronger model when the judgment is actually worth it. Check your reset times and any credit or spending controls before
[63:10] assuming the task can continue indefinitely. If you sign in with an API key, billing will follow the API arrangement rather than your chat GPT subscription. Use the current information for the way you're signed in and make sure you know what it is. I have been the person that went to bed with a $300 budget allowance for a task I was working on and all of it got burned overnight and the task still didn't get done. I was not happy. Okay. How do you stop chat GPT work or fix a bad result? If a task seems stuck, you want to check whether it's working, you
[63:41] want to wait for a decision and you want to report a failure. Those call for really different responses. So, we're going to go through each of them. A quiet interface by the way does not necessarily mean nothing is happening or nothing bad is happening. So if you're not sure what's going on the first thing I would do is ask chat GP2 worker codeex for a status. What is going on? What is complete? What is left? What needs you? If it's waiting on an approval, inspect that specific approval request and figure out if you want to approve it. If it has the wrong file or it has misunderstood the outcome, you need to
[64:12] correct that before telling it to continue. If it's looking like a bad result, point to the real problem and be as specific as possible. This total doesn't match the source is really useful. This paragraph changes the promise that I made is also useful. If you're just saying this is bad, that's less useful. Tell it what you don't want it to change so that it only touches the part that it needs to change. Now, if an external action has failed or if an external action has an unclear status, check what is going on outside codeex or
[64:44] chat GP2 work before you retry. An event or file or message may already exist even if the task didn't receive a clean confirmation because of a system integration issue. So, repeating the whole action in that situation can actually create downstream problems. One way to handle all of this if you're just not sure what's going on is to stop the current run. If you want to pause, you can pause a goal and resume later. And you can also schedule runs into the future if you just can't do work now. Disconnecting an app will remove a
[65:14] connection, but not reverse the changes that were made through that connection. So, previous work isn't affected. Whatever you're choosing to control, choose the kind of stop that you actually need to stop what is really breaking for you. Choose the kind of control that matches what you want to actually stop. Now to recover, use the mechanism that belongs to what you're looking to recover. Right? Uh you want to document revision history, a saved copy or version control for code. And then you want to ask the task to bring
[65:46] back that particular earlier state in a way that does not start undoing everything else. And so be as clear as you can about your versioning even if it's docs and not code. And then ask for that specifically. One more instruction it's worth calling out here. Material that chat GPT reads supplies evidence but not permission. However, a web page or attachment can contain instructions that chat GPT can read and interpret as instructions. Those should not override
[66:17] your request or authorize a different action. And you want to be sure that you are sending chatgpt work and codecs good content that would not be at risk of overwriting your instructions if it was read. This is how prompt injection attacks occur. And a lot of it comes down to having the system ingest work that could be read as a prompt that is really supposed to be context. So be careful what you tell your system to ingest. All right, you want to get started. Choose a job that you
[66:49] understand well enough to have a strong opinion about it. Put it in the right environment in chat GP2 work or codeex. Give it sources that actually make sense and ask for something that if you could use it, you'd be delighted with it. And then make it work, right? Correct it where it needs correcting. Edit it as you go. Save the method that works. Tell it when to stop if you need to stop. And start over as you need to start over. Ultimately, what I want you to do is start with the feature that would take work off your plate today. You can add
[67:20] everything else once you start to get a feel for it. I know I talk about a lot of advanced technical stuff on this AI channel. I want there to be space for folks who are just learning AI as well. If that's you or if that's someone you know, I have an entire playlist full of initial zero to AI videos and it's right here on this channel and you can just go and grab that and you can share it with folks in your life who just want to watch through a YouTube playlist and actually learn about the basics of AI. That's why I created it. I hope you have fun. Share it with someone who needs it and I'll see you there.