Full transcript
0:00You pay for an artificial intelligence
0:02subscription every month , but you end
0:04up just using it to ask questions and
0:06get answers , right ? Like Google , only
0:09smarter . So , this is the costliest
0:13mistake people using AI are making
0:15right now , because the same tool for
0:18the same price can do two completely
0:20different things . One version answers
0:24you inside a little chat box . The other
0:27opens your computer , analyzes files ,
0:29reads your emails , builds and finishes
0:32a task on its own that would have taken
0:34you an entire afternoon , while you can
0:37just sit back and have a coffee . The
0:40difference between these two worlds
0:42isn't the tool ; it's a key that almost
0:44nobody knows exists . And in this video ,
0:49I will show you exactly how to turn
0:51that key , which four tools are actually
0:53worth it today , and most importantly ,
0:56the right way to provide prompts to
0:58them . Because if you provide the prompt
1:02in the old way , you are throwing away
1:04time and money . By the end , you will
1:07leave here knowing how to put an agent
1:09to work for you for real , and not just
1:11chat . My name is Tainara Shmite and
1:14welcome to the Negócios em Mente
1:16channel . Here you learn everything you
1:19need about artificial intelligence .
1:21Before we get into the practice , I need
1:23to explain one thing to you . If you
1:26skip this part , the rest of the video
1:28won't make sense . So stay here with me
1:30because it will be quick . And look , I'm
1:32telling you this because it took me a
1:34long time to understand this myself . In
1:37the beginning , I used these tools like
1:39a normal chat and kept thinking it was
1:41all the same thing . Well , it turns out
1:44I learned the hard way that it isn't .
1:46The structure is simple . First , I'll
1:49explain the difference between a
1:51so-called chatbot and an agent . Then ,
1:53I'll show you the four tools and which
1:56areas each one is best in . And at the
1:58end comes the secret trick . I'll show
2:01you the formula to create prompts that
2:03make the agent deliver exactly what you
2:05want . So let's go . A chatbot and an
2:09agent can use the same brain inside .
2:12The engine can be the same . What
2:14changes is what is connected around
2:17that brain . Think of a company . A
2:20chatbot is like a very intelligent
2:22attendant , but one who stays stuck
2:24behind a counter . It cannot enter the
2:28company's system , it cannot open files ,
2:31it cannot send any emails , it cannot
2:33access a website , and much less
2:35continue a task after you have left . It
2:39chats , but does not execute . Now , an
2:42agent is like that same attendant , with
2:45the same knowledge . Get instant access
2:49to your company computer , including
2:51systems , files , your browser , calendar ,
2:54and a clear list of everything that
2:56needs to be done . And that , my friends ,
2:59changes everything . Instead of just
3:03saying how to do it , it can actually do
3:06it : research , compare info , fill out a
3:09spreadsheet , organize files , create
3:12reports , send messages , test different
3:15paths , and keep trying until the final
3:18result is reached . If you're looking
3:23for AI to get more done in less time ,
3:24then you have to check out Monday . Even
3:28before AI appeared , Monday was already
3:30one of the best productivity tools on
3:33the market , used by companies of all
3:35sizes , freelancers , or even people just
3:38wanting to be more organized and
3:40productive . Now , with its new AI
3:44features , it has reached another level .
3:47Beyond helping you organize your work
3:50and life , it's ideal for those who want
3:52to get projects off the ground and
3:54reach goals with much more agility . And
3:57best of all , there's an excellent free
4:00plan for you to test all the features .
4:02Click the link in the description and
4:04try it out now . For this agent to work ,
4:09it needs four things . First , of course ,
4:13the brain , which is the AI itself , the
4:15part that understands , reasons , and
4:17makes decisions . Second , the tools ,
4:22which are the access points that allow
4:24it to act in the real world , like
4:26opening a browser , editing files ,
4:28running commands , or using apps . Third ,
4:33the memory , which acts as a history or
4:35a context notebook so you don't have to
4:38keep asking the same things . And fourth
4:41, the goal . A clear task with an
4:44expected result . Not something vague
4:48like " help me , " but a mission with a
4:50beginning , middle , and end . When you
4:54put all this together — brain , tools ,
4:57memory , goals — and run it in a cycle ,
4:59you stop having a common chatbot that
5:02just answers questions . Bringing those
5:06four things together , running in a loop
5:08, and there you go : you have an agent .
5:10And this cycle is quite simple , just
5:13three steps . Look how easy it is . It
5:17observes what is happening , such as
5:19which files exist or what appeared on
5:21the screen . Then it thinks of the next
5:25step based on what is true at that
5:28moment , then it acts , uses a tool ,
5:31changes something , and repeats
5:34everything until the goal is met .
5:37Observe , think , and act . This is the
5:41engine behind every agent anywhere in
5:43the world . Keep this information in
5:46mind because now I’m going to tell
5:48you about the tools . But first , go
5:51ahead and like this video , subscribe to
5:53the channel so you can stay up to date ,
5:55leave a comment below , tell me what
5:57you're thinking , and most importantly ,
6:00tell me a task you'd like to perform
6:02with an agent . Before I tell you about
6:05the four tools , I want to show you the
6:07scale of this in practice . Come with me
6:10to the screen . Look , this folder here
6:13is a real mess . It's got screenshots ,
6:16PDFs , receipts , photos , travel plans ,
6:19study PDFs for different languages ,
6:21weird file names , really just
6:23everything thrown in . And I bet your
6:26downloads folder is like this too . Now ,
6:29watch what I'm going to do . I opened
6:32Claude Code , selected the messy folder ,
6:35and here I’ll give it a command
6:37written in normal language , as if I
6:39were asking a person . You'll see that
6:42it's nothing complex . I asked it the
6:44following : open and read all the files
6:46in this folder . Separate the folders by
6:50category , rename each file with the
6:52date and type , and finally , create a
6:54spreadsheet listing everything . The
6:58spreadsheet must be in XLS format with
7:00a professional design , formatted
7:02headers , adjusted columns , and a
7:04summary tab . Do not delete any files .
7:08And now just watch . Done . In a few
7:12minutes , it informed me of the
7:13structure it used for the folder and
7:15created a spreadsheet listing every
7:17item and detail on its own . And by
7:20opening the folder , we can see the
7:22separation of the categories . I opened
7:25a few folders here so you can see which
7:27files are inside . And notice that it
7:30changed the names following the
7:32formatting requested in the prompt , in
7:34the command . And for sure , what I liked
7:37most was the spreadsheet , which looks
7:40beautiful , super organized , with
7:43category , new name , original name , type
7:46, date , the folder it was added to , and
7:50the description , plus a good summary
7:53tab . Notice the difference . An AI
7:56chatbot teaches you how to organize
7:59that folder , how to create that
8:01spreadsheet , all step-by-step , like a
8:03recipe , but it's for you to do manually
8:06. The agent simply took that task and
8:10said , " Alright , I'll do it . " And it did
8:12. And look , this command I provided
8:15wasn't luck , no . There's a formula
8:17behind it , which is exactly what I'm
8:19going to give you later in this video .
8:21And when you learn it , I'm sure you
8:24will never use commands the old way
8:26again . Well , let's talk now about the
8:29four agent tools that are worth it . But
8:33first , if you want to get ahead and get
8:35the latest news on AI before everyone
8:38else , then I recommend you join our
8:40free WhatsApp group . And look , don't
8:43worry , because there are no " good
8:44morning " messages there , or people
8:46sending stickers . It’s just news ,
8:48actually hand-picked by our team . To
8:51join , just point your phone at this QR
8:53code appearing on the screen or click
8:55the link in the description and the
8:57pinned comment . There you’ll also
9:00find our free materials , like courses ,
9:03prompts , spreadsheets , plus discounts
9:05on tools we recommend . So , if I were
9:08you , I’d take a look . Well , let's
9:10talk about the tools . There are a ton
9:12of agent tools out there , but the vast
9:15majority are just noise . After testing
9:18several of them , four remain that truly
9:20deliver good results . And what’s
9:23interesting is that each one is very
9:25good at something different . And I'll
9:28go through each one quickly , because I
9:30know how valuable your time is . By the
9:32way , if you want a video dedicated
9:34solely to any of them , just leave it
9:35here in the comments . The first one is
9:38Claude Code , which is Anthropic’s
9:40agent , the one I used to show you the
9:42example of organizing my folder . It
9:45runs as a program that you install on
9:47your computer . It doesn't matter , it
9:49runs on your computer , so you'll
9:51download it . It can be Windows or Mac ;
9:54it works on both . The name has " Code , "
9:57so many people think it’s just for
9:59programmers or that it’s very complex
10:01, but it isn’t . It can rename files ,
10:04read PDFs , extract data from
10:06screenshots , edit things on your
10:08computer , in short , anything you can
10:10describe in plain language . The limit
10:14is your imagination , not the tool . The
10:17good thing about Claude Code here is
10:20that you have a much wider range of
10:22things that it can execute . About the
10:25price , pay attention here because many
10:27people get it wrong , okay ? The free
10:29version of Claude does not include Code
10:32, so you need to have a paid plan .
10:35However , the same subscription covers
10:37the regular chat , which is what many
10:38people use , and the agent as well . So
10:40it’s all bundled together . Plus , the
10:42paid version also includes Claude
10:44Workspaces , which is another very cool
10:45tool . And where does it stand out ? It
10:48lets you see its reasoning . So you can
10:51follow the AI step-by-step and you can
10:54interrupt it at any moment to correct
10:56and redirect it mid-path to the right
10:59direction . It's ideal when the task is
11:03complex , you want to follow it closely ,
11:05and you need a thought partner . And why
11:09is it good to stop and adjust ? Because
11:11folks , it consumes tokens , right ? Which
11:13is like the gasoline it needs . And then
11:16you might exceed your daily , weekly , or
11:19monthly usage limit . The second tool is
11:22Codex , the agent from OpenAI , which is
11:24the same company behind ChatGPT . So , if
11:28you already pay for ChatGPT , Codex is
11:30already included in your plan . Same
11:33account , same login , no new
11:35subscription , no new tool to learn ,
11:37just install and use it . And it has a
11:41trick that others don't : a cloud
11:43version . So you submit a time-consuming
11:46task , close your laptop , go about your
11:48life , do other things , and when you
11:50return , the work is ready for you to
11:52review . So it's useful for those who
11:55don't want to sit in front of the
11:57computer or keep it on while the AI
11:59works . So you can delegate and come
12:02back when it's ready . In short , if you
12:04already use ChatGPT and want to start
12:07with as little friction as possible ,
12:09Codex is the best choice . And actually ,
12:11look , I opened it here just to show you
12:14, Codex is also available for your
12:16phone . So , you can continue your
12:17project from where you left off . If you
12:20started on the computer , you can
12:22continue on your phone , you stay in the
12:24loop , so you receive notifications when
12:26it completes a task or needs your
12:27attention for something . And you can
12:30start something new as well , initiate a
12:33task on your phone and continue later
12:35here on the computer . It's going fast ,
12:37isn't it ? Because we’ve already
12:39reached tool number three , which is
12:41Open Claw , a real-life automation . And
12:44what makes it special ? It doesn't stay
12:46trapped in a website or program . It
12:48lives inside your messaging apps . So
12:51this is very useful . Where does it live
12:53? Oh , Telegram , WhatsApp , and so on .
12:56Besides that , another advantage is that
12:58it is free and open-source . So you
13:00install it on your own machine . So
13:03think about it . You send a message to
13:05your assistant just like you would to a
13:07friend . The work is done on your
13:09computer at home , and it replies when
13:12it's finished . You could be in the
13:14grocery store line and send a task from
13:16your phone . So , real-life use cases :
13:19summarizing emails and creating
13:21reminders , building documents and
13:23organizing ideas , saving video ideas
13:26sent via message , organizing notes
13:28automatically . This is the agent that
13:32finally delivered what we were promised
13:35: a real assistant that lives inside
13:37the apps you already use . In fact ,
13:41Rodrigo already made a video talking
13:43specifically about Open Claw . I’ll
13:45leave the video linked in the cards
13:46here , and it will also be in the
13:47description below . And our fourth tool
13:50is Antigravity , Google’s agent for
13:52visual tasks . It is Google’s agent
13:56platform built on top of Gemini . Right
13:59now , it’s in open beta , so it’s
14:01free for any individual without needing
14:03a credit card . I actually
14:06double-checked before recording this
14:08video , but depending on the date
14:09you’re watching , that might have
14:11changed , but for now , it's free . And
14:14what are the advantages ? Gemini can
14:16actually see what it’s doing . So it
14:19looks at a screenshot , compares two
14:21designs , and generates images with real
14:24visual context . So , without a doubt ,
14:27it’s the best for creating screens ,
14:29making layout adjustments , and working
14:30with images . Ideal for those who work
14:33in design and marketing . So , if your
14:36work involves design , marketing , or
14:38anything visual , this is your agent . It
14:41shines where others are left in the
14:43dark . Do you see how you can even use
14:45more than one , how they complement each
14:47other ? That’s not a problem either ,
14:49but I’ll talk more about that soon .
14:51So , here’s a quick summary for you .
14:53Which one to choose ? Choose the one
14:55that fits your daily life best and dive
14:57deep into just that one . You don't need
14:59to learn all four at once . This is
15:01great for you to start getting
15:03comfortable with agents , then you can
15:05start adding others . So , Claude Code is
15:08better for following reasoning closely
15:10and using it for complex tasks . Codex
15:13is for those who already pay for
15:14ChatGPT , which is the easiest path .
15:16Open Claw is for those who want to
15:18automate real life using message
15:21windows . And Antigravity , for visual
15:23work , design , and marketing . Every tool
15:25has its strong point . Identify what
15:28fits your routine best and start with
15:31that . Depth also comes with use , folks .
15:35Now , pay full attention here , because
15:37nobody is telling you this . You can
15:40choose the best tool in the world , but
15:42if you give the wrong prompt , it will
15:44deliver something terrible . And what
15:48ruins results with an agent most is
15:49exactly that — the way you give the
15:51command . For a chatbot , the command is
15:54a description of what you want . But for
15:57an agent , the prompt has to be almost
15:59like a contract , a simplified one , of
16:01course . But you get the analogy . A
16:03briefing of what it is required to
16:06fulfill cannot be short and vague . Why ?
16:09Because the agent truly has autonomy in
16:11its hands . With a vague sentence , it
16:13will start to make things up . It will
16:16choose a color you didn't ask for ,
16:17create a section you don't want , and 10
16:20minutes later hand you something that
16:22wasn't what you imagined . And all of
16:24this while spending resources , spending
16:26tokens , for example . The solution isn't
16:28to write a longer prompt , it's to write
16:31a structured prompt . And it only has
16:34four parts , so it's very easy . It’s
16:36even easy for you to take notes , which
16:38I highly recommend . So , memorize these
16:41four words , because every prompt you
16:43give an agent for the rest of your life
16:46will use these four things : goal ,
16:48limits , format , and output . Let's go
16:52one by one . But hey , if you haven't
16:54liked it yet , please , I'll ask you to
16:56like this video , subscribe to the
16:58channel , talk to me in the comments ,
17:00and let's see part one , starting with
17:02the goal . The goal , folks , is different
17:05from the action , okay ? So , the key
17:07question here is the following : for you
17:09to be able to define your goal . What
17:11does the finished work look like ? Again
17:14, the goal is not the action , it's the
17:16final result . It will define the finish
17:19line for the agent . So , what would be a
17:21bad example ? Create a page . This is an
17:24action . How to turn this into a result ,
17:27right ? For you to add to the goal
17:29section . Create a single-column page
17:31with an email capture at the top , ready
17:33for me to review before publishing . The
17:36second sentence shows the agent how it
17:38knows when it has finished the task .
17:41And if you don't give it a finish line ,
17:43the agent will invent one . And once
17:45again , you probably won't like it . So ,
17:47remember , the goal answers the question
17:50: What does the finished work look like
17:52? Part two , limits . And these are the
17:56safety rails . The limits are everything
17:58the agent cannot do . This part here is
18:01what avoids disasters . There is a very
18:03simple analogy for you to understand
18:05better . Think of a child playing near
18:07the stairs . You put up a gate , not to
18:10stop them from playing , but so they
18:11don't fall . So the limit is that gate .
18:15So examples of limits : do not touch any
18:18file outside of this folder . Do not
18:21publish anything . Just let me review .
18:24Don't copy text from a competitor's
18:26site or delete files from this folder ,
18:29which is what I used in that example I
18:31showed you . Your list of constraints is
18:35your learning scar . It only grows , and
18:38each one saves you a headache in the
18:40future . So , ah , you've created your
18:43constraints there , and you're talking
18:45to your agent , it's performing the task
18:47, and then you see it did something
18:49wrong again . That will become a
18:51constraint , and you'll keep updating
18:53your list of constraints . Part three is
18:56the format . The format is the exact way
18:59you want to receive the result . This is
19:01where most prompts fall apart
19:03completely . Either the agent does the
19:06work right , but the delivery is in a
19:08format you can't use . So say exactly
19:11how you want to receive it , for example
19:13, in bullet points , in a table , in a
19:15single file , or in a folder with a
19:17specific name . In your prompt , you need
19:19to remember to be specific . The work
19:22can be correct and still be useless to
19:24you if , for example , it's in the wrong
19:27format . So , if you can't describe your
19:29format in one sentence , you still don't
19:32know what you want . Therefore , you're
19:34not ready to hit enter yet . You need to
19:36think a little bit about how you want
19:38this delivery . And you can use a
19:40regular chat to help you choose the
19:43best format for that task . And that way
19:46you don't spend tokens and resources
19:48with the agent . With the agent ,
19:50everything has to come ready . And
19:51finally , our part four is the exit ,
19:53which is what it should do when it gets
19:56stuck . Almost no one writes this part
19:59here , but it’s the one that saves the
20:01most tokens and time , okay ? It defines
20:03what the agent should do when it gets
20:06stuck or when information is missing .
20:08Without instructions , the agent chooses
20:11the worst possible outcome , okay ? I
20:13swear it’s like this : it keeps trying
20:15, it keeps running in cycles — in
20:17circles , actually — and it keeps
20:19spending resources , trying , " I'll try
20:21this way , that way , this way . " When it
20:23could simply ask you , right ? So , one
20:26sentence solves it . You can add . If any
20:29important information is missing , stop
20:31and ask me before continuing . Have you
20:34defined how it will handle a doubt ? So ,
20:37with the exit defined , the agent stops
20:39and asks for help . Without it , it tries
20:42to guess and enters this infinite loop ,
20:44wasting your tokens . Either you define
20:47the output , or we define it for you .
20:48And usually , it decides the most
20:50expensive way , right ? So , what does
20:52this complete formula look like ? Our
20:55contract with the agent for objective ,
20:57limits , format , and output . For example
20:59, the objective would be all of this ; I
21:02would combine it into one prompt , okay ?
21:04I can add the objective and then write
21:06it next to it , all in the same prompt .
21:09And then skip a line for limits , then
21:11format , and then output , or I can just
21:13send the white text without the headers
21:15. All at once . So , the prompt would
21:19look like this : create a one-column
21:21landing page for my product launch ,
21:22focused on capturing visitor emails
21:24right at the top , ready for me to
21:26review before publishing . Use only one
21:28file , don't touch anything outside the
21:30page folder , and don't publish it
21:32anywhere . The page folder should have
21:35the page file plus a short summary ,
21:37listing each section that was created .
21:41If you don't know who the product is
21:43for or what its unique selling point is
21:45, stop and ask me one question instead
21:47of guessing . So , this right here is a
21:50complete prompt . You define the result ,
21:52the guardrails , the delivery format ,
21:54and what to do when information is
21:56missing . Adding to this , what would I
21:58do in this real case ? Obviously , I
22:00would feed the agent information about
22:03my product . So , a document about the
22:06product , including the differentiators ,
22:08the marketing terms we're using , and
22:10everything else . Test this out in your
22:13commands before you hit enter . Write
22:16out all four parts . The first time
22:19you'll feel like : " Wow , this is extra
22:21work that seems unnecessary . " By the
22:24third time , you'll be wondering how you
22:26could ever send commands any other way .
22:29So remember , for a chatbot , you use a
22:32description ; for an agent , you use a
22:34contract . Now , let me tell you where
22:38most people give up on using agents .
22:42They ask for something , the agent
22:43messes up , they correct it , then they
22:45ask for something else and it repeats
22:47the same mistake again . And again , and
22:50again . And then the person thinks the
22:52agent is broken , that it's bad , that
22:53it's not worth it , and they just give
22:55up . But it's not broken . It's you who
22:58forgot to tell it to remember that
23:00information . Remember that memory
23:03notebook for context I mentioned ? It's
23:05this right here . Every serious tool has
23:08a text file that the agent reads before
23:11starting any task . Everything written
23:14there becomes a rule it will always
23:16follow . You place this file in your
23:19project folder and , voilà , you've just
23:22given your agent long-term memory . So
23:25you need to ask it to create this
23:26little notebook of memories . You need
23:28to ask it to remember this , that , and
23:30the other . And as the interactions
23:32happen , there will be more things it
23:34needs to remember . I'll give you a
23:35simple example . Imagine you're creating
23:38something serious and professional , and
23:40the agent keeps putting emojis in all
23:42the text ; you remove them , and next
23:44time it puts them back . It's tiring ,
23:47right ? At least I think so . But then
23:50you say just once : " Create a rules file
23:52and take note . Never use emojis in this
23:56project . It is professional work .
23:58Remember this in all our future
24:00conversations for this project . " Look
24:02how simple , look how wonderful , guys .
24:04It takes less than a minute and done ,
24:06it will create it , save it , and that
24:08problem will never return . Either in
24:10that project , or if you want it to be a
24:12general rule , it won't return in any
24:14project . You can set that too . Look ,
24:16I'm excited today , so I'm going to give
24:18you one more trick . You can ask the
24:21agent to update its memory on its own .
24:25You say : " Every time I correct you , you
24:27must note this correction as a new rule
24:29at the end of the file . " Again . Quite
24:32simple , right ? Then , look what happens .
24:34The first time you have one rule , but
24:37by the fifth time , you already have
24:39about 20 rules noted down . And a few
24:41days later , there comes a moment when
24:43the agent almost never makes mistakes
24:45anymore , because you created , see , a
24:47perfect customization for your case ,
24:49almost as if it could read your mind .
24:52It's like a new employee who learns the
24:54company's way of working and never
24:56forgets it again . And so you don't
24:59leave here with just theory , here is an
25:02action plan for today . You will choose
25:05one of the four tools , select a boring
25:08task , that task that is truly boring ,
25:10not a test for fun , okay ? Something
25:13boring that you do there every week .
25:16You will write the command in the
25:18contract format with those four
25:19elements that I hope you wrote down . So
25:21, if you wrote it down , it's already
25:23there for easy access . Otherwise , you
25:25can just rewind this video , ask it to
25:27create a memory file , and then you add
25:29at least three little rules and run it
25:32from beginning to end . And later , if
25:34you feel comfortable , come back here
25:36and let me know in the comments how
25:38your experience went . Because the truth
25:40is this : you can watch 1,000 videos
25:43about AI agents , and I love that you're
25:45here watching mine , but the real
25:47breakthrough only happens when you open
25:49the tool and execute . You only learn
25:53when you actually do it in practice . So
25:55stop reading and watching videos and
25:57start executing . And then you can do
25:59things in parallel . Oh , but now I want
26:00to learn this other thing . And of
26:02course , folks , keep improving , keep
26:03watching more videos , but remember to
26:05always keep practicing as well . And if
26:07you choose to start with Claude Code ,
26:10there are seven secret commands that
26:12are easy to apply and will help you a
26:14lot , including saving tokens . And I
26:17showed them to you in this video
26:18appearing right here .