so today we'll do something a little bit different it's a new format we're trying right now which is a sort of live build so instead of going through a presentation or talking about use cases of a specific function in mind Studio we're just going to build something together I will show you what we're trying to mimic before we start to build it but the idea is that I will build it you can ask any questions during the build you can ask any questions at the end of the build and let's try to actually all understand how to build this sort of workflow in M Studio so let me share my screen very quick claim okay so this is what we're trying to build it's a more complicated version of what we're trying to build we will actually focus on something a little bit simpler today but this is a Content writer that we built for one of the people that are basically partnering with my studio and the idea is having a user input block here where we ask two things we ask for the keyword and we ask for the human touch so sort of something that the AI doesn't know about that you want to infuse in the content as the human and that's very important when you build any sort of content writer because content writing can be a bit um stale can be a little bit boring if you at dii do all of the work all the heavy lifting if you do provide some sort of context or if you do provide some sort of personality to the AI then it will behave a bit better and brighter better copy so that's why we have the human touch second we have the Deep scrape so this is a custom function actually I can share the custom function with you if you're interested to do it um this does consume a lot of tokens though so I will give you the cheaper way to do it and a more expensive way to do it which is to consume a lot of tokens this scrapes the top 10 content pieces on Google for that particular keyword which we entered in Step One it takes the content and it saves the content in bright content or results now again as this consumes a lot of tokens because you can imagine 2,000 words per article for every ranking results on Google they're talking about a lot of tokens about 30,000 40,000 tokens so what we try to do in the Second Step here is to take all of the content from the Google Surf and we try to summarize it to to make it a little bit more acccent still give the AI all of the information it needs to create a Content piece that's unique and interesting and follow the same guidelines as the top ranking articles but doesn't look too copycat it doesn't need to have the ENT context of the other articles and then we do some processing so the first part of the processing is ignore the list of keywords I want the full context here so this is the content in markdown from the Google Ser search engine result page then we're getting all the keywords in the top Ser so we want to understand what are the competitors talking about in Google what are they talking about for a specific keyword let's say that the keyword is dog we might find that most of the topic and most of the content pieces on dog also include the breeds of dogs in a particular country or the type of different dogs that you can have in apartments in different kinds of buildings so all of those things are things we want to know because those are the topics that we also want to cover in our piece then we create a data source this is mostly for titles so we query a particular data source we have here for the best titles this one where we have the title what makes the title good full reasoning on on what makes the title good and the scoring the scoring for clarity curiosity benefit driven urgency emotion and uniqueness finally we generate the outline and then we pass the out we generate the outline with gp4 turbo which is the best model for reasoning right now it's debatable you can use different models as well it's not necessary to use gp4 turbo but gp4 turbo is usually the best model at reasoning so it's the best model at actually outlining what the content should talk about coming from the scraping of the Google search result coming from the list of keyword they talk about coming from all the summaries Etc I'm asking now gp4 turbo to take all of that into consideration and write a great outline including talking points and then the talking points are passed to CLA and clonet is writing the actual piece of content this is just one of the ways you can build a Content Builder content writer in mind Studio there are all sorts of different things you can do like you can make it much much easier and just have the prompt saying write a Content piece on x I need to write it you can have just scraping Google without having the content of the pieces you can just have the titles of the different content pieces on Google or you can ask another external API like we do here we have a AI keyword research module where we are asking for a keyword research this keyword research is happening outside of Mind studio and then we're having the result in mind studio and using that for the content writer that being said probably very confusing right now if you never built something in M Studio or if you never built a Content writer in particular so the idea now is to rebuild it together we are going to skip the custom function right now we are going to use a custom function that you have in the community functions but if you're interested in this particular custom function which again takes all of the content on Google so for a particular keyword it finds the top 10 links ranking for that keyword and then it extracts the full content of those 10 if you are interested in this particular function I can give you the link to this app and you can remix it in your workspace okay so I'll just take a look at the chat very quickly in case there are quick questions okay Rob wants a de scrape function yeah no problem I can send it later okay so let's create a new AI when you create a new AI you know that you can start from one of our templates we do have a Content writing template which is the blog post generator this is a great starting point if You' never built a blog post generator in the past this is probably what you should start with and let's start from this one actually because it's what you should probably do but another option if you already know what you're doing if you've done it already a few times and you want something a little bit more custom you can start with generate prompt or you can start with blank a but let's start with the block post generator even if you start with a template you don't need to stick with that template you can literally delete all of the inputs all of the blocks and you can start from scratch okay so this is the prompt it says SE optimize block post generator this is the assistant role you want the content to be SEO optimized so optimize for ranking on Google search results or Bing but you know most people use Google tone and style for the particular content the output the response formatting imagery and links and call to action so right out of the bat one thing that I would remove from this template in my personal experience thei is not very good at suggesting images so if you are writing about something that's very very generic let's say you're writing about dogs then sure the AI will say okay in this particular part of the blog post you should include an image of a dog that's obvious you don't really need an AI to tell you that but if you're building something a little bit more custom like how to how to build this content writer for example I say that I'm writing this as a blog post for M Studio how to build a Content writer in mind studio thei is not going to be very good at suggesting what to include in terms of imagery which in this example would be screenshot of the tool right it to be screenshots of this particular section it will be screenshots of the automation stab it will be screenshots of different parts of the application it will not be able to suggest it because it doesn't understand what my studio is and it will not be able to do it on a consistent basis so I would actually delete this part the image and links and the call to action part is also debatable let's remove that as well just because the call to action is probably very specific to the particular topic we're writing about and we don't want it in the system message which is the same for the entire workflow Philip also wants to cust some function yeah that's fine this is what the flow looks like the menu we have the start block the menu block and then we have the the link to the jump the jump redirects you from the menu from that particular workflow to another workflow in mind Studio we are dirting to the blog post flow in this example and blog ideas here so the menu is redirect to I want to write a blog post R direct two I want to write a blog post top I want blog post topic ideas just to show you what this looks like let's go in preview Draft when we start a new Trad you have the entire on boarding which explains what the template is about we're going to delete all of that and then you have the menu so what do you need a blog post a list of topic ideas if I click on a list of blog post topic ideas I can say what should the blog post be about let's say the blog post is about dogs explain the talking points we want to talk about breeds we want to talk about size we want to talk about um personality maybe and these are some ideas on blog post I could write you can always see what's happening by opening the debugger so if you click on this question mark icon and then show debugger you will see what's happening here on the right we have all of our inputs we have the programmatic message and write a list of blog post topic about dogs and this is the response from GPT 3.5 turbo so I think this is fine we're not going to touch this workflow too much I think this works the only thing I would do here in the block post ideas workflow is to change the model right gbt 3.5 turbo is not the best model anymore at pretty much anything you can think about so if you need a fast model CLA Trio is faster if you need a very high intelligence model gp4 TBO is more intelligent than GPT 3.5 if you need something that's more creative clonet is more creative than GPT 3.5 so likely anything you want to do GPT 3.5 turbo is now very unlikely to be the model for you it was in the past it was good on a lot of things it was what started the revolution in a way but it's just not the best anymore so I would go here in model settings you can switch it very easily to CL three IU the temperature at 0.5 is fine the maximum response size we can bump that up to 496 so that's all for the content ideas pce now let's delete the onboarding as well so as as SE before this is on boarding and the on boarding are just a series of display text and the display text is letting you know what this template is about we don't need that so you can click on the first item you can hold shift on your keyboard click on the last items and you can bulk select them and then you can click on the minus icon to delete them you can also add inputs in bulk in case you didn't know it's a nice shortcut so if you want to add from talking points to work count you select talking points hold shift click on work count and now you have the four input in in the block you can do this in on boarding and you can do this in the user input field as well so if you go here in blog post automations user input you can also do that the same thing you can add inl you can remove inul one cool thing you can also do for shortcuts purposes is click on space and if you click on Space it will always Center the canvas so wherever you are in the canvas you can click on Space you go back to the center and it also zooms to 100% on your screen so here we are asking for the the topic the keywords the talking point call to action bandw word count and tone so from this list I would probably remove the bandw just because most AIS are just not able to not use those words one thing you can say to chpt or to other models is don't use delve into don't say um let's kick it off those sort of things because they use it very very often and it sounds aiish so you don't want the content to necessarily feel like it's being written by an AI and you might want to ban those words and that's the the reasoning behind this B wordss this will work if you use a high intelligence model so if you use gp4 turbo or if you use CL 3 Opus even CL Sonet this will work if you use CL IU or if you use T 3.5 or Gemini Pro I'm not sure that would work like I'm not sure that would make any significant difference so for now we are going to remove band wordss that's all we also want to remove the tone it's fine we're going to pick up the tone from the other articles in Google Search right now we want this to be up to date and always follow the trend and we're also going to with the work count simply because the I is terrible counting it's just not going to be able to do that work count so even if you say I want a 2,000 words article it's not going to be a 2,000 words article the only way to instruct thei to actually write something of a specific length is to provide an example of that specific length so if you want 2,000 words you should provide as example in the prompt an article of 2,000 words then the I will try to replicate that and also output 2000 words but it's all going to be mimicking some sort of prompt it's not really thei understanding okay this is award and I want 2,000 of these I think the others work I think the other works okay so then we have a message block which is write a blog post about that and the second one is optimize the blog post for the user I actually don't think we need the optimize for SEO it's fine we can just add it to the main prompt here but what we do need is a separate block for generating the outline I almost never asked the I to generate the blog post right now because it's it's not good at reasoning especially if you use a model like CL 3 ion if you add an intelligence layer before that an intelligence block you can ask for the outline and the key talking points like we've seen in the example before and that usually performs much much better same thing for clonet actually like clonet is good but it's not as good as gp4 so what I would do here is add another send message block for now we can keep it empty but I would set the model to gp4 TBO and the purpose of this we always use markdown or I always use markdown in the send message prompt it's just a little bit easier so this is going to generate the outline and this is going to generate the post okay another thing here just in terms of lexic and how you talk to the AI it's usually better to call thei as assistant so you refer to assistant wres this assistant does this rather than write this or write that it's not super important in this particular example because it will be able to pick it up but it can be important in other in other use cases so here it can be assistant writes a blog post about topic the blog must inform engage and educate readers on the subject matter you must always include talking point keywords and call to action we don't have the go to action anymore so we can remove that make sure to use the full tone we don't have the tone so we can remove this and the part before so this generate the outline could be generate an outline about the topic topic just remember one important thing when you write topic like this you are calling the variable topic you are not saying the topic is this you are literally extracting what's inside the variable topic so if you type dogs here when this is resolved when the AI actually responds to you it will be a system writes a blog post about dogs right it will not be topic dogs it doesn't have the context of what that variable is so when you do it here you can't say generate an outline about topic well in this case it can kind of work but in other examples generate an outline about dogs doesn't make a lot of sense as generate an outline about the top Toc blogs dogs I think this is a bit clearer but you see it all the time I see it all the time in customers building their AIS it can be like this is the list of homes for example or this is the list of humans do like what the human suggested thei to do this doesn't do anything because this will get us this will get replaced with the actual value of the variable for example this is the list of dogs and this is the list of proper something like that so it basically you need to think about the variable as something that will be replaced not as something that instructs thei to do anything at all so in this example let's say generate an outline about the topic topic the talking points are this if you don't remember the name of the variable which is what I'm doing right now you can go in user inputs you can expand this list here let's also remove those we don't need them So it's talking point talking points are talking point the T we don't have it anymore the work count we don't have it I think that's all we have the call to action though let's includ the call to action consider the call to action is all to action okay so this generates a basic outline let's your gp4 turbo remember to change the temperature the default temperature is the maximum temperature for that model which is clearly not what you want gp4 turbo at temperature two is Will output giberish we want a temperature of like 0.5 maximum response tokens should be the maximum for that model which is 496 for gp4 Turbo the final one is Rev document that's fine yeah I think we can give it a go it to be very basic for now but we can give it a go so we want a blog post the blog post should be about let's change a topic let's say no code AI in the workspace any relevant keyword we can say no code ai ai in workspace AI use cases artificial intelligence open AI probably want to talk about open AI anthropic Google Microsoft sounds good explain the specific talking points we want to talk about AI for specific job roles and job titles we want how to create custom AIS without coding and we want maybe a table of use cases by cost call to action is you can build your own custom AI with M Studio you can get started for free okay let's see what it does so we have the outline no in the workpace that's not exactly a good title okay the title is Bel BL title empowering your team how noore transform roles and reduces cost okay that's not too bad blog leader introduction blog body introduction to nood AI a for specific job roles table of use cases we have the conclusion call to action and suggested imagery so there's still is suggested imagery for some Reon so we can troubleshoot that then let's click on next and see what happens with the actual writer let we want to open the debug just to check which model we're using where is in gp4 Turbo for this one that's okay taking a bit too long just check oh okay we're using 2.1 here yeah is very very slow oh and we're also assigning to a variable right because before we were showing it later that's fine so you will see that it actually goes directly to the document that's because the last step is assign to variable this should be display to user so if we display to user and we publish again it might fix it in the the Trad that we already started I don't think it it will do though yeah it won't but that's fine we still don't want call 2.1 call 2.1 is old model I suggest you don't really use it anymore the real new CLA 2.1 is CLA tonet but CL IU does work fairly well for like 90% of use cases clonet is a better writer though so let's go with clonet and again temperatures should be 0.5 let's also add another layer here because this article is still taking from the knowledge base of thei model it's not particularly up to date so what we want to do here between user input and send message we can also reorder these a little bit is add a custom function you can browse Community functions and you can search for Google search basic so Google search basic will search on Google and save the results in a variable the search query is going to be the topic so you can use the variable topic and the output variable is going to be Google search here you are retrieving a variable you're asking mind Studio to basically insert the topic here in this particular input the output variable is a variable you're creating right now and you're creating it to store the Google search results generally speaking in my studio actually not generally speaking just always whenever you recall a variable whenever you need the value in a variable that already exists you use the graphs you use the graph parenthesis here when you need to create a new variable and now store the variable the value there then you just write it without the graphs so this is graph graph topic and this is Google search without the graphs so now we have Google Search now this variable exists and from this moment on we can use that variable and give it to the model as context so consider the call to action is call to action after this we can say context from the current Google search on this topic and it's Google search okay let's do the same thing when it actually writes the content because I think that's useful for the generation as well so context from the current Google search and let's also add some rules so usually I add something like this on H2 rules for assistant both in the outline and in the block post generator and here you can have all sorts of things you want the assistant to do so one thing is always output in markdown and here we are generating the outline so we want the talking points and there was no talking points before so let's say include talking points for each title create a structure that includes H1 H2 H3 H4 and several paragraphs for each remember this article has to be SEO friendly and long form match the content with the current Google search and for the writing part here we can say output in markdown long form content only match the content in current Google search always add the talking points from human and I think this is a good starting point let's specify that these are the talking points these are the keywords and this is the context so that's fine we're searching on Google for the topic we are passing it to the outline generator then we're passing it to the article generator and we have a few rules for the assistant there's an extra s there okay let's publish and try it out let's do no code AI let's list the keywords no code AI open AI Microsoft Google yeah in the workspace talking points let's actually copy them from before you want to copy it from before you can go in the debugger so the debugger is here this is our thread these are the talking points from before you can see that in user input we have the step four which are the talking points let's copy those let's paste them here list the keyword that's fine explaining specific talking points okay for some reason it starts from the beginning let's do block post again nood AI the keywords are nood a open AI Google it will pick them out from the content anyway I'm not worried about that if you have a list from an SEO tool something like samash or HRA you can copy paste it here these are the talking points the call to action is try my studio for free let's wait let see what happens this is better and power your career with no AI that makes sense you can see that it does include the talking points now so there is H2 H3 these are the talking points and it is taking context from Google you can see that because it talks about things it might not know about without searching on Google like mid-range AI Solutions on Enterprise level AI it's still suggesting the imagery which means that it's probably in the main prompt we can check that lat later and then it's writing it out with CLA 3et I show the debugger this was gp4 turbo and now it's writing it out with CET so already after 20 minutes even less including the explanation we have a brighter that's significantly better than chpt like if you go to chpt and you ask can you write an article on no AI it will not write this thing you can still make it better of course this is not a perfect article and uh the thing I I showed you before that generator is obviously better because the prompts are very very long but that's actually that's actually the the main benefit of using mind studio right like if you don't like how something comes up and we didn't like the first output we just change that you don't need to wait for a software developer to go ahead and do it on their software you don't need to wait for specific imp inputs or for specific features requests or stuff like that you can literally go in the back end of your application here and change it whenever you want and you can change it in one second if you don't like something for example right now we don't want the imagery anymore that's probably in the main prompt here so we have imagery and links yeah so we removed it from the blog ideas before I think and we didn't remove it from here so if I want to remove it I just go go here remove it and that's literally it I can also remove the call to action that's it if I want the outline to include the talking points like I did I go there I say include the talking points that's all so it's not really about what the the end result is after hours and hours it's what you can do in a few minutes to just change the ey whenever you want without any developer help so to make this a little bit better you should also tune these prompts a little bit more I know that the box is a little small so it doesn't suggest that you can write more but you definitely can like you can write very very long prompts in here in the assistant I showed you before this is the block to generate the outline and you can see that we have the same basic structure but we do use more markdown so this is a title one generate an outline this says topic identification content inspiration and this thing is massive right like it's a much much bigger prompt and the final one is also much bigger so why is it big it's not because I came up with all of this out of my mind as soon as I was writing it it's because I tested this writer many many times like we did right now like I tested with a few keywords I tested with a few topics and I've seen the results and I've noticed stuff that I didn't like in the results so what this prompt tries to address is after testing it many times I believe this is the list of things that the will do that I don't like and I want to include them in the prompt as exclusionary stuff or stuff that I want to to clarify for example R for here is the article should be very long and I bold that very because in some cases even if I said I want the article to be long or I want the article to be 2,000 wordss long it would still write a very short article assistant will use multiple titles that's also very important and assistant will not output H2 title but just the title because it would say instead of just using the H2 it would say H2 and then the title of the H2 and I you like that as well so what this list here does is basically check whatever I don't want the ey to do and whatever I want I want the ey to do so assistant must bold a few important keywords assistant will generate the content now it will not show the outline etc etc so one thing you can do other than having the Deep scrape instead of Google Search is to tune these prompts a little bit more and to not display it to the user every time the only reason you would do that is that when you have a very long prompt and a very long response that enters the tread and once it's in the tread every subsequent message with the AI will keep it in context so the reason this blog article knows about the outline is that I'm showing the outline to the user here I have a send message block the send message block is set to display to user and it's generating the outline anything that is displayed to the user enters the tread and is passed down to the next message so this outline is then passed to this to this send message block now this send message block knows about the outline it knows what happened before and can generate the article another way to do this when you have many many more prompts or many many more send message blocks is to save this as a variable which is what I usually do and the variable can be outlined and then here in the blog post generator which is the actual final send message block you can say assistant writes a blog post about topic and that's the topic and based on the following outline I forget to copy so yeah the outline is in outline and this will take the response from the previous send message block and include it here now in this example it will pretty much not change anything but if you have more prompts if you have more send message blocks piled up on each other that can be really helpful another thing that you can do if you want your article to have an image as well would be to add another custom function here and we have custom functions for th3 we have custom functions for stable diffusion and and whenever you use these functions you need your own API key so you need to include the API key here but then you can choose the model for example the three and the prompt could be the topic so now you're asking the to generate an image on this particular topic and you can then use this topic image in the final output and I don't know that that can be better if you're writing about something that can can be generated with AI for example if you're writing about dogs AI can gener generally generate dogs it cannot generate software screenshots so if you're building something technical not the best but it can generate abstract images it can generate peoples it can generate um events it can generate stuff like that so that's something cool you can do or you can do something bigger you can scrape a large amount of data something like the Deep scrape I showed you before and then you can use the send message blocks to refine the data and that's what I did in the the other AI so instead of having just a basic search you have the complex one and then you use the send message block to take this response here which is Google search and you say for example from Google search extract this whatever that is it could be the list of titles it could be the list of content it could be the main keywords it could be the topics it could be the structure it could be whatever you want extract this and then you can save it as a variable and that will be that thing you're extracting like titles for example you can do the same thing after and you can do same thing from Google search extract main topics you know and then you can save this as a variable and the variable name is topics you can even reorganize them a little bit just to understand what's going on so you could do the search then you refine it and then you pass it to the send message block and now you have the outline generator and the block post generator that ends in a document so this is a much bigger idea where you're taking a very large Corpus of data and extracting it using thei and you can use the AI that's cheaper actually which is CLA 3 I because extracting things is something that pretty much any AI in mind Studio can do you don't need an expensive AI to do extractions of any sort especially if the content is already in the in the prompt so here I'm saying from this thing here from the Google search you can actually specify what it is it makes it a little bit better you want to extract something specific you are already giving the full context of what the needs to do you're already telling the everything this is the content you don't need to invent anything you don't need to write anything you don't need to understand anything the only thing you need to understand is where is this particular component I want to extract for example where are the titles where are the content pieces then it will extract that and use it later on so I think that was the basic let's actually use that function here and then we can go through the questions in in the last 10 minutes I think so to create a custom function in mind Studio instead of using the Community functions you can click on any function and create on on click on new or you can go here in the side panel and click on the plus icon you can name the function let's name the function deep scrape and there is deep scrape scrapes top 10 results in Google saves content and titles I think it saves the URL as well so content URLs and titles I'll copy paste from the other we're not going to write it right now but I'll copy paste it so this is the code for that it limits it to the top 10 results it will not scrape more then we have the configuration tab which is where the user actually chooses what to do and then the test data which is where you test the function so here we are testing with latest marketing Trend and the output variable is full text results you can always test this let's test it it actually works so let's click on the play button will take a long time CU it's craving the top 10 results but it should work the meantime I'll just read through the chat very quickly so I know what to focus on later okay so you can see what it did is full text results this is an array of results you can see that from the square bracket and then you have the URL for each of these top 10 results and you have the content for each of these top 10 results so this is a gigantic output as you can see and that's why we're using clo clak is very very cheap so when you have a prompt analyzer that you need to understand AIT where you need to understand a very large Corpus of data C IU best choice of pretty much all the other models unless it's very specific but in this case we're not doing a high intelligence task we're simply asking quu to summarize this stuff for us so let's go back to blog post let's switch the function to deep scrape you see that this changes a little bit we have the search query and we have the output variable but it's pretty much the same as Google search basic so let's do topic here and the output variable is full scrape let's extract stuff right right now so from Google search let's actually say from this well let's generate a summary first so let's generate a summary from this list of ranking results on Google including URL title and full content assistant will generate a long summary just so that it doesn't conscribe it too much to a few sentences along summary with and let's say what we actually want so we want the title we want key talking points we want we want key data probably like statistics numbers any sort of original research main keywords in the content and that's probably it then build a topical model and find NLP LSI keywords which are just ways to say find similar keywords for SEO purposes for the topic topic that you spot in the ranking results so this will analyze the entire thing that we've just seen in the test data so it's all the top ranking results all of the content and it will find the title key talking points key data points and Main keywords then it will find similar keywords for AO purposes for us to rank on let's assign this to variable summary The Next Step will be from summary actually that's fine we don't need another step right now these the only extractions we need let's streamline this a little bit so we have the user input topic keyword talking points call to action we are scraping Google we are getting the top 10 results we're getting all the content we are getting a summary of the content built by CL 3 IQ the maximum response size token is 4,096 then we need to pass this summary here to the outline right so generate an outline about the topic assistant also knows these are the topic and keywords currently ranking Google and we paste in the summary here and same thing here in the generate post since we're assigning to variable remember that now this doesn't pass on automatically this is not displayed to the user so unless I specify this variable in the send message block here the final one it will not know about the search because it's not displayed to the user and anything that it's not displayed to the user is kind of contained in its own Department it's not displayed it doesn't actually follow the flow right just because this part is here this send message block is here doesn't mean that the send message block three way three blocks down the line knows about it so even if I have the summary here if I don't refer to the summary anywhere if I don't say it it should refer to it here and it should refer to it here the summary would basically just stay there and never get P to the the following message blocks so let's do the same here we can say assistant also also knows the these are the topics topics and keywords going on Google so that works final is document and I think we're pretty much good to go so let's try it one more time these will obviously take more time because now it will search on Google for much more stuff but let's go with no code AI relevant keywords or no AI open AI Google tools stuff like that but it will pick them up from the search anyway talking points noi is the future you can start for free next we don't really want a call to action it's fine and now it's searching on Google so it's using the custom function you can always see what's happening in the debugger so let's open it up again it will take like one minute probably so in the meantime feel free to start posting questions in the chat as soon as this is done I'll go through the questions and we can have a discussion about it okay so it's doing something I think there's probably an issue summarizing it here on the side panel but I think it did scrape it because I see some results down there we can check in we can check check in the debugger later if it worked fine we have the background message so the background message is generate the summary we are generating the summary from the Google search results and this is what I received from CL IU so this is the summary about the particular search now we need to send the summary to the outline let's see if that works fine okay we sent it to the sum to the outline we are generating the outline this is the final response for the outline generation and now it's generating the actual article and I think you can notice that the article itself is following a much better structure so here we have the generate the post this is the resolve message and this is the output so we have the title the title is 10 times better in my opinion this is the introduction you have the democratization of AI through noal platforms the business impact and you start to see some numbers some numbers could be inaccurate but since it's in the training data it's it's in our prompt it's likely true you can you need to obviously fact check this but you will start to see some numbers pop up so 13.2 billion in 2020 to 45.5 billion in 2025 300% Roi you have some tools finally not only the ones we mentioned but also other so other tools there are numerous other tools such as ACO lob nanonets each with with its own features this is again coming from the Google search it's not in the training data it's not in something I said it's coming from there even the structure itself is better and it's coming from the other H2S and final things in the Articles and then we the conclusion of the call to action so personally I still think this is short so we definitely need to tune it a little bit more to generate longer outputs but the quality of the outputs is much much better than before and we used 28,000 tokens for the final one which is not crazy we used 1.6 6,000 tokens for the outline generation and 950 tokens for the actual summary the only one that uses a lot of things is this Ser here and I think some of these articles weren't scrapable you can see it from this result here unfortunately some things are just not scrapable they block scrapers so even if you do use a scraping function in mind Studio it will not work but we can see what happened so here if you go in the debugger you can see that most of these were actually scraped just fine so we can see the URL of the content okay I should have not clicked on it I can see the URL of the content so here's the content in the HTML format here's the URL the content so all of this was fed into the AI as context so you know that the data is not just invented it's not just an hallucination it's actually coming from articles ranking right now on Google including the the whole context and the topics the final keywords Etc okay so let's go through some questions open the chat I've seen a couple before as well so idea for a potential feature it would be awesome to be able to insert a document editor into a workflow yeah that's a that's a good idea so basically you want to edit the prompt in a way before it's sent to the AI and then you want thei to respond to that instead of responding to the basic one yeah let's take good idea we can add this feature request Rob says is there any versioning happening behind the scenes for these prompts for example can you revert to a previous version if you've experimented too much not really it's um it's something that I actually would like as well so it's something we can add feure request for sure right now the best way would probably be to just duplicate the workflow so once you have a version that you're happy with you can name the workflow version one underscore name of the workflow then you can duplicate that workflow you can do your tests and then keep the one that you like it's it's not perfect but it's the best way to do it right now do you have a step-by-step guide that explains what you've done we will send a recording of This webinar or Workshop what you want to call it but we also have a YouTube tutorial so if you go on our YouTube channel there is a YouTube tutorial on how to build a Content writer it's not as complicated as this one it's not as complete but it shows the basics of how to structure the outline how to structure the blog post how to structure the the workflow as a whole so I think it's a good starting point as well Jimmy says about variables you mentioned how to pass them along within the workflow is there a limit on their numbers or Lipan yes so it depends on the type of variable the onboarding variables so the ones that you have here and I suggest that you do use them it's just not useful for this work in particular but usually you would have the name of the person their company name their website URL all of that stuff is the same every time they interact with your AI and even if it's yourself so you can add them as on boarding variables and those variables last in all threads throughout all of your interactions with thei it does expire at some point but it's basically when you cancel the data like when you remove the storage from your browser or when you you log out you know stuff like that actually not when you look out when you look out the stores so it's only when you delete the storage from your browser if it's an embedded app otherwise the global variables are tied to your account for that particular AI application and they will be there for you when you need them across the entire workflows across the entire app and whenever you need them then you have the runtime variables which are whenever you open the debugger you can see the runtime variables so we have the topic the keywords the Google search all of the stuff that we've tried during the workshop today all of those were runtime variables runtime variables last for the current thread so whenever you create a new thread those runtime variables are not available anymore and you will override whatever is available before one important thing though is that the runtime variable expires when you create a new thread not when you move to a different workflow so if from the workflow blog post ideas I want to save something and then add a jump block to my blog post generator I can definitely do that and I can pass a variable from the workflow blog ideas to the workflow blog post and it will still be there it will still be in the context because it's not a new thread I'm just jumping from one work to the other but I'm staying in the same thread and as long as you're within the same thread the variable are there the runtime variables about formatting rules can you specify them one for all the initial prompt does the assistant always refers to the initial prompt on top of the current task so the the initial prompt is a system prompt whenever you look at the API for example for open AI or anthropic this prompt is the system prompt it's not supposed to be a a complete instruction on what you want the to do it's supposed to be what thei is so the system prompt is you are a Content generator you are an SEO expert you are an accounting expert that's supposed to be the system prompt when you want to instruct the I to do something in particular you should add that to the send message block because yes this is kind of in the background whenever you do anything with thei and there are many things you you don't want the system promp to always do you don't want a block post generator to always think about some specific topics for example you want this system prompt to be very generic you want it to represent the AI so you are a content generator your responses are in markdown you are professional you don't swear you don't you use emojis or you don't use emojis that's fine those are generic inputs that you want for all of your subsequent interactions with the AI but you should keep the actual information like I want you to generate the outline in this way I want you to generate the extraction right now in this way I want you to generate the blog post and avoid too many titles Bal some examples I want you to use a lot of lists I want you to include some tables all of these specifications should be in the send message blocks because those are the prompts the actual prompts that you send to the AI not the system message which is the same across all the prompts Rob says duplicating workfl is a great tip would also be great a great way to track the state of a project at certain key points imagining a case where something is shipped and the client may have made changes they forgot about yes that's that's also a good point so like logging the changes in your workflow in case you hand it out to a client could be helpful right now we don't really suggest that you hand it over to a client that's not familiar with M Studio or that doesn't know what they're doing in the platform it's probably better if you keep control of that and they're just viewers so there's a specific role in the workspace for people to use thei but not to edit theis you should probably have your client with that role like user instead of being an admin where you can edit the applications so says will we be able to pull in all of these user inputs through a form and fit it into the same AI using zapier so that those specific inputs connect to the variables we've set in our AI in mind Studio yes you can do that so you can trigger my studio workflows from wherever you want I have a blog post generator that works instead of slack so I just slack it something like a keyword list of URLs it scripts the URLs it follows the keyword rules and finds different topics for the particular keyword and then it PR the content all of that is doable as long as you send the variables in mind Studio through zapier so as an example here in the final send message block I need the topic di needs the topic the needs the summary the I needs the outline it needs the talking points the keywords and Google search in this particular my studio workflow I'm sending all of this stuff from our UI from this particular workflow it doesn't need to be that way if you want to send it through the API that's fine but these things need to exist and if you use the API the syntax is a bit different for the variables so if you're sending the topic to the Mind Studio UI you just need to write it up like this if you're sending it through the API you need it like this so launch variables topic and then you would go in zapier and map this topic with whatever the topic field is so if you have a form and topic is cell number one then you would map launch variables topic to cell number one and use that it's a bit complicated to explain in in two minutes but we have a full tutorial on how to use zaper we have an example on how to deploy um an application in mind Studio inside of slack we have an example on how to deploy it in Microsoft teams we also have a tutorial on how to use it with make.com where we go through the full API implementation so how to call the API how to retrieve a value from the API it basically teaches you how to call this these kind of workflows from wherever you want it can be a chatbot it can be a form it can be anywhere where you have a connection through zapier or make and those integrate with thousands of applications so they're very flexible and you can definitely expand this across your applications and and your text tack even if you don't use the my studio UI Jimmy says it is great to know that variable keeps alive between workflows is there a size limit there probably is just because there is a limit to everything but I'm not aware of that and I don't think that it's it's a limit you should worry about um I have very very large ones like for example right now we have this debugger and the uh full scrape as you can see here it says the string is too large to display but it doesn't mean that it's not there like it's actually used the context and it's using everything that's been scraped it's just too long to display in the UI I'm not aware of the specific technical limitation I'm sure there is a technical one but it's very very large be free to type in more questions if you have any I will also just quickly publish this one and make it remixable so you already have the link so this is the link to the app you can already remix it so you have the functions ready in your workspace okay um no problem just say it was helpful and and I hope that's how you pronounce your name so if there are more no more questions I'll I'll let you go but we can meet up in different workshops we have more coming up and you can find them on blog. feel free to sign up to the next one and Jimmy says can you show some insights about the document you mean the document Terminator the document Terminator is yeah I can so the document Terminator is basically um a sort of Google doc where you can write up whatever you want the extra power here comes from using AI within this document editor let's say that you don't like one of these titles you can select the title and you can highlight it and then you can click on this m magic wand there on the right it will show it in yellow so the Highlight will turn yellow and you can revise the selected text for example write an alternative title if you click on submit to take a little bit of time and then you have democra democratizing ai how no platforms are power in businesses okay so you can click on yes and as you can see it replaced it you need to change the formatting back to H2 but you can basically get alternative text to a lot things you're doing also the this is coming from The Prompt so if you go back to the blog post flow the document revisor has a revision template and the revision template says Rite the following text respond with all in the revised text and nothing else Etc then you have the text which is select the text this the default variable it's whatever you select in the document Revis Terminator and you have the instructions which is what you write inside of that block so whenever you select something this is the selected text when you click on this magic wand what you write in here are the instructions so you have those two variables that you can play around with and you can change this Rite Rite um template so if you don't want it to rewrite something but you want to write something you can do that so you can say write the continuation of this paragraph for example the text can be paragraph and the instructions could turn into guidelines for example and if I publish this again I can go back here refresh and let's say magic wand write the key talking points so now instead of rewriting the introduction it should actually output new content and that new content should be the key talking points I'm not sure it actually did very well I think they are the talking points it just didn't l them out very well like these are probably the talking points in the AI mind but they W format it properly so the only format it has as you could see from the title Ry as well it seems to be the standard text it doesn't actually output a formatted document but it does follow your instructions okay uh thank you everyone again for participating feel free to register register to our next workshops we have other life builds we have different experiences we're trying experiments we want to find the best format for you all to to understand more about my studio play out with the tool and have more features so catch you on the next one and have a good one bye
