스크립트
00:00:00you're using cloud code in obsidian the wrong way or at least you're barely scratching the surface
00:00:05we can use obsidian to do a lot more than just track markdown files for us instead we can use
00:00:11it to create a command center for our entire cloud os a system that includes a fleet of skills and
00:00:16automations a memory layer and a visual wrapper on top of it all that gives us insights that you
00:00:22just can't get with the terminal or the desktop app did i also mention that we can give it fully
00:00:27local voice capabilities but how do we do this in a way that actually provides value and doesn't
00:00:31devolve into a zero substance visual spectacle well that's exactly what we're going to cover today and
00:00:37by the end of this video you're going to understand how to build this yourself now if you follow this
00:00:42channel for some time you know that we've gone through a few iterations of this clod os not too
00:00:47long ago it looked something like this which was a web-based version and before that we had an obsidian
00:00:54version now the cool thing with this web-based version we had was the voice capabilities a fully
00:00:59local voice setup that allowed us to get very quick responses and control some things through our
00:01:03microphone now what i've done is we've essentially upgraded the obsidian command center and we brought
00:01:10in the best parts of the web version and combined it into what you see here this gives us all the power
00:01:16of obsidian i've included a voice setup which was very much inspired by what we see from codex and chat
00:01:23gpt's voice mode and we now have a one-stop shop for everything we want to do with cloud code but what
00:01:29does this whole thing actually buy us that we don't get inside the terminal don't get inside the desktop
00:01:32app we have to answer that question or else this whole thing is kind of pointless let's start on the
00:01:37visual side with what we're actually looking at first things first it gives us a simple glance at
00:01:41whatever single metric we want what i really want to hammer in here is that this is 100 customizable
00:01:46so whatever you want to be able to see you can put inside here this is just a custom obsidian plugin
00:01:51that cloud code actually built for me i want some sort of insight on what i'm doing with my token usage
00:01:56i can see all of my metrics across social media channels i can see my schedule that is pulled from my
00:02:01google calendar from that it breaks down specific tasks for me and then i also get some insight in terms of
00:02:07my morning headlines right here you can see different automations and skills that i can run at a click
00:02:12of the button and i have different tabs here that dive deeper into these different reports that i run
00:02:18daily for example this is my research report this comes from a morning intel automation which gives me
00:02:24all the github trending repos based on the last seven days the last 30 days and then i can also see
00:02:29outlier videos on youtube different content angles and then i get sort of a running look at what's trending on
00:02:35hacker news when i look at my audience tab this is taking a look at how my specific content is doing
00:02:40across youtube what's working what isn't and why and since this is obsidian i'm not limited just to
00:02:45this i can use the terminal inside of obsidian and this is partly why i like this voice modes because i
00:02:51can pull this up at any time and say something like pull up the terminal
00:02:57and cool there it is now this thing does have actual voice stuff so if i say something like hey what was
00:03:03the biggest ai news story for today claude got a model level watermark rolled out globally yesterday
00:03:13invisible survives copy slash paste triggered by the eu ai act deep seek harness also exploded to 109
00:03:23k stars in two days and is dominating github trending so that's kind of cool that i can like ask it a to do
00:03:32commands inside obsidian like i said pull up my terminal or whatever and i can ask it questions
00:03:37about reports i've already run that's why it was relatively quick it didn't have to like take my
00:03:41command go out on the web find that information process it and then respond it's been wired to
00:03:47look at these reports that i've already run again something like my research report and then quickly
00:03:50answer questions about it the other cool thing about voice mode is this works even when i'm not inside
00:03:54obsidian so let's say i'm tabbed out i'm doing some other work elsewhere like for example i'm inside the claw desktop app
00:04:00yet i want to ask it a question about something i essentially want to call on claude to audibly
00:04:05talk to me again very similar to what we do with gpt live mode which you can do inside of codex this is
00:04:10a really good tool so in this case i just key the hotkey and i can ask something like what's the biggest
00:04:16thing on my schedule for today
00:04:23your biggest commitment today is the sponsor deliverable review for warp at 11 a.m then the chase
00:04:30ai plus coaching call at 3 p.m but the whole day is really built around the voice orb demo record block
00:04:37at 9 30 which is the main focus and so like you see there i'm able to ask questions about different
00:04:46things and it can quickly respond now this also the sort of voice mode thing it doesn't just answer
00:04:52questions about you know reports i have going on it can actually call on claude in the background and
00:04:57run a headless version of it so if i was doing something else and then i actually wanted cloud
00:05:02code to go ahead and do something like relatively deep like hey go ahead and start you know planning
00:05:07this project or start you know doing deep research on x y and z it would go ahead and do that it would
00:05:12do it in a headless mode now the voice response wouldn't be as quick because it has to you know
00:05:16actually go and do it and then respond but for these quick tasks that are pulling on information
00:05:21i've already found it's relatively smooth and again this is all local this is a local voice mode
00:05:27running on my gpu and i'll show you kind of how it's set up but overall i think in terms
00:05:31of the visual page that's sort of the value this sort of like wrapper os obsidian setup is bringing
00:05:37you which i think is pretty cool especially the voice mode when it's tapped out because
00:05:40being able to audibly interact with cloud code even when you're tabbed out and stuff in a way that is
00:05:44similar again to that gpt voice mode is something i wanted to bring over because the voice mode inside
00:05:48of like the cloud desktop app just isn't really there so that's what the user interface and voice mode is
00:05:53actually bringing to the table the second thing that's actually providing value in a cloud os setup
00:05:57is that skill and automation backbone you see all these skills right here that have been you know
00:06:03essentially moved into buttons that i can click at any one time or i can have it run it being voice
00:06:07mode and that's great the real value is that i have taken everything i do in my day to day and i have
00:06:12turned them into skills and automations which is a simple enough concept but very few people do it
00:06:16correctly what you're looking at here is a visualization of all the skills and automations
00:06:21i run and the concept is relatively simple you have a number of tasks that you run on a day-to-day and
00:06:28week-to-week basis both personally and with your job these tasks are broken out into different domains
00:06:34for me that includes things like memory productivity research content my community my agency sales so on and
00:06:41so forth within each of these domains there's specific things that i do for research for example i have a
00:06:48pipeline dedicated to youtube i deep research specific topics i have stuff that goes into my light rag
00:06:53database i have my morning trend scan which is essentially that report you saw earlier i want to
00:06:57be able to look at my competitors etc etc etc i could do all these manually or after i figured out how i
00:07:04actually want it to work manually i turn them into skills and then once they've been turned into skills i ask
00:07:09myself does it make sense for this to be an automation if yes that's what i do now doing this has nothing
00:07:14to really do with like a cloud os in the obsidian dashboard you said this is just like best practice
00:07:18every everyone should be doing however building this is what essentially is the foundation of the
00:07:26cloud os if you don't have this creating something like this is kind of pointless like what are we even
00:07:31doing here there's no like information that's going to travel here you're not going to be able to call
00:07:34on specific skills and automations regularly this becomes a visual spectacle
00:07:39if we don't have this foundation so this is the actual substance and we'll go into the best
00:07:44practices of creating something like this for yourself because it's not really hard
00:07:47and then last but not least we have sort of the obsidian memory layer which i think people get
00:07:52confused with the whole point of the obsidian memory layer isn't that it's necessarily super
00:07:57charging cloud code like we see all the stuff and we get like super hyped up about oh my gosh look at
00:08:02like this graph this looks super cool it's like a graph rag like it's not like a graph rag it's not like
00:08:06graph rag at all actually what obsidian does is it organizes information and markdown files for you
00:08:12the human and there is a small tangential benefit for cloud code itself that tangential benefit for
00:08:19cloud code comes into effect when we have thousands of files and we have set up our obsidian ball in a
00:08:25way that actually makes sense it makes it easy for cloud code to navigate to give us accurate answers
00:08:29quickly essentially saves us on tokens and boosts the performance a little bit and that is entirely a result
00:08:34of how you set it up what you see here is that carpathy obsidian rag system you see everywhere and this is sort of the
00:08:40template for it and again we'll go into deeper detail of how you would set it up because you don't have to do it
00:08:45exactly like carpathy did but there are some fundamentals we can pull from it so that's what the
00:08:50subsidian plus cloud os does that's what it buys you and what you can't get inside the
00:08:54terminal or inside the desktop app and now let's talk about how you would actually create this for
00:09:00yourself but first a quick word from today's sponsor me so i just released a brand new version of my
00:09:06cloud code master class and is the number one way to master this tool especially if you don't come from
00:09:11a technical background we focus on real practical use cases so you can actually learn the fundamentals
00:09:17and then turn around and apply this to whatever personal project you want we're currently running
00:09:22a sale on the membership so if you want to join check the link in the pinned comment you can find
00:09:26it inside of chase ai plus now before i explain how to create your own skill and automation backbone
00:09:31for your cloud os let me walk you through the voice setup so you sort of understand how it's actually
00:09:37working so in this example you are saying hey what's today's biggest ai news this voice command
00:09:46goes to the voice module call jarvis whatever you want and that voice recording is then transcribed
00:09:53via faster whisper which is a open source repo it's going to live on your computer does a transcription
00:09:59from there your voice got turned into words and those words get sent to haiku 4.5 i used haiku
00:10:06because it's the smallest cheapest fastest model we also have the ability if you want to swap out haiku
00:10:12for any sort of local model the reason i don't have the local setup is because
00:10:16i give this exact setup away to people you can find inside of chase ai plus and i don't know
00:10:21necessarily what their hardware is going to be so the power of their local model can change a lot so
00:10:26haiku is like a nice standard but if you have the hardware for it you could change this for really
00:10:30whatever you want throw in quen throw whatever whatever so what's the biggest ai news it becomes text
00:10:35that text goes to haiku and haiku now figures out where to route this and there's really three tiers
00:10:40this is broken into so tier one is skills if i told you know jarvis hey i want you to run the morning
00:10:47intel report well that's going to get routed to tier one and it knows it's just going to have to execute
00:10:51a skill it doesn't have to add any other words to it just execute the skill just do what you're told
00:10:55right and then let me know when it's done tier two is hey i just want you to pull metrics so that's
00:11:01what we're asking here i'm saying what's today's biggest ai news i wanted to just look at a report that
00:11:05already exists so if i'm asking about news it knows it can go to essentially the research tab take a
00:11:11look at our morning intel report and let me know these things we break out this specific tier because
00:11:15we want this to be fast i don't want it to do any like searching or research on its own just find
00:11:19whatever exists and tell me so that would be what this command goes goes into tier two then we have
00:11:24tier three and i alluded to this earlier if i pull up voice mode and i say hey i want you to go ahead
00:11:28pull up fable 5 i want you to start doing deep research on task x y and z and then coming up
00:11:32with a plan for action well that isn't a specific skill that's not a metric we need to actually like
00:11:38pull up an actual instance of clawed code and that's what this is going to do it's going to do
00:11:42it completely headlessly it's going to run all the things and then it's going to report when it's
00:11:45done so this becomes kind of the catch-all this requires the most power and obviously it's going
00:11:49to be the slowest now once it executes any of these tasks whether it's in tier one two or three
00:11:54it's obviously the response is going to be in text but we want that to be a voice response
00:11:59so that is where kokoro comes in again an open source tool which then gets sent back up to here
00:12:06and then it responds to us with you know whatever it's going to say so that's how the voice mode works
00:12:11you say command it gets transcribed it gets routed command is executed turn into a voice sent back up
00:12:18and it's actually simple enough and like i said you could turn this 100 local if you wanted to
00:12:22by simply swapping out haiku for your local model of choice now let's talk about the skills i have
00:12:28some of my most commonly used ones right here so at any time i want to run them and i have this tab
00:12:32open i can just click on them and like you saw as we went through the voice mode we can easily trigger
00:12:37them just by saying the name through the voice system don't even have to be tabbed in but the
00:12:42question becomes how do we actually create these skills and what should those skills look like luckily
00:12:46it's complicated this looks the workflow for doing this actually pretty simple step one is defining
00:12:52the tasks what is it you actually do in your day to day that we can codify remember one of the issues
00:12:59with ai is that it is inherently non-deterministic i tell it to do one thing and i tell through that one
00:13:05thing 10 different times it's going to go about it 10 different ways if i can instead specify the way i
00:13:11want things done and specify an end state we can then turn that into a skill and essentially codify
00:13:16it make it as deterministic as possible but we have to know what those tasks even are so the way we
00:13:21can do that is well luckily we actually have a few options to do this the first option is we just kind
00:13:27of give it a stream of consciousness all right you're going to turn on your microphone and you are just
00:13:33going to yap the cloud code and explain what you do every day and what you do every week based on that
00:13:39conversation we're then going to ask cloud code hey i just gave you a bunch of information about what i
00:13:44do which parts of that can we turn into skills and then it turns them into skills for you the second
00:13:50way to do this is you have it take a look at your actual logs everything you do inside of cloud code
00:13:55every single tool call every single command every single text is logged on your computer we can have
00:14:01cloud code take a look at these logs over the last 30 16 90 days and then pull skills out of that this
00:14:07is really nice because we have an idea in our mind right about what it is we actually do
00:14:13but what is the reality of that right what do you literally use cloud code for we're going to be able
00:14:18to see this in the logs and ideally you would actually do step number three which is a combo of both of
00:14:22these things right you do the screen of consciousness you tell cloud code what you think it is you actually
00:14:27do and it creates skills from that then we take a look at the logs themselves and pull that out and
00:14:33in combination with those two things you begin to build this repertoire of skills now which skills
00:14:40do we turn into automations oftentimes that's going to be pretty self-evident right is this something you
00:14:46need to run every single day to specific time well then turn it into an automation turn it into routine
00:14:51which is one command away you just tell cloud code to do it for you now when it comes to automations and
00:14:56skills really the best practice is to keep it as a skill first for a while actually make sure this is
00:15:01working the way in the manner you want and then once you're happy with the outputs you're getting
00:15:05with the skills then we go ahead and turn it into a full-blown automation that we don't run manually
00:15:10it just does its thing and we can go even deeper with this and i'm not going to do that in this
00:15:15lesson and essentially start adding some like loop engineering principles where we're taking a look
00:15:19at these automations and that run we compare them with their past outputs and we build some sort of
00:15:25system where it becomes self-improving right we give it some sort of specific goal some sort of
00:15:29criteria trying to hit and if that automation doesn't hit it we can continue to adjust but that's
00:15:34a little more advanced just understand that's where you can go with this but it's all built upon the
00:15:39skills right can we codify what you do every day into a skill the answer is yes and it is simple from
00:15:46there integrating that into your cloud os is really just as simple as this like which ones do you use
00:15:50daily and in terms of the voice mode section it already knows what skills exist so there's nothing you
00:15:56really even have to do there remember like it's haiku it's clawed it has a list of everything you do
00:16:00in the background and those skills are also going to form the basis of what your command center even
00:16:04looks like here right everything you see here is essentially some sort of skill or some sort of
00:16:08automation whether it's pulling my schedule from google whether it's the morning headlines whether
00:16:13it's pulling metrics from social media whether it's this research section this is all just outputs and
00:16:18data from automations that's all this is now yours is going to look different and that's a good thing that's why
00:16:24i think the system is great because it is fully customizable whatever it is you want to see
00:16:30should be an output of a skill and then we throw it on here and boom it's yours now let's talk about
00:16:36the memory side of the equation and what obsidian actually brings to the table in terms of quote unquote
00:16:40improving clod's memory now this is the tweet from carpathy that made everyone lose their minds
00:16:45over obsidian and sort of that clod code connection like this was out in april nearly 22 million views
00:16:51and sort of the system he laid out is relatively simple and this is what it looks like our obsidian
00:16:56command center our clod os primarily lives in a folder known as the vault doesn't literally
00:17:02have to be called the vault but that's sort of the naming conventions that obsidian uses it's just a
00:17:06folder on your computer within that folder we have three subfolders we have raw we have wiki and we have output
00:17:15so i want you to imagine we have claude open up in this vault folder and we say i want you to do some
00:17:23research on rag systems so it goes out and it finds a bunch of data a bunch of sources a bunch of raw
00:17:31information about ai asians all that raw data is essentially put inside the raw folder it lives here
00:17:37right we have article about rag now let's say you want to go a step further and you want to take all
00:17:44that raw information you want to turn it into like a wikipedia style article well that thing goes into
00:17:49the wiki section and you can see here we have a sub folder in the wiki section for rag systems so all
00:17:56that data gets turned into a new article about vector databases and we have a second article that has to
00:18:02do with chunking strategies right we're essentially just synthesizing all that raw information we found
00:18:07now let's say we want to turn that into a slide deck i want a slide deck talking about rag systems
00:18:12well then that slide deck is going to get put into the output folder right we're turning into an actual
00:18:17deliverable and it is now a slide deck so the simple idea is we have a section for raw information
00:18:24we have a section for sort of just like synthesized data and then we have the deliverables we create
00:18:28from the synthesized data now the important thing here isn't that we have that sort of you know
00:18:34breakdown it's that it gives ai a clear path forward to find information and the way it does it isn't
00:18:41just because we have the folders it's because at every step of the way we have essentially a master index
00:18:48or an index.markdown file which acts as a table of contents that says hey once you go into the wiki
00:18:53folder you can now read this markdown file that says here's the three subfolders inside here we
00:18:59have folders for ai agents rag systems and content creation does ai need a master index a table of
00:19:04contents when you have three subfolders absolutely not but what if you have three thousand what if you
00:19:08have thirty thousand what if you have three million right it becomes a little bit easy for everyone
00:19:12involved if there is a table of contents and that idea is repeated as we go one layer deep so
00:19:18i went to wiki i saw the master index i wanted to learn about rag systems will i go into the side
00:19:23the rag systems what do i want to learn about rag systems well guess what there's another index file
00:19:27in here that says hey vector databases chunking strategies etc and the whole point of this is
00:19:33is there's really two points one for you if you the human being wants to be able to find stuff it's
00:19:38pretty easy two there's a clear path we give cloud code a map to find things and because it has that map
00:19:44it gives us answers quicker and it's going to burn less tokens doing so and so that's the whole purpose
00:19:48of the obsidian memory it's not really giving it more memory it's giving it a map now in terms of
00:19:52practical application do you have to set up exactly this way with raw wiki and output absolutely not you
00:19:57do not but you need to set up something that makes sense for you you can see here over on the left i have
00:20:03a folder for content daily notes inbox ops project systems and wiki so it is a little bit different and
00:20:09slightly more convoluted if you make something that is slightly more convoluted how do we
00:20:14still get the benefits of a quote-unquote map well you can just spell it out to claude and that's
00:20:18what i do inside my claude.md the biggest thing you're going to do inside of your claude.md that
00:20:22lives in your vault is just spell out the vault structure and how it needs to navigate things because
00:20:27with a navigation pattern and a vault structure i've given it a map and instructions of how to read it
00:20:32that's all you got to do and guess what if you're confused with all that you know what you can do
00:20:36you can literally just point claude at your vault you can copy this entire tweet from carpathy and you
00:20:41can say hey can you set this up in accordance with the principles spelled out here now let's talk about
00:20:45how you bring this all together and actually build this sort of ui again completely custom and it is
00:20:49literally just a plugin that claude code creates itself and installs into obsidian on its own and you
00:20:54can see that here under community plugins i have the chase command center of note one other community
00:21:00plugin you definitely want to have enabled when you're doing all this is the hot reload community plugin
00:21:05you're not going to be able to find it at least you haven't been able to find it through the official
00:21:09obsidian community plugin sort of like user interface but you can find it on github just
00:21:14search for hot reload and it's right here just point cloud code at this and it'll install it into obsidian
00:21:19for you so all you have to do is have a conversation with cloud code and say you want to build a
00:21:23custom obsidian plugin that acts as a command center and then you're just going to spell out what you
00:21:29want it to actually look like a great way to do this is to actually use claw design and have it come up
00:21:35with a bunch of different visual mock-ups for you you can see here how many different iterations
00:21:40i sort of went through so i would give it some inspiration like you can literally just go on
00:21:46something like pinterest and look up like you know dashboard or something like command center to get you
00:21:52started drop in those screenshots in here and then say hey i'm trying to come up with a mock-up
00:21:57for obsidian plugin i want it to bring in these certain metrics that certain metrics will depend
00:22:02on what you want to actually see i want it to show my calendar i want to show these skills i want to
00:22:07show it my token burn whatever it is you want and then you're going to say i want you to create five
00:22:12very distinct variations of this and if you do that inside of claw design it will put them all on the
00:22:17same page like this all of them are probably going to be relatively different and then when you find one
00:22:22you like i suggest you then double down on it and have it create multiple variations of that command
00:22:28center so i kind of went like with this clay hardware one at first then i found this cool like background
00:22:33i was like okay what if we put a background inside of here and then had it layered on top so then i started
00:22:38doing variations of that eventually i landed on this one where we kind of have that glass morphism effect
00:22:43so you can kind of see the clouds sticking in through the back here and then if you look closely i had
00:22:49it also add almost like this like star twinkle effect it's pretty subtle so because i didn't want
00:22:55to be overwhelming um but that's what i did i added a little bit of motion so when i click the overview
00:23:00you can see it tick up but again like you can do whatever you want you make this as cool or as lame
00:23:06or as boring as you want to see totally up to you but if you're doing this inside of claw design once
00:23:12you land on a sort of design you want to move forward with you just tell it that and it's going
00:23:17to take all the code it's created for these mockups it's going to put it in a zip file and then you can
00:23:22just take that zip file like you see right here drop it inside of claw code and say hey build the plugin
00:23:28based on this it's going to build it it will install it and then you'll be able to see it inside of
00:23:32obsidian itself if you want to go a step further you can actually enable computer use inside of
00:23:37clawed code and you can have it sort of do all the design iterations while it's taking screenshots
00:23:42of itself so it kind of makes it pretty hands off but that's essentially the process i went through
00:23:47to create mine and so that's the workflow for creating your own obsidian clawed os command
00:23:51center with some voice capability like i said you need to customize this to get the most out of it
00:23:56but if you want my exact setup including the whole voice thing and everything i have that inside
00:24:01of chase a plus and you can find that in the pink comment so as always let me know what you thought
00:24:06i would love to know what sort of command center sort of setups you've been able to create and what you
00:24:11did to customize yours because you can get kind of crazy with this but besides that i'll see you around
커뮤니티 글
아직 글이 없습니다. 이 영상에 대한 첫 번째 글을 작성해 보세요!
이 영상에 대해 글쓰기