21:38Jay E | RoboNuggets · Tutorial
The ARMS Framework: A 4-Part Agentic OS for Claude Code
A creator walks through his own Claude Code command center and the four-layer system underneath it: Skills, Memory, Routines, and Applications.
August 21stThe argument in one line.
An agentic OS pays off when every layer is tuned for cost and self-improvement: skills routed to the cheapest model that works, routines that stay quiet, memory Claude can map in plain text, and apps that move repeat work out of chat.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude Code or the Claude desktop app daily and keep hitting your five-hour or weekly usage limits.
- You already have a pile of skills and scheduled tasks and no idea which ones are expensive or never used.
- Your workspace has grown to thousands of files and Claude is getting slower or worse at finding the right one.
- You want concrete ideas for micro apps, mods, and Chrome extensions that take routine review work out of the chat window.
- You build agent setups for clients and want a teachable structure to organize them around.
SKIP IF…
- You've never written a Claude skill or CLAUDE.md file. Start with a basics tutorial; this assumes you already have a workspace.
- You want a step-by-step build of one specific tool. This is a tour of principles and examples, with the actual prompts living in an external PDF guide.
The full version, fast.
The video argues most agent setups still run on advice from before the Claude 5.5 models, mods, and the Jev decision model, and that the fix is organizing context in four layers called ARMS: Apps, Routines, Memory, Skills. Skills split into database, artifact, and self-triggered types, get routed to the cheapest capable model (Jev routed 4,451 prompts for 18 cents), and improve themselves from feedback. Routines should mostly be passive, capped on spend, and include one that proposes new routines. Memory works as a plain-text map through MEMORY.md, CLAUDE.md and router files, plus disciplined compacting and keeping the prompt cache warm. Apps prefer CLI over API over MCP, extend tools with mods and Chrome extensions, and move recurring review jobs into micro apps built from a reusable component library.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:10
01 · Intro
Claude 5.5 models, Claude mods and the Jev decision model just landed, but most agent setups still run on old advice. Promises four parts with three levels each.
01:10 – 04:03
02 · My Agentic OS v2
Tour of the home dashboard: micro apps, calendar, routines, email triage, a skills deck with model and effort pickers, an artifacts inbox, and a 3D second-brain view of roughly 400,000 files with token-usage views for routines and connected apps.
04:03 – 04:59
03 · The Framework
Introduces ARMS: Apps, Routines, Memory, Skills. The key to any agentic OS is how context and the workspace are organized, learned from the bottom up.
04:59 – 05:12
04 · Guide, send it to your agent
All prompts live in a free PDF guide meant to be handed straight to your agent.
05:12 – 15:35
05 · Part 1 - Skills
Level 1: database, artifact and self-triggered skill types, plus wildcards against overfitting. Level 2: route skills to cheaper models, including with Jev, and watch per-skill spend. Level 3: skills that update themselves from review feedback.
15:35 – 19:45
06 · Part 2 - Routines
Level 1: active versus passive routines. Level 2: capped routines sorted by weekly token spend. Level 3: a Friday routine that reads past sessions and suggests new routines.
19:45 – 31:11
07 · Part 3 - Memory
Level 1: long-term memory as a plain-text map via MEMORY.md, CLAUDE.md, router files and database skills. Level 2: short-term memory, compacting with purpose and keeping the prompt cache warm. Level 3: self-evolving memory via /note, /calibrate and suggested fixes.
31:11 – 37:38
08 · Part 4 - Apps
Level 1: find connectors in the order official, CLI, API, community MCP. Level 2: mods for Claude Code and custom Chrome extensions. Level 3: micro apps for recurring jobs, built from a reusable UI component library.
37:38 – 38:09
09 · Wrap up
Recap of the ARMS framework and a pointer to the PDF guide.
Lines worth screenshotting.
- A skill doesn't have to be a procedure; it can be a database of knowledge, like every principle and example of how you write.
- Reference-heavy skills overfit, so have them return 8 options from your references and 2 wildcards that ignore them entirely.
- HTML is the one output format that covers infographics, slide decks and PDFs, and a brand book in HTML keeps every artifact consistent.
- A grammar check doesn't need a frontier model; clear-cut skills should run on Haiku by default.
- A fast classification model routed 4,451 prompts in about 0.6 seconds each for 18 cents total, and sent roughly 4 in 10 to Haiku.
- Fifty routines that each hand you an output means fifty things to review; most routines should stay silent unless they find something.
- Ask Claude to cap an expensive scheduled task at a dollar figure, like $4 a week instead of $10, and it will usually cut the fat.
- The best routines are jobs you already do by hand, so schedule one that reads your past sessions and suggests what to automate.
- A second brain is only worth it if Claude finds things faster or cheaper; the 3D graph is optional.
- Write CLAUDE.md as a map of your workspace that routes to department files, not as a rulebook.
- A 420,000-file workspace can run on plain-text router files with no vector embeddings, and Anthropic reportedly found the same internally.
- Claude's prompt cache goes cold after about an hour idle, and re-sending one message to a 600,000-token Opus chat cost roughly $5 at API rates.
- Having idle sessions ping themselves hourly for four hours would have saved an estimated $437 in API costs over 30 days.
- When auto-compacting, tell Claude what to keep, or it decides for you and may drop the one detail you needed.
- If you correct Claude on the same thing twice, your memory system missed something.
- CLIs are the lightest way to connect an app to Claude, APIs are heavier, and MCP servers cost context on every new chat even when unused.
- Once a task is routine, a chat window is often the wrong interface; a small local app for reviewing that work beats scrolling the conversation.
Tune every layer of your agent for cost.
WHAT TO LEARN
A strong agent setup isn't more tools; it's skills, routines, memory and apps each designed so Claude spends fewer tokens and gets a little better every time you use it.
02My Agentic OS v2
- Build one home view for your agent's outputs, schedules and artifacts so you can find past work yourself instead of paying tokens to ask Claude.
- Track token spend per routine and per connected app so you can optimize the expensive ones and retire the ones you never use.
03The Framework
- Organize the workspace in four layers, apps, routines, memory and skills, and set them up from the bottom so each layer builds on the one below.
05Part 1 - Skills
- Treat skills as knowledge stores, not just procedures, and add a default wildcard quota so reference-driven output keeps producing fresh ideas.
- Route clear-cut skills like spell checks to Haiku or a fast classifier; most work never needed the biggest model and nobody notices the downgrade.
- Wire your review tool so feedback on an output also updates the skill that produced it, turning every correction into a permanent improvement.
06Part 2 - Routines
- Make most scheduled tasks passive so they only alert you when something needs attention, otherwise fifty routines become fifty daily reviews.
- Put a dollar cap on any routine that burns heavy tokens each week and ask Claude to rewrite it to fit the budget.
- Schedule a weekly routine that reads your recent sessions and proposes new automations for the jobs you keep repeating by hand.
07Part 3 - Memory
- Write CLAUDE.md as a map that routes to department files; plain text navigation is enough even at hundreds of thousands of files.
- Tell Claude what to keep when compacting, and keep long chats' prompt cache warm, since a cold cache re-bills the whole conversation.
- Park mid-task improvement ideas in a per-skill backlog and apply them in a batch when your usage limit is about to reset.
08Part 4 - Apps
- When connecting an app, look for an official connector, then a CLI, then an API, and only then a community MCP, because MCP tool lists load into every chat.
- Build small mods or Chrome extensions for tools you live in; they run as code and add capability without spending tokens.
- Move recurring review work out of chat into small local apps, and give Claude a reusable component library so it never rebuilds the same UI twice.
Terms worth knowing.
- Agentic OS
- A personal workspace and dashboard built around an AI agent, organizing its skills, scheduled tasks, files and connected apps so the agent can do ongoing work with little hand-holding.
- Skill
- A saved set of instructions or knowledge, stored as a file, that Claude loads and follows when you call it by name, usually with a slash command.
- Jev
- A fast, cheap decision (classification) model released alongside the Claude 5.5 family, useful for quick choices like which model a prompt should be routed to.
- Model routing
- Sending each task to the smallest model that can handle it, such as Haiku for spell checks, instead of running everything on the most expensive model.
- Overfitting
- A machine learning term for a system that sticks so closely to its training examples that it stops producing anything new. Here it describes skills that only echo their reference files.
- claude -p
- Claude Code's non-interactive print mode, which runs a prompt in the background without opening a chat, so a skill can be triggered from a button or script.
- Passive routine
- A scheduled task that runs quietly and only alerts you when it finds something worth acting on, as opposed to an active routine that delivers an output every time.
- Context window
- The amount of text an AI model can hold in one conversation at once. The Claude 5.5 models hold about 1 million tokens.
- Context rot
- The drop in an AI model's ability to recall details as a conversation fills up its context window.
- Prompt cache
- A short-lived store of a conversation that lets the model re-read it at a fraction of the normal token price. It expires after a period of inactivity.
- Compacting
- Summarizing a long conversation into a shorter one so it fits back in the context window. In Claude Code it happens automatically or via the /compact command.
- MCP
- Model Context Protocol, a standard for connecting AI agents to outside apps and data. Each connected server's tool list loads into every new chat.
- CLI
- Command line interface: a text command tool an agent can run directly and get a short answer back from, which keeps token use low.
- Claude Code mod
- A small add-on that changes how the Claude Code app looks or behaves, similar to a browser extension, and runs as code without spending tokens.
Things they pointed at.
03:37toolClaude Opus 5.5, Sonnet 5.5, Haiku 5.5
12:24toolJev (Anthropic decision model)
10:30toolclaude -p (Claude Code print mode)
24:30toolAGENTS.md support in Claude Code
37:20talkKatie Dill (Stripe) talk on a shared UI component library
Lines you could clip.
00:00
“Something big just changed with Claude recently. Because not only did we get the seriously capable family of 5.5 models, but we also got entirely new tools.”
07:26
“You can have it pitch eight of those according to your references through your database skill. And two of them are wildcards, which doesn't use your references at all.”
12:24
“All 4,000 plus of those together cost me only 18 cents. And about four in 10 of them went to Haiku instead of a bigger model.”
16:30
“Just imagine once you have 50 routines, then you'll have like 50 things to review every single week or even every single day.”
24:51
“Even in their workspace, so this is Anthropic with probably millions of files already, they're also finding that plain text is much better versus previous complex setups of vector embeddings.”
28:10
“If I were to send a message here, it would recache around 600,000 tokens. And given the model, which is Opus, that would account for around $5 in API credits.”
29:50
“If you have to correct Claude on the same thing twice, then your memory system probably missed something.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
Something big just changed with Claude recently. Because not only did we get the seriously capable family of 5 .5 models, but we also got entirely new tools. Like for example, Claude mods that literally let you change how the Claude app itself behaves.
And even the decision model Jev unlocking entirely new use cases because of how fast and cheap it is. But most people's agent setups are still built from outdated advice from a few months ago. So today I'll share with you exactly how I set up my agentic operating system and how I take advantage of these new tools to work more effectively with agents without draining my usage to the point that I can even create videos like these all edited with Claude while still running a business on the side and even having time for my personal life.
This will be a longer than usual video, but if you watch through it, I'm sure you'll pick up some nuggets and foundational principles on how to make your systems run faster, cost less, and get more done, especially as we head to 2027. We'll go through it in four simple parts with three up -to -date levels each so that by the end, you'll be using agents better than 99 % of people.
And if you're new, my name is Jay. I spent over a decade working with brands you probably know, have been in AI since my master's in data science, and now I'm leading our AI business in one of the largest AI communities globally. So what you're seeing here is my agentic operating system.
I use mainly the Cloud 5 .5 models with it, but of course you can run it with any other model. And obviously the aesthetics are just us having a little bit of fun with it, but it's really the architecture and UI underpinning this that makes it valuable to me personally. Because from this home dashboard, I can access the micro apps I've created to help me work better with my agents, such as this content system that lets Claude generate and also show images in the videos, which I pretty much use to replace Six Field.
A video editing app where my agent shows me the animations it's done so I can give comments to it much easier. An Instagram carousel app where Claude parks posts that I can just review and approve. This simple recommendation app where my agent tables improvements to our system that I can just apply at the end of the week if I have extra usage remaining.
It does serve as a homepage, so I also see what's next in my calendar, the next scheduled routines that I have, emails, the cloud things I should take note of or reply to, and also some skills that make sense to trigger on demand while letting me select the model and effort directly from here. I also make a lot of artifacts as part of my work.
And so Claude just parks them here when it's done. So let's say I want to find a deck that I made last week related to a client. Let's say beta shares.
Then I can just easily find that here without needing to ask Claude and spend tokens. And finally, if I click through here, that just shows me my second brain system, which I guess is just the industry's fancy term for the files in my workspace, for which apparently I now have around 400 ,000 files in here. So if I zoom in, you can see each of these nodes represent a file.
And if we get out of the fancy 3D view, you can see a lot of these files are interconnected just so my agent can find what it needs faster and easier. And if I need to search for a file, like let's say I want to check or show the contents of one of my skills, I can do that here, which is just much faster and definitely a nicer UI than my default file explorer.
Finally, I can also have a timeline view of which files, skills, and routines my agent last edited, a routines tab that shows me my scheduled tasks in this week view, and more importantly, which ones are consuming a lot of tokens so that we can optimize them. And this also gives me a view of the skills and applications I have cloud connected to, to see which ones are consuming the most tokens so that we can optimize them and which ones we don't use at all so that we can retire them.
Now that is a lot of information and you definitely don't need every single thing or feature that i mentioned just now but the last time that i took you guys through my agentic os that seemed to help you guys a lot in terms of thinking of your own systems and so that tells me that you might be interested in an update especially since in the past days not only did we get the cloud 5 .5 models which are leaps better than their previous counterparts but we also got new powerful tools like the new decision model called jev and cloud mods which can literally modify the Cloud Code desktop app.
So I'll take you through those things and everything else you need to know. And my intention is by the end of this session, you will also get novel ideas for your own agentic operating system so that you can build something similar to what you see here, either for your business or to even resell to clients. Now, there are many ways to design an agentic OS, but in my view, the key is always how you organize your context and your workspace.
And at least the way that I think about it and teach it in our community is what I call the ARMS framework. And it's quite easy to remember because it's like giving your agent ARMS to reach out to your workspace. It stands for apps, routines, memory, and skills.
And the point being is that if you set these four up the right way, then you'll be ahead of 99 % of people when it comes to using AI agents. And the best way to learn it, in my view, is from the bottom up. So the plan for this session is I'll just go through each one from the bottom up with three up -to -date levels for using it, given all the updates and improvements to the models that we've seen in the past few weeks.
That way, you can just skip ahead freely if there's things that you already know. But given the depth of what I plan to cover, for sure, you'll pick up a few nuggets along the way that you probably haven't known yet. And by the end of this video, if you watch the whole thing, then you'll be able to level up your adjunct skills.
to the point that you can also build an OS like the one I have here. Now, before we dive in, just know that all of the resources and prompts that I'll say here, I put them all in this PDF guide, and you can just send that whole thing to your AI agent and pick out the stuff that would be useful to you. So that's just down below if you need it.
So first, let's start with skills, which are these saved instructions that Cloud follows whenever you call them. And level one here is knowing your skill types. Now, these skill types, they do vary depending on the work that you do.
But just to give you an idea of how to think about them, here are three examples of skill types that I have observed in our own work, which may be applicable to you as well. The first one are what I call database skills. Because when you talk about skills, usually you think of them as SOPs, right?
Standard operating procedures where you have clear -cut steps for the agent to follow. But you can actually also think about skills as holders of knowledge. And just to give you an example, I have this skill, which if I just go and find it, it is scald slash tone.
And it essentially holds all the knowledge and principles around how I write. Because in my work, I write a lot of emails. I write a lot of posts for our community.
And so it just makes sense to have a database of examples, principles, and guidelines of my own tone of voice so that my agents know exactly how to write first drafts for a lot of things that I write. And so this skill, whenever I call it in any cloud chat, instantly my agent would know all of the principles we have in our tone of voice because of this skill.
the great thing is you can also do this for non -text databases so for example because i do a lot of stuff on youtube i also have a database of thumbnails in here which is all associated with this skill called slash thumb and that just includes all the references and file paths to thumbnails that we have shipped and even some inspiration ones that would be useful for our videos and because thumbnails need to be visual and i need to look at it then i also have this artifact which holds the latest references that that skill has so the point being is that if you produce or or reference the same type of output throughout the week, it is probably a good idea for you to have a database skill that you can just have your agents refer to.
Now, one call out that I need to make when it comes to using database skills is the problem of overfitting. And overfitting, it's one of the very first concepts that you learn if you ever do machine learning or data science. And it also applies to a degree here.
Because if you give your agent references, the problem with that sometimes is that it will only stick to those references and you actually prevent it from giving you fresh ideas, which are sometimes good. And so you overfit your system to just this one corpus of references that you give your agent. And so to prevent against that, one idea is to have your database skills introduce what I call wildcards by default, which let's say Claude will pitch you 10 titles for your article.
You can have it pitch eight of those according to your references through your database skill. And two of them are wildcards, which doesn't use your references at all. And so that way you can get fresh ideas every time.
And some of them, if they are good, you can just ask Claude to include those as part of your database skills as new style and tonality references from then on. To give an example, here on the screen, you'll see on the left some of the thumbnails that we have in our references. And when I asked Claude to create wildcards, here are some examples at the right.
I also do this when it comes to video animations. So we have this one style for Robo Nuggets at the left, which is more tech focused, I would say. But on the right, I actually sometimes ask Claude to create wildcard styles.
And some of those have been quite interesting to experiment with. And that is how I get new styles and new ideas by essentially just letting Claude cook and suggest new novel ideas, which I can just pick out to see which ones are good or not. In fact, I use this idea so much that I actually have a shortcut for slash wildcard.
And this is just a short, really short skill that just defines what I mean when I say or when my other skills say the cloud needs to generate wildcards. So that is something that you can adopt if you want fresh ideas and you want to avoid overfitting from your data set. Now, the other skill type that you can think about are artifact skills.
And I use artifact skills because a lot of the times the output that I want actually isn't optimized to be presented in just the chat interface in the chat window of your cloud. So let's say you want to create an infographic, a slide deck or a PDF. What I actually found is that a lot of these artifacts can be created using just one format that agents are really, really good at, which is HTML.
That stands for hypertext markup language. And that's basically the coding language that powers a lot of the websites that you see in the internet. Now, what I do is I have this artifact skill called slash robo, which is also my design system.
And it's what I use to make these PDF guides for these videos, for example. And the secret is that those PDFs that you see, they are actually created with HTML first before Claude exports them as a PDF. And the reason why those PDFs and pretty much every artifact that I make is so consistent in design is because this skill is associated with our brand book which is also in html and just lays out the guidelines of the colors that we use fonts that we use and so on and so if you don't have something like this yet for your business I highly recommend creating one so that you can avoid all those vibe coded sloppy looks that your AI agent will usually default to.
Now, another pro tip is to have your AI agent build you an inbox specifically for your artifacts. So for example, similar to what I showed earlier, this ring at the top is sort of my artifacts inbox. And whenever my cloud creates an artifact, it just puts it in here.
And I can also search through these artifacts if I do so wish. And that just makes it easier for me to. up older artifacts like this one for example was created two days ago so instead of asking claude to open and find it for me which costs a lot of tokens sometimes i can just type it in here and immediately find it and the last skill type that i use that you can think about are self -triggered skills these are skills that sometimes make sense to trigger outside of the chat To give one example, here in my skills deck, I have all these self -triggered skills that I have.
And one of the ones that I use a lot is this slash cleanup skill. And I just run that when I find my computer is slowing down because of some temporary processes and files on my computer that maybe Claude has left. And if I want to trigger this, I just click on this instead of asking Claude to do it for me.
What you can also do in your OS is to give yourself a capability to switch the model as well as the effort level that you want to run that skill in. And if you want to set up your self -triggered skills, you only need to be aware of this little known Cloud feature called Cloud -P, which essentially just lets Cloud be activated directly by sending a prompt to it in the background without having to open up your chat.
And if you want to try this out for yourself, I'll just put this prompt on the screen for you to get started. You can screenshot this, but again, you can just find it in that PDF guide down below. And you can see once that skill finishes, I just have it give me this simple report of what it fixed and its output.
So if level one is being conscious of skill types, which make you more effective with your output, level two is actually optimizing your skills to save you money and time. And this is about model routing, which is essentially just routing each skill to the right model that you need to use. Because a common mistake that I see a lot of beginners make is they create skills, maybe they forget about them, and they don't even know which skills they use the most.
But the problem is some of those skills, they may be running it in bigger, pricier models than they need. But a lot of them can actually run on lower models like Sonnet or even Haiku, especially the ones where there's already a clear -cut process embedded in the skill. So for example, I have this simple skill called slash grammar and that only checks for spelling.
So every time I call it, it just runs on Haiku and answers much faster and much cheaper. And what I actually started to do and started to experiment on is to use Jev, which is this decision model that is really fast. I think up to 200 times faster and 400 times cheaper versus frontier models.
But basically, since it's a classification model, it can quickly decide for you on several things, including which model to route your prompt to. And just to give you some data in my workspace, it made apparently 4 ,451 of these goals in the past weeks. Each one of those model routing took only 0 .6 seconds, so almost no time at all.
And all 4 ,000 plus of those together cost me only 18 cents. And about four in 10 of them went to Haiku instead of a bigger model, which saved me a lot of tokens in usage. And at least from my perspective, I never felt the pain of being routed to a lower model in my work.
And so that is a simple good test if there's... active model routing to Jeff would be right for you. And again, if you want to get started with that and with a lot of the techniques that I'll show here, all you need is to just be aware of it and have a prompt similar to what I have here on the screen so that you can work with your agent with setting that up.
Now on this same thread, what's actually useful is if you have a dashboard like this that shows you what are your top use skills. And more importantly, if you're seeing that you're draining your usage quite a lot, you can actually inspect which of your skills is spending the most and see if they need to be optimized. For example, the one that I use the most apparently is this slash animate skill because that is the one that I use to create a lot of these videos.
And so that is quite token intensive. But with a list like this, then you can quickly eyeball just which skills may be costing you money that you probably don't remember using. So that might be cost for optimizing it.
Now, level three, when it comes to skills, is designing them so that they are self -improving. Meaning that the more that you use these skills and you give feedback to your cloud, you need to make sure you have a system so that these skills self -improve whenever you use them to give one example of how i do it whenever i create videos with animations i actually use this application which we built for our community called rubric editor and essentially this just allows me to add in comments like for example this one you can see i pinned a comment here and i just gave feedback to it i think through voice transcription which apart from the comment what makes this rubric editor effective is that it also includes the timestamp as well as the location of that pin and then when i'm done giving comments and essentially directing this whole video I just copy all the comments for all clips to my clipboard and then send that to Claude.
But what I wanted to show here is that a part of this custom application is this toggle where I can teach my slash cut skill and my slash animate skill whenever I give comments. And when I have that toggled on, you can see if I paste those comments on my clipboard or to cloud rather, you can see it automatically attaches this part of the prompt where it's asking my agent as well to teach or update my slash cut skill as well as my slash animate skill.
And that way I've made a system so that my skills will self -improve every time I give feedback to it. And so the main takeaway for you here is that if you know what skills you use a lot through a view like this, then you can also think about how you can automate the self -improvement of that skill so that the more that you use it, you polish it more and more so that it gives you better and better quality output every time.
And really, you can just think of this as a really simple version of reinforcement learning, which is anyway what the top AI labs use in order to make their models better. So if you take that learning and actually try to apply it to your own operating system, then you'll be able to become more effective when using skills as well.
So now let's go to routines, which are essentially just tasks that Claude runs on a schedule. And the first level here is knowing first what are active routines versus passive routines. And just to differentiate between the two, active routines are one that does a job for you and provides you an output that you can review, while a passive routine just runs in the background and only pings you or disturbs you when it finds something worth flagging.
For example, if I go to my cloud app and go to my scheduled routines in here, you can see I have this active routine, which is a daily inbox digest that just reads my emails and gives me a digest over a telegram every day. And this is an active routine because it hands me an output for review. But the thing is, you can actually have routines where you don't necessarily need an output every time.
To give one example that you can apply for your business, what I have here is a routine that's called a brand watch. And essentially, every week, it just checks the internet for any websites or domains that are using the RoboNuggets brand without our approval, and then it just pings us and alerts us if there is.
But if there's none, then it's not going to give us an output at all, which is important because the more routines that you run, if each and every one of those actually are active routines and are giving you an output every time, then just imagine once you have 50 routines, then you'll have like 50 things to review every single week or even every single day, which is just not optimal.
And so once you understand how passive routines work, you can now start to become more creative with your routines. So for example, you can have one that scouts domain names or TLDs that are relevant for your business and just paints you if a good one becomes available. You can have one that just looks after your machine.
Like for example, if you want to clear old files or just let Claude flag if any applications are slowing down your device, those are all good candidates for passive routines. And now if level one is knowing about passive or quiet routines, then level two becomes important, which is all about knowing about capped routines.
Because every scheduled task that your agent runs obviously costs you something, either plan usage or API credits, right? But once you've got a lot of them, it's actually quite easy sometimes to lose track of the expensive ones. And so here in my What I did here personally is to have this routines tab, which just shows my scheduled tasks across the week.
And more importantly, I can sort them based on how much tokens they are using up per week so that I can monitor and see if any of them are draining my usage without really providing value to our business. And also through this view, you can also see which routines run on which time, if in case that's something that you want to manage.
And so when you find that a particular routine is actually costing you a lot of tokens and money, all you really need to do is to ask Cloud to optimize that scheduled task and see if you can bring down the token usage to more manageable levels. And one way by which you can do that is by setting a cap. to that particular routine.
So for example, if this routine is costing you around 18 million tokens or $10 in API credits per week, you can just ask Claude if we can cap that at, let's say, $4 per week in API credits. And a lot of the times it'll be able to do that and ship out a lot of the fat in your scheduled task. And finally, the third level to this is to have a routine that suggests routines.
Because really, at the end of the day, the best routines are actually just jobs that you already do by hand. But the problem is sometimes you don't notice that you actually just repeat them.
So for example, for my workspace, you can see every Friday morning I have this scheduled task, which essentially just looks back through my sessions and spots the jobs that I keep doing myself. And so just routines in this micro app that I made called Rubric Recos. And this Rubric Recos micro app is just, again, one of the toolkits that we made for our community.
And how I use it is that at the end of the week before my usage resets, I usually just look at this and see. If there are particular improvement points that Claude thinks would be useful or beneficial to us, and I can just add them to cart if it's reasonable. And once those have been finalized, I sort of just check out and copy that for my agent and then send it to Claude so that it applies it for my own setup.
And you can see it's a browser -based app that runs on my local host, which essentially means it just runs on my computer. And if I go to routines in here, it just gives me a list of ideas of routines and tasks that might be worth automating. So this way, I let Claude suggest ways by which we can automate our setup.
And so every week, going back to that self -improvement cycle that we talked about, your operating system just gets better and better the more that you use it. And again, all you need to get started is a prompt like this one, which you can just screenshot and send to your agent. So now let's go to the next layer of the framework, which is memory.
Now memory, simply put, is everything that Claude knows about you and your work. And just as importantly, how fast and how cheaply it can retrieve it. Because the more that you use your agents, the more files eventually pile up.
And searching all of them every time can be slow and can burn through your usage. And in my view, when we talk about memory, you can split it into three levels. And the first level is long -term memory.
which is essentially everything that is saved in your workspace. And when people say second brain systems nowadays, that's really just a fancy term for those files and those data that is contained in your workspace, which is probably just a folder in your computer. Now, my second brain system, or more specifically, the visual to my second brain system would be this.
And from here, you can see all of the applications I have cloud connected to, my routines, my individual files split by departments, and also the skills there in the center. that is all routed to by my claw .md here in the center. Now, to be clear, you don't really need a fancy map or visual to have an effective second brain system, really.
In fact, this is useful to me because I do a lot of AI education and training. So if I need to show my files and skills like what I've been doing in this video, it's just much more understandable and also faster for me to search for them here versus, let's say, I'm using my normal file explorer, which is just slower and is less engaging.
And this folder, by the way, that is the exact same files and folders that are represented here also in this map. So if you want a nice visual, if it's relevant for your work, or you think it's better or a bit more fun, then go ahead. Because honestly, with today's models, Claude can create something like this for you pretty quickly.
But I'd argue that the main value of all second brain systems really is if it gives you either or both of two things. Either you have a system that gives you a faster speed of retrieval or a cheaper cost of retrieval. Or simply put, can Claude quickly find your stuff?
without burning through your usage limits. Especially as you use agents more and more, you can see my workspace here has something like 420 ,000 files as of the moment. So you definitely would need to have a system of retrieval for this if it balloons to this size.
And from what I've found, really my setup and my system here is quite simple, at least simpler than most, because there's really just three layers here that I use. And so I'll go through them and see if that works for you as well, because I just find that even with that big workspace, just having these three simple layers of long -term memory is already enough for me.
And the core idea that cuts across all these three is that we are basically just giving Claude a map. of our entire second brain or workspace. So for example, the very first layer here, which is default for a lot of cloud code, is memory .md.
Now memory .md, by default, cloud would already keep this file for you. And by default, it would just give you one short line for each thing it remembers about you. And how cloud code is set up is that it actually loads this file for you at the start of every chat.
So cloud has already read through this memory .md when you start your first prompt. So you can think of it like a library's card catalog, for example. where instead of Cloud having to find and walk every shelf, it would check the catalog first for a map of your entire library.
And you can see for my workspace, this is my memory .md. And the reason why it is connected to all of these other nodes across my different departments, like in content, in community, and so on, is because what I'm trying to do is take advantage of the fact that it gets preloaded on every session. And so that way it has linkages to each and every part of my workspace, which just helps Cloud navigate the entire map better.
Now for a small starting workspace or second brain, that is usually good enough. But as your workspace grows and you're finding that Claude is having difficulty retrieving information for you, what you can do is to use the next layer, which is your Claude .md. Claude .md, that is a file of instructions that Claude reads at the start of every chat.
And although the previous guideline to using Claude .md is to set it up as a sort of a handbook of how things run within your workspace, I actually set mine up as a map of the whole workspace of the entire second brain. And the reason why I have it here in the center is because I actually have it route to the different departments of my work.
Like if we're operating on content, then these are the files that it can work on mainly. If we're operating on personal stuff, then these are those files. And by the way, just to mention, Cloud Code actually now reads agents .md too.
So if your workspace already has one for other AI tools like Codex, then you can just have one central agents .md as your core operating rules. Now, like I said, as part of my call .md, I have there a map of my whole workspace. Then what you can do is have each department have its own file that routes against.
So for example, I have a business .md that routes to the sub files and folders within it. And I find that that is usually enough when working with agents because today's 5 .5 models are really smart enough in order to just use plain text to find what it needs. So if you're more advanced, you can see I'm not even using any vector embeddings, which I may do that if I feel some sort of pain with retrieval.
But as of now, it works for me. My usage limits don't get drained up whenever I use Claude. And so there's no cost to overhaul or upgrade this system, at least as of yet.
And in fact, just to mention, Boris Cherny, the creator of Cloud Code and who leads the team now, he also mentioned that even in their workspace, so this is Anthropic with probably millions of files already, they're also finding that plain text is much better versus previous complex setups of vector embeddings and RAC databases, what was popular just a while back.
So if it works for them, then most likely it will work for your business too. And the third layer, which is just a supplementary, is obviously skills. Because as we mentioned earlier, you can actually have database skills, right?
So these database skills, it just tells Cloud where your information lives. So you can literally have a skill called slash name of business, then data. And that skill, whenever you reference it in the chat, that just points to where the up -to -date information or even dashboards of that particular business or client is in.
And the skill .md of these database skills, they also adopt the same principle where they are routers so that when Claude reads them, it can actually infer where to find the information you're asking for, depending on your prompt. But the point being that you need to utilize each of these three layers and just make your whole second brain searchable.
That's number one. CleanText turned out to work really well for these LLMs. And if you have a lot of metadata, especially for visual assets, that just makes it easier for your agent to search through your entire second brain.
And if you need a prompt to get started with everything I covered just now, then you can try it out using this one, which I also included in that guide below. Now, if level one is long -term memory, level two is something that I often don't see being talked about as much, and it is Claude's short -term memory, which is what Claude can see in your chat to it in that particular session.
Now, technically, the shorter memory, it's called the context window, which is essentially Cloud's working memory for this one conversation. And the new Cloud 5 .5 models, they have a 1 million token context window. And just to put that in perspective, that's about the first four Harry Potter books in a single chat.
And when it comes to shorter memory, i .e. the context window, there's two problems that usually arises. One is the fuller that the chat gets, the worse that Cloud gets at recalling what's inside it.
And Tropic actually has a name for this. They call it context rut. You may have heard of it.
And the second problem is that everything in that window, every prompt that you send to Cloud, that actually gets reread every time. And so a bloated chat also costs you more and drains your usage faster. And so to address each one, the first step is to compact with a purpose.
So when a chat gets long and you reach that 1 million token limit, Cloud will automatically compact your chat. which is essentially just it's sending a prompt to itself to summarize the chat so far. But what's important here is that you can actually tell it what to keep.
Like for example, you can type slash compact and just keep the business decisions or the templates that we've created as part of this chat. Because if you don't do that, then cloud will just decide what matters and it might drop the one detail that you might actually need. And if you want to guide cloud with compacting with a purpose, meaning you give it some principles of what to always keep whenever compacting, then that's just one simple prompt away and cloud will probably amend its cloud .md.
reflect your wishes. And the second tip with regard to that rereading every conversation, every prompt, is to watch what we call your cache. And the prompt cache, that's essentially Claude keeping what's in your chat in its shorter memory so that when you send a prompt, when it has to reread that same whole conversation, that only costs a small fraction of the normal price.
But if you leave a certain chat idle for about an hour, that cache goes cold and your next message would pay to load that whole chat. all over again. To give you an example, in one big chat of mine, you can see here this chat has been idle for 11 hours and 30 minutes.
And if I were to send a message here, it would recache around 600 ,000 tokens. And given the model, which is Opus, that would account for around $5 in API credits, which right now, since I still have my usage limits, it will just drain this usage by quite a lot. And by the way, this is a mod that I created, which if you're new, mods are essentially sort of like Chrome extensions, but for your cloud code.
So it adds capabilities like these that don't cost any tokens. And you can see in my operating system, I consider the prompt cache so important that apart from my seven day and five hour usage limits, I make sure to always keep that visible to me.
And to give another session example, you can see this one, the prompt cache is still warm. So even if I send a message in here, I won't have to pay the full price of these 186K tokens that are currently in this Claude's short term memory or context window. And you can see here, this chat has been idle for around nine minutes.
So if I wait for around 51 minutes more, this whole chat will get cold. And so the other thing that I do is I actually have a directive for each of my sessions to self ping themselves for the next four hours, around every hour. way I can extend my one hour limit of the cash expiring into four hours.
And so that's something that you can do as well. Or you can also even run an analysis, like have Claude look at the last 30 days of your chats, how much would it have saved you. So I did this analysis myself, and it concluded that if I extend my prompt caching duration through this self -pinging system, then it would have saved about $437 at API prices.
And so you can try that yourself if it makes sense for you. And to do that, again, you can just send this prompt on the screen. And also, the usage mode that I have here, I'm open sourcing it.
So you can just grab that down below if you need it. And finally, level three is self -evolving memory, where you want to make sure that your setup gets better every time that you use it. Because if you have to correct Claude on the same thing twice, then your memory system probably missed something.
And personally, I use three things for this. The first one is a backlog of fixes. And I just have this simple skill.
It's called slash note. And when I notice mid -job, let's say that a skill or an application should do better or become better, usually fixing it right then and there doesn't really make sense because it would lose your thread or confuse your agent. So I just type slash note and give my feedback there and use this haiku to just park that in a file just for that skill.
And whenever I'm ready or my usage limits are about to reset, I just ask Cloud to go to that list and apply every single fix. And this works for me because throughout my work in my operating system, I usually think of new ideas that I think I would want Claude to do next, but just not right now. The second thing that I use is a skill called calibrate.
Now by default, Claude code already saves some memories on its own when it thinks that something's worth keeping. But sometimes if a piece of feedback is really important, you usually don't want to leave that to chance. So when you want to teach Cloud something, I just type in slash calibrate and it saves my feedback, links it to related memories, and even fixes the skill that it belongs to.
And the third one, which we've already slightly covered, is letting Cloud suggest fixes for your memory too. So like I mentioned, I have this micro app called Rubric Recos that just routinely provides me with ways to improve my setup. And this also has a section for memory.
And so I just pick the ones that I like here and ask Cloud to improve that part of our operating system. Now let's go to the last layer, which is apps, which is all external applications that Claude can reach outside the chat and any applications that you actually build yourself. And level one, of course, is connecting to the apps that you use.
But the main unlock here is you can connect to even applications that have no official connector. So even if you don't find them in a Claude desktop app, you can actually have Claude find it yourself. So instead of digging through settings, what I usually do if I need to connect something is to let Claude search it for me.
I use this skill called slash search connectors. which for example i've used that in my workspace to connect to things like adobe premiere pro which doesn't have an official connector even personal apps you use like my fitness pal again which doesn't have an official connector and what this skill does basically is it scours the internet to check for connectors in a specific order so first of course it looks for the application makers official connectors first then if that's not available sometimes the open source community actually publish a cli which is a command line interface tool which cloud can run directly then it checks if there's options for the app applications API and last is it checks if there's a community MCP.
Now, the reason that order matters is because generally CLIs would be your lightest option to give to LLMs because at their core, how they work is Cloud can just type one short command and just gets back the answer that it asked for. So it's really lightweight and it helps save up your usage too. APIs, application programming interfaces, usually you get this by getting an API key from a certain website and giving that to Cloud.
When Cloud uses those, they are a step heavier because the answer that comes back whenever Cloud calls that API is often long with lots detail that cloud doesn't actually need and so because cloud has to read all of it every call that uses up some more of your usage and finally you have mcp connectors which among them generally would be the heaviest because every time that you start a chat cloud actually loads your full list of what each mcp can do so even if you never use it whenever you start a new session it uses up your space before you've even asked anything so i just find that it's good to keep mcp connectors light and that's why generally we go in that order Now, the second level, which is pretty new, is mods.
So mods changed Cloud Code itself. And I'd argue that you can take the concept of mods, not just to Cloud Code, but to your other applications as well. But first, just to define it, a mod, essentially, that's just a small add -on, like a Chrome extension for Cloud Code, which can change how Cloud Code looks and works.
So a good example of that is my usage RoboBar here, which I've already demonstrated earlier. So it shows my weekly and five -hour limits, as well as important info about my prompt cache. And I've also done another video which got a lot of attention on some more Cloud mods that you can actually use for setup.
So you can check that out. I'll link it somewhere in this video in description as well. But like I said, you actually don't need to be limited to just modifying Cloud code.
And one good example is apart from Cloud, I bet that you have another application that you always have open and you probably have open right now. And that is your browser. So even in the era of agentic AI, we still use our Chrome browser quite a lot.
At least for me, that is where I review a lot of the artifacts that Claude makes and even my whole agentic OS dashboard that is also a browser based application. And the great thing about Google Chrome is that the concept of mods has always been there to Chrome extensions. So you can actually build Chrome extensions to shape the way that you work, especially in websites and browser -based applications that you use on the daily.
I have this skill called slash Chromeify, and that just encodes some best practices on how to build Chrome extensions. And to give some examples of what I have here, I have this Chrome extension called RoboCapture, which I use that whenever I find a part of a website that I like front -end wise. I then use that extension, click that part, and it just grabs the underlying code of how that part is built.
Then I just hand it to Claude to reverse engineer how it works and builds that exact look instead of me just passing on a screenshot, which a lot of people do. Another extension that I've made because I produce a lot of educational content on YouTube, and whenever I see a good title or thumbnail that I like, I can actually right -click on that on my browser, pick what to save, either the thumbnail or the title or both.
And that just goes into my database skill for feature reference. So that way I never leave YouTube. The Chrome extension also runs via code, so it doesn't cost tokens.
But the next time that I need to ID it on a thumbnail or title, then Cloud already has my references to work from. And so the main takeaway is if you're using Chrome, build Chrome extensions. You can use Cloud to build them and just go to Chrome extensions, toggle on developer mode, and you can then load your own extensions straight from your computer.
And so building on this, level three is now creating a micro application for the big jobs or tasks that you do. Because once a task is something that you do routinely, like for example, for me, when I need to review a whole video being cut by my agent, it makes sense to think about if a chat window where you talk with your agent is still the best and most efficient way to work with your agent to get to the output that you want.
And so now I find that a lot of my work has now moved out of the chat and into apps like these that were all created by Claude and I can just access from my local device. I already mentioned about Rubric Editor, where I review every cut with the actual clip. And also I can provide comments and direct much easier versus having to type it in chat.
I also made a small app recently because we're trying out carousels on Instagram. And here what's good is I can quickly view and review each of the carousels that Claude is making for me. It's much easier to manage and review versus having to look for it in chat.
I also have this micro app called Generations, which is just all of the images and videos I've asked Claude to generate. And I do that via a more direct connection to the image and video model. So not using Higgs field.
And I actually made a video on this separately. So I'll just link that somewhere in this video if you want to learn how to set this up as well. And so I encourage you, you can actually try building these applications yourself as well, especially since Claude has been really effective with building these out with just a few prompts.
Now, a good tip if you're going down this route is to actually start to build your reusable UI components library. So you can see here as part of my slash robo skill, we have this component library where Claude can just pick through some of these pre -built components so that it doesn't have to remake each one. And this is just important so that you don't have to spend tokens rebuilding everything from scratch, especially if it matters for your company that a lot of these UI components that for sure your employees and your teammates will start to be making are consistent as well.
So for example, Stripe, their VP of design, Katie Dill, also talked about this in her recent talk where they basically made this similar idea of having a UI library for their employees and their agents to grab from in order to create Stripe branded material, which is cost efficient and is also just going to... unify the look and feel of every application that your team builds.
So there you go. That is all the components of the ARMS framework. And as mentioned, if you want all the prompts and resources that I mentioned in this video, you can just grab that guide below, which you can send to your agent and pick out the nuggets that you need.
Thanks for watching till the end. And usually I do sessions like these to help you use AI better for your business. So if that's your thing, then feel free to subscribe as well.
I'll see you next time. Cheers.
The Hook
The bait, then the rug-pull.
The Claude 5.5 models, a near-free decision model, and app mods all landed within weeks, yet most agent setups still follow older advice. This breakdown walks through a live agentic OS dashboard, then the four ARMS layers with three levels each, and the numbers behind them: 18 cents for 4,451 routed prompts, $437 saved by keeping the cache warm.
Named ideas worth stealing.
04:15acronym
The ARMS Framework
- Apps
- Routines
- Memory
- Skills
Four layers for organizing an agentic OS, learned bottom-up from Skills to Apps, each with three levels of maturity.
05:33list
Three skill types
- Database skills (hold knowledge and references)
- Artifact skills (produce HTML-based decks, PDFs, infographics)
- Self-triggered skills (run outside chat via claude -p)
Ways to think about skills beyond step-by-step procedures.
07:26ratio
8 references + 2 wildcards
- 8 options drawn from your reference database
- 2 wildcards that ignore references
A default split that keeps database skills from overfitting and surfaces new styles worth adding back to the references.
21:40stack
Three layers of long-term memory
- MEMORY.md (preloaded index of what Claude remembers)
- CLAUDE.md or AGENTS.md as a workspace map routing to department files
- Database skills that point to where live data sits
A plain-text map Claude reads first so it can find files fast and cheap without vector search.
31:40ordered-list
Connector priority order
- Official connector
- CLI
- API
- Community MCP
Search order for connecting an app, from lightest to heaviest on context and usage.
How they asked for the click.
VERBAL ASK
37:38link
“If you want all the prompts and resources that I mentioned in this video, you can just grab that guide below, which you can send to your agent and pick out the nuggets that you need.”
Soft lead-magnet CTA at the framework intro and the outro, pointing to a free PDF guide in a Skool community.
MENTIONED ON CAMERA
Visual structure at a glance.
