14:18Jay E | RoboNuggets · Tutorial
STOP Prompting Claude
A 14-minute tutorial on the three tiers of self-running Claude Code workflows — and why the creator of Claude Code stopped prompting it manually.
June 12thThe argument in one line.
A well-structured, visually-mapped memory system lets an AI agent retrieve the right file through deterministic scoring instead of brute-force reading, cutting both response time and token cost by roughly 40 percent.
Read if. Skip if.
READ IF YOU ARE…
- You run Claude Code (or a similar agentic coding tool) against a large personal or business file system and want faster, cheaper retrieval.
- You already have dozens of connected tools, scheduled automations, or skills and have lost track of what's actually wired up.
- You want a repeatable way to audit which integrations are safe to keep versus which should be disconnected.
- You're building a memory or context system for a client and need a way to visually explain it to a non-technical stakeholder.
SKIP IF…
- You use Claude for occasional one-off tasks with a small, simple workspace where retrieval speed isn't a bottleneck.
- You're looking for a plug-and-play open-source tool rather than a custom system built through extended back-and-forth prompting.
The full version, fast.
The core claim is that most personal knowledge-graph setups (the Obsidian-style node view) are cosmetic and don't actually make an AI agent faster or cheaper to query. The creator had Claude map his real workspace into four layers -- connected applications, scheduled routines, memory files, and skills -- and built a small deterministic script that scores and narrows candidate files by keyword before ever invoking the model, rather than letting the agent grep and read its way through everything. In a live side-by-side test, the session using this system answered faster and used about 30,000 tokens versus about 50,000 for a default session, a roughly 40 percent saving. The actionable path is two-pronged: have Claude research current best practices for memory systems via a web-research skill, and feed it references to existing open-source memory/graph projects as design inspiration before asking it to build a version tailored to your own files.
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 – 00:41
01 · Intro
States the thesis: while Claude Opus is included in the subscription, have it build a real second-brain system, not a cosmetic Obsidian-style graph.
00:41 – 01:53
02 · The second brain graph
Tours the Rubric node-graph UI mapping business, content, personal, and community files into visual departments.
01:53 – 02:24
03 · The ARMS framework
Names the four layers of the system: Applications, Routines, Memory, Skills.
02:24 – 03:56
04 · Applications: power and risk
Explains that connected-tool nodes signal both automation power and security/trust risk, and should be audited periodically.
03:56 – 05:55
05 · Routines, memory and skills
Covers scheduled routines (retire unused ones), the 35,466-file memory layer, and the skills layer with direct file access from the graph.
05:55 – 06:54
06 · Using it with clients
Suggests the visualization is a clean way to walk a client through their own agentic operating system.
06:54 – 07:56
07 · Why the graph is only 30%
Argues the visual is only part of the value; the other 70% is whether retrieval actually gets faster and cheaper day to day.
07:56 – 09:32
08 · Live test: faster and cheaper
Runs the same prompt in a second-brain session and a default session simultaneously; the second-brain session answers first and uses about 30K vs 50K tokens.
09:32 – 11:02
09 · How it works under the hood
Explains the brain.js retrieval script: strip keywords, score candidate files without reading them, open only the top match, check the relevant section, follow pointers.
11:02 – 12:12
10 · Build your own second brain
Gives the two prompting techniques used: a 'last 30 days' web-research skill for best practices, and naming reference open-source memory projects (QMD, g-brain, Graphify) as design inspiration.
12:12 – 13:41
11 · Verdict and free guide
Closes with an overall endorsement while the model is included in the subscription, and points to a free PDF guide in the pinned comment.
Lines worth screenshotting.
- A visual knowledge graph that just shows connected nodes without changing retrieval behavior is, by the creator's own admission, mostly for show.
- The real gain from a second-brain system is deterministic retrieval: scoring candidate files by keyword before opening any of them, rather than letting the model grep and read everything.
- In a live test, a workspace with this system answered a query using about 30,000 tokens versus about 50,000 for a default agent session -- a roughly 40 percent reduction.
- The number of connected applications in an agent's workspace is a double-edged signal: more nodes means more automation power, but also more surface area for risk if a connected tool can take real-world actions like sending emails.
- An unused scheduled automation you don't remember creating is a candidate for retirement, not just a leftover to ignore.
- Every retrieval starts by stripping a query down to its real keywords and discarding filler words like "which" or "do we use" before scoring any files.
- The scoring step assigns a likelihood score to candidate files without opening or reading any of them first, then only opens the highest-scoring file.
- Even after opening a file, the system checks for the specific section likely to hold the answer instead of reading the whole document.
- If a section is just a pointer to another location, the deterministic code follows that pointer before ever falling back to the model.
- Prompting Claude with a web-research skill that scans Reddit, Twitter/X, YouTube, and Hacker News for current best practices was one of the two core techniques used to design the system.
- Naming specific existing open-source memory or graph projects in a prompt (rather than describing the concept abstractly) is a concrete way to steer an agent's system design toward a proven pattern.
- Setting an explicit performance goal, such as requiring a visualization to reload in under ten seconds, gives an agent a target it can check its own work against instead of stopping at 'looks done'.
A second brain only pays off if retrieval gets cheaper and faster.
WHAT TO LEARN
A visual map of your files is worthless on its own -- the value is a deterministic scoring layer that finds the right file before the model ever reads anything, and that's what actually cuts cost and latency.
01Intro
- A pretty node-graph view of your files is mostly cosmetic unless it changes how an agent actually retrieves information.
02The second brain graph
- Structure your workspace into distinct layers -- connected tools, scheduled automations, memory files, and reusable skills -- so each can be audited on its own.
04Applications: power and risk
- Treat the count of connected tools in an agentic workspace as a risk signal, not just a power signal, since more connections mean more that could go wrong if misused.
05Routines, memory and skills
- Periodically review scheduled automations you no longer remember setting up and retire the ones you don't actively rely on.
09How it works under the hood
- Score candidate files by keyword before opening any of them, so the agent narrows down to the most likely source without reading everything.
- Once a file is opened, check for the specific relevant section rather than reading the whole document, and follow internal pointers instead of re-searching.
08Live test: faster and cheaper
- Measure any memory system you build by two concrete numbers -- response speed and token cost -- compared against a default session, not by how the visualization looks.
10Build your own second brain
- When prompting an AI to design a system for you, point it at a live web-research pass on current best practices and name specific existing projects as references rather than describing the goal abstractly.
- Set an explicit, measurable performance goal, like a maximum load time, so the agent has a concrete target to verify its own work against.
Terms worth knowing.
- Second brain
- A structured external memory system -- files, folders, and notes -- organized so both a person and an AI agent can quickly find and reuse past information.
- Node graph / knowledge graph
- A visual map where dots (nodes) represent files, tools, or concepts and lines represent relationships between them.
- Deterministic code
- Code that always produces the same output for the same input without invoking a language model, used here to score and locate files cheaply before the model reads anything.
- Context window / token cost
- The limited space and dollar cost of everything an AI model has to read to answer a query; fewer tokens read means faster and cheaper responses.
- Skill (in an agentic tool)
- A reusable, named capability or script that an AI agent can invoke to perform a specific task, such as running a web search across specific sites.
- MCP connector
- A protocol-based connection that lets an AI agent read from or act on an external application, such as a calendar or CRM.
- Semantic search
- Search that matches meaning and context rather than exact keywords, letting a query find relevant results even without matching words.
Things they pointed at.
11:02toolLast 30 Days skill (Matt Van Horn)
12:12toolQMD (Query My Docs, Toby Lutke)
12:25toolg-brain (Garry Tan)
12:32toolGraphify
02:05linkSecond Brain PDF guide
Lines you could clip.
01:05
“A data visualization chart like this isn't really useful if it doesn't tell you any information about what you're seeing.”
12:06
“It only captures around 30% of the value of a second brain system, because the other 70% has to do with the day-to-day work you ask Claude to do for you.”
15:16
“At least for speed, this one already has the answer while the default Claude code is still channeling and thinking through it.”
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.
metaphoranalogy
00:00Cloud Fable five is the most powerful model in the world. And one of the best things that you can do right now while it's still part of your subscription is to point it to your workspace and have it create for you a proper second brain system. And no, I'm not talking about a simple Obsidian graph that doesn't really do anything except for show.
00:17I'm talking about a proper system, one that maps your whole workspace, cuts your token cost, and finds anything faster than your default Cloud Code. So that is exactly what I did for my setup.
00:28And today, I'll show you how it works, how to have Fable build one for your workspace, and how to get the same system running for you before this pricing window closes. Let's dive into it.
00:41So what we're looking at right now is the entirety of the second brain that Claude Fable five made for me. So I have files for my business work. I have files for content because I do a lot of content, stuff for personal, as well as for my community.
00:54And you can see in the second brain system, you have these departments similar to what you would expect, let's say, if you were building out a company second brain. But when I was building out this system along with Fable, I actually wanted to make sure that this system and this visual is useful as well. Because let's say you're using a default tool for graphing your nodes, your second brain.
01:14Similar to what you would expect from something like Obsidian, which is probably the most popular one in the market right now, to me a data visualization chart like this isn't really useful if it doesn't tell you any information about what you're seeing here. More often than not, what I'm finding is for people using Obsidian as their second brain in this graph view, it is merely just for show and just for illustrating that you have this fancy connections in your nodes.
01:36But for this version of the second brain, apart from it just being a bit more visual and giving me clear view of, let's say, how many more files I have in my business department versus that of content versus that of my personal workspace, It also gives me a clearer view of my entire AgenTeq operating system, which at least for me, the way I teach it is if you get these four aspects right, your applications, your routines, your memory, as well as your skills, then you are definitely going to be ahead of 99% of the users of agentic tools like Cloud Code.
02:08And as always, to make this easy for you, I've created this PDF guide that just provides all of the learnings as well as all of the principles that I will talk about in this lesson so that you can either read through it or provide it to your Cloud Code, and it will understand how I made this system. And you can just find this in the link down below.
02:24And so here on the blue line, you can see which particular applications I have connected to my Cloud Code, whether that's via an MCP connector, an API, or a CLI tool. And you can see when Fable five made this for me, it identified that I have it connected to Google Calendar, I have it connected to Google Drive, and so on and so forth.
02:40And the reason why that's important is because this application view, it gives you visibility on two things. The first is if you have an application that you are using constantly that is not yet in this blue circle, then that might mean that you would need to do some work in order to plug that in to Cloud Code in order to automate that tool.
02:58And so in principle, the more nodes that you have in this application layer, the more powerful Cloud Code becomes. But that also has a flip side to it.
03:05Right? Because the number of nodes that you have in your application layer here doesn't only signify how powerful Cloud Code is, it also shows the level of trust and correspondingly the risk that you would have when using these agentic platforms.
03:18Because with a lot of these tools, let's say you're connected to HubSpot for instance, what that basically means is that if Cloud Code can connect and control a 100% your HubSpot workspace, then it'll be able to send emails and marketing campaigns to your mailing list, which can be risky by itself if you don't know what you're doing.
03:36And so the second way by which this application layer is going to be useful is if there's an application that you are actually not using anymore, then it might be worth disconnecting that to Cloud Code and this nice graph view actually gives you that whole picture so that you can just quickly eyeball which particular applications that yourself or your business or your client would need to disconnect to.
03:56Now, there's the same principle when it comes to routines. Because routines, what are those? They are basically your scheduled tasks that are running in the background.
04:03And I've talked about this in several other videos before, but the use of this yellow area is quite similar to what you would expect with the applications layer as well. If you have a lot of routine setup, then that means that you have a lot of automation setup. So it may mean also that you are saving a lot of time.
04:18However, if you have a lot of routines in here and you actually don't remember anymore you setting this up and you actually don't need this particular automation anymore, then it might be worth retiring it as well. That is, at least for me personally, is running on my Hermes agent, which if you need a quick reminder of what that does, you can see that it is connected to this skill for the daily log skill that actually I can open that file directly from my file explorer here on Windows.
04:43Or if I zoom in here and actually look at the skill, then I can actually read the skill and what it contains straight from this second brain view. And then, course, you have your memory layer or the context that you have built up for your business over time. So for me personally, I have roughly, and I didn't know this, apparently, I already have 35,466 files in this folder, and Fable was able to map those out and actually create relationships between those nodes and those files.
05:11So So now let's say I'm looking for a particular file that I can just fully replace my file explorer or my finder in order to just find the specific document that I'm after straight from within this second brain system. So let's say I'm looking for this document or this photo, which I already know the file name of and I don't really need Claude to find it for me, then it can also show me that and I can even open it straight from here.
05:33And it will open that on my device. So you can see this is just a five by five grid of profile photos in there that we can use. And then obviously, and the last layer, which is probably the most important and most useful, are the skills.
05:44So in this one view, you have a clear visibility of what your skills are. And similar to what I showed before, you can also open them and see which files or routines they are connected to. And so with that framework, if you're building this for someone else or for a client, then you can cleanly and clearly illustrate to them how their whole AgenTeq operating system works, how their second brain system is set up, and what are the different departments or fields of work that their current workspace or second brain is operating against.
06:14And you can even take them through actual files in here and click through it in order to expand any of the folders that you may be interested in. And if it's a sufficiently good visualization like this, just like what CloudFable did for us, then I think it is also pretty useful.
06:28Right? Because this is probably a much better way to communicate the idea of a second brain system and how your whole workspace looks like versus just showing literally a folder in your workspace, which for me, that whole second brain system that just represents all of the files and all of the folders that I have in this robo workspace.
06:45So if you're taking someone through their workspace, it is probably much better and much easier to take them through those ideas in this view rather than that folder structure. But no matter how good or fancy or flashy this visualization is, I would say that it only captures around 30% of the value of a second brain system because the other 70% has to do with the day to day work that you ask Claude to do for you.
07:09And the question you should ask yourself is, with this second brain system that I'm implementing, am I actually getting Claude to retrieve documents from our memory, from our second brain much faster? And often more importantly, is it getting those at a lower token cost?
07:24Because if you set up this whole second brain system, but it either slows you down or it ends up costing you much more just to find files in here, then it's probably not worth setting up. And at least for me with this system that Fable has set up, it was able to do those two things. It was able to retrieve things faster as well as for cheaper.
07:41And I'll show you how it works and how I guided Fable to create this system for me in a bit. But just to show it with a quick test, what I have here are two Cloud Code sessions. This one already has that second brain system implemented and this blue one doesn't.
07:53So what I'll do now is just send these two messages, these two promise at the same time, and then by the end of it, we will see which of them will get the answer fastest. And number two, if we are saving tokens using this second brain system. And there you go.
08:06At least for speed, this one already has the answer for us while the one that is using the default cloud code without that second brain system is still and basically using these default finders like grep and glob in order to find your files for you. And when both of those tasks have finished, what I can do now is just type in slash context in order to just find out how much tokens each of these have used.
08:31So the way to read this in case it's your first time seeing it is that these system prompts, tools, memory files, and skills, those are pre injected on every session that you start. That's why they are mostly fixed as you can see. But the messages here, this purple area, that is what Claude actually did in that session.
08:47And you can see for me, with this second brain system, it only took around 30,000 tokens to get to the answer, while for the default Claude code, it took around 50,000. So there's roughly around a 40% savings there, which is pretty good. Now I did also ask Fable to, of course, do a few more tests just to verify that, and it basically spun up that same test for default Cloud Code versus the second brain system that it built and is passing those criteria for each of them.
09:11Now because each of our workspace is actually quite different, so the way that I use Cloud Code and the files and the folders that I have in that Robo workspace is probably different versus yours. So I think what's most helpful is if I give you the principles of how you can guide Fable to create the perfect second brain for your system.
09:26But in case you're curious, the design system that it created for me is also pretty interesting, so I'll just go through that for a bit. Because basically what it did is every time I interact with it or I ask it to retrieve a piece from our memory, it uses this custom code that it created called brain dot j s. Stands for JavaScript.
09:41And if I pull back that session from earlier, you can see here that it is invoking that brain dot j s at the beginning of this retrieval task. And what it does after is quite simple. Let's say we ask it which TTS voice we use.
09:53It essentially just strips down the keywords there, which is just TTS voice in this case and other filler words like which do we use, those are discarded. And at this point for all of these, the reason why it is able to save in tokens is because this is deterministic code. So it's not really invoking the model in order to do all of these things.
10:12Because I have a lot of indexes and reference maps in my own workspace, in my own second brain, it can actually do this where it gives the possible sources a score without opening or necessarily reading any of them yet. And then when it finds the highest score, basically the file or the markdown file that has the highest likelihood where that particular information is stored, then it opens that, but it doesn't read the whole thing.
10:35It actually checks if there's a particular section where that answer may be obtained. So it works in multiple layers in order to just prevent it from reading irrelevant files. And then it says here, if the section is a pointer to somewhere else, then the code just follows that pointer.
10:49Again, this is deterministic code. And then once it has that answer, it now provides that to the LLM, which is now the model answers back to you. But like I said, that's the system that works for me.
11:00It may not work for you necessarily, and so the principle of how you can guide Fable five in order to do this for you is probably more useful. And the core of it really is two things.
11:09The first one is I just prompted Cloud Fable five to do this slash last thirty days skill, which is this skill by Matt Van Horn, who is the cofounder of Lyft. And what that does essentially is it will scour the web across Reddit, across Twitter, YouTube, Hacker News, and other several sources in order to find, at least for this case, the best practices when it comes to second brain system.
11:29So I just had it pull and do a deep research on that information and use that information as well as scanning our workspace in order to create that system for us. The other thing is because I'm a bit closer to this space, I actually know of a few memory projects, which are these open source repositories that I think would be good resources for Fable.
11:46So if you want to use these as well, you could just take a screenshot of this and include that in your prompt to Fable five. To give an example, this QMD project, which I think stands for query my docs. That's from Toby, the founder of Shopify, which essentially just applies the principle of semantic search to your memory system.
12:01There is this system by Gary Tan, the CEO of Y Combinator on what he uses as his second brain, which he calls g brain. And then there is another open source repo that is now actually funded by Y Combinator called Graphify that just creates smarter connections between your files and folders.
12:16So I just gave all these three to Fable and to ask it to go through these repositories, consider them, and build out the perfect second brain system for my workspace. And the other helpful thing when it comes to creating a piece this visual is that you can actually ask Fable to set a slash goal to make sure that, let's say, if you're changing layouts in here or you're moving nodes around, it can actually check its own work to make sure that there is no lag for the user.
12:39Similar with every time that I reload this and it refreshes, I also set a slash goal there in order to make sure that it loads under ten seconds and that it should continuously optimize the whole interface up until it gets to that point. So that's just one way for you to improve, I guess, the user experience if this visual view is something that you are after.
12:57But there you go. This is my second brain system. It is, I think, a good way for you to use Cloud Fable five while it's here in the non usage based pricing.
13:07It's a really powerful model. And if you're looking to upgrade your second brain system and the way that your context and your memory is set up, then I think you can offload a lot of that work while Fable is part of your subscription. And, again, the whole PDF guide of how I set this up so that you can just give that to Claude and go through this exercise yourself as well, you can just find that in the pinned comment below.
13:26But that's it for this one. As always, thank you for watching until the end, and I'll see you all next time. Cheers.
The Hook
The bait, then the rug-pull.
A creator claims most personal knowledge graphs are cosmetic -- pretty node diagrams that don't actually change how fast or cheaply an AI agent finds your files. He sets out to prove his own system is different, with a live side-by-side speed and cost test against a default session.
Named ideas worth stealing.
01:53acronym
ARMS (Applications, Routines, Memory, Skills)
- Applications
- Routines
- Memory
- Skills
The creator's framework for the four layers that make up an agentic operating system for Claude Code, each rendered as a distinct colored cluster in the node graph.
Steal forauditing any AI agent workspace -- what it's connected to, what it automates, what it remembers, and what reusable capabilities it has
How they asked for the click.
VERBAL ASK
02:05link
“I've created this PDF guide... you can just find this in the link down below.”
Soft-pitched twice (once early, once at close) as a free lead magnet in the pinned comment/description; a paid import-ready version of the system is also linked in the description but never mentioned on camera.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
AFFILIATECommission earned if you click.
Visual structure at a glance.
