Latest / AI Fire Daily / #529 Neil: Build A Second Brain With Claude Fable 5 In Claude Code
Transcript
- 0:00Intro music. There's this silent, creeping problem
- 0:04in how we use AI today. We give it absolutely
- 0:08everything. Oh yeah, we really do. Just hand
- 0:10over the keys to the whole mess. Right. We take
- 0:13years of chaotic notes, nested folders, random
- 0:17PDFs, and we dump it all into the context window.
- 0:20And we expect the model to somehow make perfect
- 0:22sense of it. Which is wild when you actually
- 0:24think about it. It is! Because instead of getting
- 0:27smarter, the AI just gets overwhelmed. It drowns
- 0:29the noise, it gets slower, and inevitably it
- 0:32just becomes far less useful to you. It is definitely
- 0:34a slow, creeping kind of failure. You don't usually
- 0:37notice it on day one. But eventually, the sheer
- 0:40weight of that digital chaos always wins out.
- 0:44Welcome to this deep dive. Today we are taking
- 0:46on a very specific mission for you. We are dissecting
- 0:49a fascinating guide on building a fable five
- 0:51second brain. And the really interesting twist
- 0:53here is that we are building it entirely inside
- 0:56Claude code. Exactly. So we are going to explore
- 0:58the four foundational layers that make up this
- 1:01brain. We'll examine why the system absolutely
- 1:03must live in your local environment. Right, and
- 1:06we'll also walk through the exact prompts required
- 1:08to launch it properly. And finally, we will break
- 1:10down how to audit the system. Because if you
- 1:13ignore these setups, well, they degrade rapidly
- 1:15over time. Oh, for sure. A second brain is never
- 1:19just a, you know, set it and forget it tool.
- 1:21It requires intentional structure. It demands
- 1:25real maintenance. So before we can build a better
- 1:27system, we really need to understand why our
- 1:30current habits break down. And it basically comes
- 1:32down to context overload. Yeah, context overload
- 1:35is the perfect term for it. I mean, if Claude
- 1:37Code is just sitting on my hard drive, how does
- 1:39it know the difference between a random brainstorming
- 1:41document from three years ago and an actual critical
- 1:44workflow I wanted to run today? Well, That is
- 1:47exactly the problem the Fable 5 methodology solves.
- 1:50And for those who might not be familiar, Fable
- 1:525 isn't some new piece of software you have to
- 1:55go buy or download. Right. It's just a method.
- 1:57Exactly. It's a highly specific prompting methodology.
- 2:01It is a structural framework for organizing your
- 2:03local files so Cloud Code can act as an actual
- 2:07autonomous agent. Which is so important because
- 2:10normally the problem just builds up so slowly.
- 2:13You create a new project folder over here. You
- 2:16drop a quick meeting note over there. Yeah. And
- 2:18crucial long -term decisions end up buried inside
- 2:22random unnamed text documents. It feels completely
- 2:26harmless in the moment. I mean, we all do it.
- 2:28But consider what happens when you ask Claude
- 2:30a really complex question a month later. It has
- 2:33to guess. Right. The AI has to dig blindly. It
- 2:36starts searching through this massive unstructured
- 2:39mountain of context. It wastes processing time,
- 2:42and more importantly, it wastes tokens. Meaning
- 2:45the digital budget the AI spends to read and
- 2:48process your text. Yes. And when Claude reads
- 2:50hundreds of totally irrelevant files just to
- 2:53find one single answer, it burns through that
- 2:55budget fast. The context window just fills up
- 2:58with junk. Exactly. The active memory limit the
- 3:00AI can process at once gets maxed out. And as
- 3:02a result, the answer you get is shallow or just
- 3:04flat out hallucinates facts. And the solution
- 3:07provided in our source material is incredibly
- 3:09pragmatic. It pushes back against this recent
- 3:11trend of drawing massive complex visual graphs
- 3:15of your knowledge. Yeah, those big web -like
- 3:17graphs that look cool but are practically useless
- 3:20for an AI. Right. Instead, it argues for splitting
- 3:24your system into four distinct layers. Layers
- 3:27that an AI can actually parse logically. Because
- 3:30the AI simply does not think like human beings
- 3:33do. No, it really doesn't. It doesn't care about
- 3:35pretty folder colors or visual nodes. It needs
- 3:38structured, hard -coded categorization. So the
- 3:41Fable 5 guide... breaks your digital life down
- 3:44into apps, routines, memory, and skills. Let's
- 3:47explore how that actually functions in practice,
- 3:49starting with that first layer, which is apps.
- 3:52OK, so these are outside systems, right? They
- 3:54are the data sources and external tools that
- 3:56your local AI can connect to. Exactly. Think
- 3:59of apps as giving Claude a pair of hands to reach
- 4:01outside your hard drive. Right, like it can pull
- 4:03live data from GitHub, or read a Slack channel,
- 4:06or query a database. Yep, and it usually does
- 4:08this using MCP. Which is basically a bridge letting
- 4:11AI use your out - some tools. Precisely. It stands
- 4:13for model context protocol, so that handles all
- 4:16your external connections. Now the second layer
- 4:18is routines. These are your repeated workflows.
- 4:21Yeah, the actions you run every single day or
- 4:23week. Things like pre -commit checks on your
- 4:25code or weekly report generation. And in the
- 4:29Claude code environment, these routines are often
- 4:31managed using hooks. Right, and hooks are just
- 4:34small scripts that trigger automatically during
- 4:35a specific task. So when you start a new feature
- 4:38branch, a routine layer tells Claude exactly
- 4:41which scripts to run in the background. That
- 4:43makes a lot of sense. Then we move to the third
- 4:46layer, which is memory. And this one feels incredibly
- 4:49crucial. Both of the core of it all is the context
- 4:52that gets reused across all your different sessions.
- 4:55Because Claude Code actually uses a dedicated
- 4:57file for this called Clauye .md, right? Yes,
- 5:00this layer acts as your ongoing project history.
- 5:03If you made a major architectural decision last
- 5:06month or established a specific brand voice that
- 5:09lives in the memory layer so the AI never forgets
- 5:12it. And finally, the fourth layer is skills.
- 5:14These are your highly reusable instructions.
- 5:17Standard operating procedures for very specific
- 5:19processes. You would store these in a dedicated
- 5:22skill .md file. So when you give Claude a task
- 5:26that matches a known requirement, it just calls
- 5:29up that exact skill. Right. It does not guess
- 5:32how to format your code or draft your email.
- 5:34It opens the manual and follows your rules. You
- 5:37know, it makes me think of a library. Right now,
- 5:40most of us hand our AI a search warrant for the
- 5:42entire building and say, uh, go find something
- 5:45about Project X. And it takes forever. Exactly.
- 5:48It's so inefficient. But this four -layer system
- 5:51is like stacking Lego blocks of data instead
- 5:53of tossing them all into a giant plastic bin.
- 5:56You are narrowing the search space so the AI
- 5:58doesn't exhaust itself before it even begins
- 6:00the actual work. That Lego analogy hits the nail
- 6:03on the head. When the AI needs a specific workflow,
- 6:06it goes straight to the routines layer. It ignores
- 6:09the memory layer completely for that specific
- 6:11query. But I have to ask a practical question
- 6:13here. Does setting up these four layers mean
- 6:16we have to completely abandon our traditional
- 6:18folders? Not at all. That is the beauty of this
- 6:21approach. You keep your current chaotic folder
- 6:23structure exactly as it is. You don't move anything.
- 6:27We are just creating a new navigational lens.
- 6:30So, no, it just maps your existing mess so the
- 6:33AI can navigate it. Exactly. You don't have to
- 6:36change your daily habits at all. The AI just
- 6:38gets a much better map. Okay. Now that we understand
- 6:41the architectural blueprint, we really need to
- 6:43talk about the construction site because... A
- 6:46great blueprint is completely useless if you
- 6:48can't access the actual building materials. This
- 6:51brings us to a major technical distinction in
- 6:54the source guide. We have to address the difference
- 6:56between clawed chat and clawed code. And the
- 6:59guide is completely uncompromising on this point.
- 7:03This entire second brain workflow belongs exclusively
- 7:06inside Cloud Code. Right. Not the standard web
- 7:10browser interface. Which I think is a tough pill
- 7:12to swallow for some people. Oh, totally. Cloud
- 7:14Chat is wonderful for plenty of things. You can
- 7:17brainstorm ideas there. You can draft the initial
- 7:19prompts there. But for building a functional
- 7:21second brain, the web interface is the wrong
- 7:24tool entirely. I get the logic, but I have to
- 7:27push back a little on behi - on behalf of the
- 7:28listener, there is a real, undeniable comfort
- 7:31in a standard browser window. Oh, for sure. It
- 7:34feels safe. It is familiar. Opening up a terminal
- 7:37environment to run cloud code can feel incredibly
- 7:40intimidating for non -developers. I completely
- 7:42sympathize with that. The command line is stark.
- 7:45It doesn't have friendly little buttons. Right.
- 7:47But comfort simply cannot be the primary goal
- 7:50here. Right. Autonomous execution is the goal.
- 7:53If you want a second brain that actually works
- 7:55for you, it must live where the actual work happens.
- 7:58Because the AI fundamentally requires access
- 8:01to the real workspace. It needs local permissions.
- 8:04Exactly. It has to actually read your local files
- 8:06and scan your messy folders. And it needs to
- 8:09utilize those skills and routines we just mapped
- 8:11out. Yes. It needs to check its own token usage
- 8:14in real time. It has to run retrieval logic against
- 8:17your actual hard drive. Cloud Chat is isolated
- 8:20in a browser. window. It is trapped in a sandbox.
- 8:23You cannot reach out and touch your files. So
- 8:25is the standard chat interface totally useless
- 8:28for this kind of project? I mean it serves a
- 8:30purpose very early in the process. You can use
- 8:33it to research how to write the initial instructions,
- 8:35but the moment you actually start building the
- 8:37system, the web check cannot execute a single
- 8:40required action on your files. OK, so chat is
- 8:43for brainstorming. Code is for executing inside
- 8:46your actual workspace. You have to put the AI
- 8:48in the room with the raw materials. OK, so we
- 8:51are in the terminal. We are running Claude code.
- 8:55How do we actually instruct Fable 5 to build
- 8:58this thing? We obviously can't just type, ah,
- 9:00make me a second brain. Oh, if you do that, the
- 9:02AI will just hallucinate a bunch of generic templates.
- 9:05You will end up with an even bigger mess. Right.
- 9:07So the source guide provides a highly specific
- 9:10multi -step prompt to prevent exactly that. And
- 9:13the sequence of that prompt is fascinating. It
- 9:15really is. The very first instruction is vital.
- 9:18You command the AI to inspect the entire workspace
- 9:21structure, but explicitly tell it not to change
- 9:24a single thing. Look, but do not touch. Exactly.
- 9:26Look, but do not touch. It needs to comprehend
- 9:29the current reality first. It scans the folder
- 9:32depths. It reads the scattered notes. And it
- 9:34locates where the actual project history lives
- 9:37today in all its chaotic glory. And once it finishes
- 9:40that scan, it moves to the division phase. You
- 9:43instruct Fable 5 to mentally categorize your
- 9:46workspace into those four layers we discussed.
- 9:48Apps, routines, memory, and skills. Yeah. You
- 9:51are demanding that it identify patterns in what
- 9:54already exists. Right. It locates your repeated
- 9:56workflows, it finds your API connections, and
- 9:59it flags your reusable operating procedures.
- 10:02And then we get to the core engine, you instruct
- 10:05it to design the retrieval flow, and this is
- 10:07where the whole concept either succeeds or fails.
- 10:09This is the mechanism that changes everything.
- 10:12The flow must execute a very specific sequence
- 10:14of logic. First, it takes your conversational
- 10:17request and extracts the underlying intent to
- 10:20generate search keywords. So it doesn't just
- 10:22blindly search the exact sentence you typed.
- 10:24It interprets what you actually need. Right.
- 10:26And here's the brilliant part. It generates a
- 10:28list of candidate files, and then it scores them
- 10:31for relevance before it ever opens them. Whoa.
- 10:35Wait. Imagine it scoring candidate files and
- 10:37surgically reading only the exact right section
- 10:40instead of blind ingestion. That is incredible.
- 10:43It really is a paradigm shift. It checks file
- 10:46names. It parses the metadata. It essentially
- 10:48builds a hypothesis like, I bet the context for
- 10:51this design choice is in this specific archive
- 10:54folder. And it tests that hypothesis before burning
- 10:57tokens to read a 50 -page document. And if the
- 11:00document is huge, it looks for the most relevant
- 11:02section inside the file. It passes only that
- 11:05specific useful context back into your working
- 11:07conversation. Which completely bypasses that
- 11:10context overload problem we started with. The
- 11:12final step of the prompt focuses on the interface
- 11:14layer. But the guide is very particular here.
- 11:17It advises you to ask for interface improvements,
- 11:19like clear labels and simple navigation. But
- 11:22it explicitly warns against visual mapping. Yeah,
- 11:25very strongly. So why does the guide specifically
- 11:28advise against creating a visual map at this
- 11:30stage? Because human beings get easily distracted
- 11:33by aesthetics. Massive interconnected visual
- 11:37graphs look incredibly impressive on a monitor,
- 11:39but they add massive computational noise for
- 11:41an AI. A pretty graph does not help Claude find
- 11:45a text file one millisecond faster. So maps look
- 11:48pretty, but add noise. Retrieval speed is what
- 11:50actually matters here. Function over form. every
- 11:53single time. We'll dive into the next phase of
- 11:55this right after a quick break. Mid -roll sponsor,
- 11:58Reed, placeholder. And we're back. So we've covered
- 12:03a lot of the theory so far. Yes. We have delivered
- 12:05the instructions. We have successfully mapped
- 12:08the retrieval flow in theory. But a theory always
- 12:12needs to be tested against harsh reality. Oh,
- 12:14absolutely. We cannot trust this system with
- 12:17our daily workflow yet. So after Fable 5 presents
- 12:20its initial plan, The very next tangible step
- 12:22is creating an inets .md file. Just a simple,
- 12:26flat text file. Right. Keep it as simple as possible.
- 12:28This index file lists your four sections. It
- 12:30includes the exact file paths. It provides short,
- 12:33punchy descriptions of what lives in those paths,
- 12:35and it dictates exactly when Claude should use
- 12:37them. It functions like a hyper -intelligent
- 12:39table of contents. It gives Claude code one definitive
- 12:43place to check before it does anything else.
- 12:45which prevents it from blindly scanning your
- 12:47entire hard drive every time you ask a basic
- 12:49question. Exactly. Now once that index file is
- 12:53built, the guide demands that you test it. Yes,
- 12:56and it suggests running three... highly specific
- 12:59real -world retrieval questions. They have to
- 13:02be real questions from your actual work. Hypotheticals
- 13:05will not stress test the logic at all. Right.
- 13:07So you ask one question about a prompt or a past
- 13:09architectural decision. You ask one question
- 13:12about a daily routine. And you ask one question
- 13:15regarding deep project history from the memory
- 13:17layer. You submit the query. and then you closely
- 13:20evaluate how the AI processes it. You monitor
- 13:22the terminal. Did it extract the correct search
- 13:25keywords from your prompt? Did it open the correct
- 13:27file on its very first try? Exactly. And most
- 13:30importantly, how many files did it blindly inspect
- 13:32before it finally found the right answer? I have
- 13:34to offer a vulnerable admission here. I still
- 13:36wrestle with prompt drift myself when I try to
- 13:39manage too many loose notes. Oh, it happens to
- 13:40everyone. It really does. I will ask my AI to
- 13:44draft an email based on last week's meeting and
- 13:47suddenly it is hallucinating details from a totally
- 13:50different project I completed three years ago
- 13:52simply because it read the wrong file first.
- 13:55We have all experienced that exact frustration.
- 13:58That hallucination is the exact pain point this
- 14:02index file is designed to cure. By using the
- 14:05index, you force the AI to show its mathematical
- 14:07work. Because you can literally watch the terminal
- 14:10to see where its logic fails. Yeah. So what is
- 14:12the immediate fix if the test shows Fable 5 opening
- 14:16way too many files? Well, if it opens too many
- 14:19files, your categories are simply too broad.
- 14:22The AI is guessing. You need to restrict the
- 14:24search parameters. Right. You either tighten
- 14:26the rules inside the index file, or you rename
- 14:28your source files so their actual purpose is
- 14:30brutally obvious to a machine. You need to build
- 14:32a narrower index or write much clearer file labels?
- 14:36Yes. Make it mathematically impossible for the
- 14:38AI to misunderstand what a file actually contains.
- 14:41Okay, let's say you have built it, you have tested
- 14:43it, and it works beautifully. But a second brain
- 14:46isn't a stone monument that lasts forever. No,
- 14:48it really isn't. It is much more like a living
- 14:50garden. And if you don't actively tend to that
- 14:53garden, the weeds are going to come back remarkably
- 14:55fast. Because digital workspaces change constantly.
- 14:58You try out a new task management app. You alter
- 15:03your daily coding routine. Or you just get lazy
- 15:05on a Friday and stop labeling your files properly.
- 15:07Exactly. Entropy is a fundamental law of digital
- 15:10environments. Without scheduled maintenance,
- 15:12your beautiful lightning -fast retrieval system
- 15:15will decay. The paths will break. the context
- 15:18will get messy again. To combat this, the guide
- 15:21provides a specific audit prompt. It is a routine
- 15:24check focusing on four critical areas of the
- 15:26system. Area one is app access. You need to verify
- 15:30what external tools Claude can still actually
- 15:33reach. Are those API connections still valid?
- 15:36Do you even use those apps anymore? Because sometimes
- 15:38an API key quietly expires in the background
- 15:41or a service gets deprecated. The AI needs to
- 15:44be updated on that reality. Totally. Area two
- 15:46is routine. Are your scheduled automated workflows
- 15:50still matching how you actually do your job today?
- 15:53And then area three is retrieval quality. This
- 15:56is where you sit down and review that index file
- 15:58all over again. You look at the logs from your
- 16:01recent work. Where did Claude open too many files
- 16:03yesterday? Right. Where did it grab a chunk of
- 16:06text from the wrong section last week? Where
- 16:09do you desperately need to write clearer labels?
- 16:11Yes. And that leads directly into area four,
- 16:14which is test limits. This is absolutely vital
- 16:17for long -term success. You have to ensure you
- 16:20are continually testing the system on real, incredibly
- 16:23complex projects. Not just tiny, isolated, perfectly
- 16:26clean sample folders. Exactly. A retrieval system
- 16:30might look like it works perfectly when you test
- 16:32it on five neatly organized text files. But that
- 16:35same logic might completely collapse when you
- 16:37point it at a six -month -old, deeply messy project
- 16:39directory full of conflicting drafts. Can a successful
- 16:42test on a small handful of sample files guarantee
- 16:45the system works? Never. Small environments lack
- 16:48friction. They artificially hide the structural
- 16:51flaws in your indexing logic. You need genuine
- 16:53messy complexity to prove the retrieval flow
- 16:56holds up under actual pressure. Right. No, small
- 16:59tests are misleading. You must verify it on real,
- 17:01messy projects. Push the system until it breaks,
- 17:04figure out why it broke, and then fix the index.
- 17:07Let's step back and look at the big picture here,
- 17:10the core philosophy driving a fabled 5 -second
- 17:12brain. It really represents a massive mindset
- 17:15shift in how we interact with local AI. It truly
- 17:18is. It is not about creating beautiful, complex
- 17:21visual graphs that make us feel organized. It
- 17:24is entirely about structured surgical retrieval.
- 17:27It is about giving Claude code. A clear, completely
- 17:30unambiguous map of your digital life. A definitive
- 17:33map of your apps, your routines, your memory,
- 17:35and your skills. When you take the time to do
- 17:38this, you stop the AI from wasting its token
- 17:41budget. You stop it from wasting your time by
- 17:43reading your entire workspace for every simple
- 17:45question. You successfully transform a chaotic
- 17:48pile of loose files into a highly functional,
- 17:51autonomous database. We have a challenge for
- 17:53you today. Open your own digital workspace. Just
- 17:55look at your desktop or your main documents folder.
- 17:58Try to view that space purely through the lens
- 18:01of those four layers. Ask yourself honestly,
- 18:03where is the memory actually stored? Where the
- 18:06routine's documented. Which leads us to a final
- 18:09thought to leave you with. If your AI is merely
- 18:12a reflection of the digital environment you force
- 18:14it to work in, what does your current folder
- 18:17structure say about how you think? That is a
- 18:20very tough mirror to look into for most of us.
- 18:22It really is. If you were an AI trying to navigate
- 18:25your own mind, based entirely on the state of
- 18:27your hard drive right now, where would you get
- 18:29lost first? Out TRO music.