Latest / AI Fire Daily / #521 Neil: Claude Code Tutorial Build a System That Never Resets
Transcript
- 0:00Think about how much of your brain you lose every
- 0:02time you close a browser tab. You spend, like,
- 0:06three hours teaching an AI exactly how your business
- 0:08works. You feed it context, you explain your
- 0:11preferences, and the moment you hit new chat,
- 0:14it just forgets you entirely. It is essentially
- 0:16a digital lobotomy. Peep neat. But imagine a
- 0:20system that remembers everything. automatically.
- 0:24Yeah, one that learns from your real work. Right.
- 0:26It actively compounds your knowledge over time.
- 0:29It's the ultimate shift in leverage, really.
- 0:31You stop doing manual labor. You stop repeating
- 0:33yourself daily. You start building a permanent
- 0:35cognitive partner instead. Welcome to this deep
- 0:39dive. Today, we are exploring a massive workflow
- 0:42shift. We are unpacking the BUILB framework.
- 0:46Build. Exactly. This mission is about transforming
- 0:49cloud code entirely. We're turning it into a
- 0:51self -improving personal knowledge base. It's
- 0:53a five -step framework, right? We have base,
- 0:56upload, inflow, loop, and drive. It sounds a
- 0:59little technical at first. We're going to break
- 1:01it down completely. Yeah, this genuinely changes
- 1:02how you interact with AI. So we need to unpack
- 1:05this carefully. Before we can automate anything
- 1:07at all, we have to build the physical house.
- 1:10This is where the AI will actually live. The
- 1:13foundation. Right, the base phase of the framework.
- 1:15You can think of the raw folder like a messy
- 1:17garage. It's full of unorganized boxes and files.
- 1:21Yeah, just raw storage. But the wiki folder is
- 1:24completely different. The wiki is a meticulously
- 1:27cross -referenced catalog. That's the perfect
- 1:31way to visualize it. Most tutorials, you know,
- 1:33they skip this foundation entirely. But this
- 1:36structure holds absolutely everything together.
- 1:38Do you need two main directories? Yes. One is
- 1:41called slash raw. The other is called slash wiki.
- 1:44I'm looking at this folder structure right now.
- 1:46And part of me thinks this is overly complicated.
- 1:48Oh, it feels that way at first. I mean, if Claude
- 1:50is so smart, why can't I just dump everything
- 1:52into one giant folder? Why add this extra layer
- 1:55of complexity? Because dumping everything creates
- 1:57complete chaos. Yeah. The raw folder holds all
- 2:01your original source material. Like what exactly?
- 2:03Well, it holds massive 50 page transcripts. It
- 2:06holds your rough notes. It holds voice recordings
- 2:09in their original form. Everything sits there,
- 2:11just waiting to be processed. OK, so the garage
- 2:14is full of giant boxes. Right. And if the AI
- 2:17has to open every single box to answer a simple
- 2:20question, it slows to a crawl. It gets completely
- 2:23overwhelmed. Makes sense. That's where the wiki
- 2:25folder comes in. The wiki folder doesn't store
- 2:27the actual content. It stores indexed references
- 2:30instead. It essentially acts as a master table
- 2:33of contents. Exactly. The AI doesn't dig through
- 2:35the heavy boxes in the garage. It just looks
- 2:37at your clipboard, finds the box number, and
- 2:40goes straight to it. It finds exactly what it
- 2:42needs incredibly fast. Super fast. That makes
- 2:44a lot of sense logically. But how does the AI
- 2:48actually know to do that? An empty folder structure
- 2:50doesn't force a behavior. That brings us to the
- 2:53clod .md file. This file represents the absolute
- 2:57house rules for your project. The house rules?
- 3:00Yeah. It tells the AI exactly how your project
- 3:02is organized. You give it a very simple system
- 3:05prompt here. You essentially dictate the baseline
- 3:08behavior for the entire system. Yes. You tell
- 3:10it to always check the wiki first. You tell it
- 3:13to analyze new content automatically. The AI
- 3:16reads this file every time you open a session.
- 3:19The clearer those instructions are, the better
- 3:21the results. Without a doubt. It sets the priorities.
- 3:24OK, so we have folders and house rules, but we
- 3:27need a way to actually move the data around.
- 3:30that requires skills. A skill is really just
- 3:33a reusable instruction set. It handles a specific
- 3:36task the exact same way every time. So if you
- 3:38do something twice, you should turn it into a
- 3:41skill. Exactly. Let's ground this with a real
- 3:43example. The very first skill you need is called
- 3:46Add New Resource. You simply drop a new PDF into
- 3:49the raw folder. And the skill automatically triggers.
- 3:52Right. It reads that dense PDF document. Right.
- 3:55And it creates a new wiki entry summarizing it.
- 3:58then it links back to the original file in the
- 4:00raw folder. It does the heavy lifting for you
- 4:02automatically. Yeah. But we should clarify the
- 4:04types of skills here. We have utility skills
- 4:07and orchestration skill. Okay, break that down.
- 4:09A utility skill does one very focused job, like
- 4:11that add new resource skill we just discussed.
- 4:13It just summarizes and links. Very simple. But
- 4:16an orchestration skill is more complex. A master
- 4:19skill that triggers several smaller tasks in
- 4:22order. A master skill that triggers several smaller
- 4:24tasks in order. That's a helpful distinction.
- 4:27Two secs silence. Before we move on, I want to
- 4:29nail this down. Why do we strictly separate the
- 4:34raw files from the wiki instead of combining
- 4:36them? Because feeding the AI massive raw transcripts
- 4:40every time is horribly inefficient. The wiki
- 4:43acts as a lightning fast index, giving the AI
- 4:46quick answers without drowning it in data. So
- 4:49raw is for storage and the wiki is for speed.
- 4:52Precisely. So we've got this beautifully organized
- 4:54empty house, but an empty house doesn't do the
- 4:57dishes or anticipate your needs. No, it just
- 4:59sits there. To make it a cognitive partner, we
- 5:01have to feed it your past, which leads us to
- 5:04priming the engine. This is the upload phase.
- 5:06A lot of people want to jump straight to the
- 5:08automation phase. That's a huge mistake. They
- 5:09want the flashy stuff immediately. Exactly. But
- 5:13before pipelines work, you need to prime the
- 5:15engine with real data. You want to start with
- 5:17real depth. You don't want to start with a blank
- 5:19page. So where do we actually find this digital
- 5:23furniture? I mean, I don't want to sit down and
- 5:24type out a memoir for my AI. You don't have to.
- 5:28The best place to start is your old Claude session
- 5:30history. Oh, interesting. Yeah, this data is
- 5:33incredibly valuable. It shows the exact questions
- 5:35you tend to ask. It highlights where you usually
- 5:38get stuck. It's a record of your actual behavior.
- 5:40It's not just guesswork about how you might work.
- 5:43Right. Claude code saves this history locally
- 5:45on your machine. You can pull this history right
- 5:48into your project. And what does it do with it?
- 5:50The system will analyze it for recurring patterns.
- 5:54It will even suggest new skills based on those
- 5:56patterns. That's fascinating. It analyzes your
- 5:59past to optimize your future workflow. But session
- 6:02history is just one part of the puzzle. What
- 6:04about the rest of my digital life? That's your
- 6:06personal ecosystem data, the files and emails
- 6:09you create every single day. The messy stuff.
- 6:11The very messy stuff. But it's a true record
- 6:14of your thinking and writing style. You can export
- 6:17this using tools like Google Takeout. Or you
- 6:19can use the standard Outlook export tool. Yes,
- 6:22either works perfectly. And I just drop these
- 6:24massive zip files right into the raw folder?
- 6:27Well, you extract them first. Yeah. But yes,
- 6:29you drop the files into the raw folder. Then
- 6:31you ask the system to analyze your writing patterns.
- 6:34What does it look for? It learns your recurring
- 6:36topics. It finds areas where it can help you
- 6:38work more efficiently. I have to admit something
- 6:40vulnerable here. I still wrestle with prompt
- 6:44drift myself. I am constantly re -explaining
- 6:47who I am to these models. It's completely exhausting.
- 6:50Oh, it's the worst. That's exactly why the next
- 6:53step is so crucial. You need to share your personal
- 6:56goals as training data. Like, explicit goals?
- 6:59Yes. Most people never even think to do this.
- 7:02They expect the AI to just guess their motivations.
- 7:05How does that actually work in practice? You
- 7:08write out your background and your long -term
- 7:09goals. You write down what you're trying to build
- 7:12this year. Could it be audio? Absolutely. You
- 7:14can even just record a quick voice note while
- 7:16walking. You drop that audio file into the raw
- 7:18folder, then the AI turns it into structured
- 7:21training data. Yes. It captures your values and
- 7:25the reasoning behind your decisions. It gives
- 7:27the AI the personal context that usually only
- 7:31lives in your head. And that personal context
- 7:33travels with you into every future session. Which...
- 7:36completely eliminates the prompt drift you were
- 7:38just talking about. That is huge. It really is.
- 7:40Once you have all these files gathered, you run
- 7:43one bulk ingest session. Using that skill we
- 7:45talked about. Yep. You run the add new resource
- 7:48skill on everything at once. It processes all
- 7:51that history and context in one sitting. Your
- 7:53project suddenly has a deeply personalized foundation.
- 7:56It actually knows you. Yeah. To sex silence.
- 7:59But regarding all this ecosystem data, I'm a
- 8:02bit skeptical. How much personal data is actually
- 8:05too much to upload? If the data doesn't actively
- 8:08help the AI understand your current workflow,
- 8:11leave it out. You want high signal context, not
- 8:13a massive dump of your entire doozle history
- 8:15from 10 years ago. Only keep data that actively
- 8:18serves your current goals. Exactly. Keep it lean.
- 8:21That brings us to the third phase, the inflow
- 8:24phase. Think of your project like a pristine
- 8:26mountain lake. A lake without fresh water flowing
- 8:29in quickly goes stale. We need to build rivers
- 8:32to keep the data fresh. This is where we finally
- 8:34talk about automation. But we have to be careful.
- 8:37Why is that? Most tutorials fail because they
- 8:41automate way too early. They try to connect 10
- 8:44different apps on day one. You end up with a
- 8:46broken pipeline that you cannot easily debug.
- 8:48It's just a tangled mess of code. The golden
- 8:51rule here is skill -driven data ingestion. You
- 8:54build the skill first, you test it locally on
- 8:57your machine. You make sure it extracts exactly
- 8:59what you want. Right. Only then do you actually
- 9:01automate it. Okay, let's break down these rivers.
- 9:03The first pipeline focuses on syncing your ongoing
- 9:06cloud sessions. Because your needs change over
- 9:08time. The questions you asked today are very
- 9:11different from last month. Makes sense. This
- 9:13pipeline pulls your latest session exports directly
- 9:16into the raw folder. It runs the add new resource
- 9:18skill to index them into the wiki. This keeps
- 9:21the system perfectly synced to your current reality.
- 9:24The second pipeline is syncing your ongoing ecosystem
- 9:27data. This pulls from places where you create
- 9:30useful information weekly. Like your meetings.
- 9:33If you use an app like Granola, it records in
- 9:35the background. You can pull those transcripts
- 9:37directly into your system. Or Slack. You can
- 9:39connect cloud code to pull your chat history.
- 9:42You can even pull captions from public YouTube
- 9:44videos. The specific app really doesn't matter
- 9:47here. Not at all. The principle is finding where
- 9:50your useful information lives. You build a direct
- 9:53path from there into your project. But wait,
- 9:56if it runs every day, you want to just download
- 9:58the same files over and over? That's a great
- 10:01catch. When you build this skill, the system
- 10:04tracks async cursor. Async cursor. Let's define
- 10:07that. A digital bookmark remembering exactly
- 10:10what was already downloaded. Right. It prevents
- 10:12the system from repulling the exact same content
- 10:15every single run. So it's smart about it. Yeah.
- 10:17It only looks for new data created after that
- 10:19bookmark. That's highly efficient. Now, the third
- 10:22pipeline is syncing curated newsletters. This
- 10:25is about bringing in high value outside knowledge.
- 10:28Almost every niche has incredible newsletters.
- 10:30but they usually clog up your inbox. Oh, constantly.
- 10:33A great trick is adding Plus Newsletter to your
- 10:36Gmail address, like johnplusnewsletter .gmail
- 10:39.com. It keeps them entirely separate from your
- 10:42main inbox. You build a skill to pull emails
- 10:45from that specific folder. And it doesn't just
- 10:48save the email. It extracts the key insights,
- 10:51finds the arguments, and updates your wiki automatically.
- 10:54The fourth pipeline might be the most powerful
- 10:56one. It's adding your personal daily reflections.
- 10:59This captures the lived experience that never
- 11:01shows up in an email. It captures the nuanced
- 11:04lessons you learn each day. The quiet realization.
- 11:07Exactly. You just use voice to text on your phone.
- 11:10You talk through a quick decision you made that
- 11:12day. You drop that short text file into the raw
- 11:14folder. A few sentences is honestly all it takes.
- 11:17The AI extracts the insights and updates your
- 11:20wiki. It's incredibly low friction. Two secs
- 11:23silence. With all these automated inputs coming
- 11:26in constantly, how do you filter out the noise
- 11:28from low -quality newsletters? You have to be
- 11:31brutally selective about what you actually subscribe
- 11:33to. If you dump garbage into the system, you
- 11:36get noise instead of clarity back out. Gerbic
- 11:39in means garbage out, so curate ruthlessly. Always.
- 11:43Mid -roll sponsor read. Welcome back to the deep
- 11:46dive. We're moving to the fourth phase of the
- 11:48framework. This is the loop phase. We've got
- 11:51a lot of water flowing in now. Right. With all
- 11:53this fresh water rushing in, we have a new problem.
- 11:55The system will flood if we don't build automated
- 11:58filters. We definitely need automated improvement
- 12:01cycles. But we have to avoid the trap of full
- 12:04automation. Full automation sounds amazing on
- 12:07paper. It does, but it's actually dangerous in
- 12:09practice. I don't buy that entirely. I mean,
- 12:12isn't the whole point of AI to automate everything?
- 12:14Why is full automation so dangerous here? Well,
- 12:17it's like an automatic gym program that only
- 12:19trains your chest for six months. Just bench
- 12:21press every day. Exactly. Technically, you're
- 12:24doing work. Technically, you're improving that
- 12:26one area. But you're quietly breaking the balance
- 12:29of everything else. Oh, I see. The system might
- 12:32optimize for the wrong thing entirely. It might
- 12:35rewrite summaries to be too short or delete helpful
- 12:38context. Exactly. You have to stay in control
- 12:41of the critical changes. To do that, we use the
- 12:44three -bucket review system. OK, break that down.
- 12:47You build a specific orchestration skill called
- 12:49improve system. It reviews all new content and
- 12:53suggests improvements. And it sorts these suggestions
- 12:56into three distinct buckets. Let's walk through
- 12:58them. What's the first bucket? The first bucket
- 13:00is auto -approve. These are low risk fixes. Things
- 13:04like cleaning up data bloat, fixing formatting,
- 13:06or repairing broken links. Basic maintenance.
- 13:09Right. The AI handles these automatically and
- 13:12just logs them in a text file. You don't even
- 13:14need to look at those. But the second bucket
- 13:16is where you maintain control. It's the need
- 13:18sign off bucket. Yes. These are higher stakes
- 13:21changes. This includes editing core skills or
- 13:23changing foundational wiki entries. So it doesn't
- 13:26just do it blindly. No. These go into a review
- 13:29file for your explicit approval. The AI prepares
- 13:31the change, but it won't execute it. It waits
- 13:33for your permission. And the third bucket. The
- 13:36third bucket is for cases requiring more context.
- 13:39The AI encounters conflicting data. Like if two
- 13:42transcripts contradict each other? Exactly. The
- 13:45AI pauses entirely and asks for your input before
- 13:48proceeding. Now we connect all these routines.
- 13:51You create a master orchestration skill for data
- 13:54ingestion, which runs all your sync pipelines
- 13:56in sequence. You schedule this to run automatically,
- 13:59maybe Tuesdays and Fridays at nine in the morning.
- 14:02Then you schedule the improved system skill to
- 14:04run afterward. Keeping ingestion and improvement
- 14:06as completely separate scheduled routines is
- 14:09smart. It makes it incredibly easy to see what
- 14:12broke if something goes wrong. If they run together,
- 14:15Debugging is a total nightmare. Separating them
- 14:18keeps the system stable. Whoa. Imagine your system
- 14:22literally suggesting its own brain upgrades while
- 14:25you sleep. It's like waking up to a smarter assistant
- 14:27every day. You just grab a coffee, review the
- 14:29file, and click approve. It's wild. Yeah. And
- 14:32that review file is your main touch point with
- 14:34the system. You can approve a change, reject
- 14:36it entirely, or say approve and don't ask again.
- 14:39That last option is powerful. It teaches the
- 14:43AI your preferences. It cuts down on repetitive
- 14:46decisions over time. Two sec silence. What is
- 14:49the actual risk of the AI overwriting a crucial
- 14:52core skill by accident? That's the exact danger
- 14:55of ignoring the bucket system. If you fully automate,
- 14:59it might rewrite a skill poorly and break your
- 15:02entire workflow. Automate the filing, but never
- 15:05automate the thinking. That's the only way to
- 15:07stay safe. This brings us to the final phase.
- 15:10The drive phase. You're finally behind the wheel.
- 15:13The engine is built. The filters are on. The
- 15:15rivers are flowing. How do you actually get behind
- 15:18the wheel without overthinking it? The biggest
- 15:21hurdle here isn't technical at all. It's psychological.
- 15:24People get intimidated. They get paralyzed by
- 15:26the sheer complexity of a blank terminal window.
- 15:29You have to remember that slow is smooth and
- 15:31smooth is fast. You don't build five pipelines
- 15:34on day one. No, you build this step by step.
- 15:36You set up the base first. You ingest the old
- 15:38data. You take your time. Brushing the foundation
- 15:40is why most people fail. They want the magic
- 15:43immediately. But you are the leader here. The
- 15:46system serves you. If a workflow stops being
- 15:48useful, you just delete it. You don't need permission
- 15:51to simplify your own workflow. Exactly. And you
- 15:53must compress your feedback loops. This is critical.
- 15:56How so? Well, if a skill produces weak text,
- 16:01don't just manually rewrite the text in a document.
- 16:04Right, because if you manually rewrite it, the
- 16:06AI learns nothing. Right. You correct it in the
- 16:08chat interface instead. You tell Claude to update
- 16:11the underlying skill directly based on your correction.
- 16:14The next run will be significantly better. Exactly.
- 16:17You fix the machine, not just the output. To
- 16:19sex silence. How do you recommend overcoming
- 16:23that initial intimidation of building this entire
- 16:26system? You stop looking at the final automated
- 16:28loop. It's too overwhelming. You just focus on
- 16:31the very first step. Create the two folders today
- 16:33and then just stop. Just build the two base folders
- 16:36today and stop. Take the very first step. The
- 16:38rest will follow naturally. Let's synthesize
- 16:40the big idea here. A working setup that gets
- 16:441 % better every week is powerful. It will always
- 16:48wildly outperform a perfect setup that never
- 16:50leaves your head. We spend so much time theoretically
- 16:53optimizing our tools, we just need to start building
- 16:56the physical foundation. You're no longer stacking
- 16:59Lego blocks of data every single morning. You
- 17:01aren't starting from zero. You're building a
- 17:04compounding intelligence that scales with you.
- 17:06It fundamentally changes your relationship with
- 17:08the machine. It becomes a partner, not just a
- 17:10tool. It's the highest leverage thing you can
- 17:12do this week. I highly recommend you start just
- 17:14step one today. Build the base. Set up those
- 17:17two folders. Write that simple clod .md file.
- 17:21You'll immediately feel the difference in your
- 17:22workflow. We're going to leave you with a provocative
- 17:24thought to mull over. If you train a local AI
- 17:27system to perfectly mirror your goals over several
- 17:29years, if it deeply understands your values,
- 17:32your decisions, and your daily reflections, what
- 17:35actually happens when that localized brain outlives
- 17:37the current AI models? Have you just built your
- 17:40first truly portable digital soul? Out to your
- 17:43own music.