Latest / AI Fire Daily / #456 Neil: Claude Projects Tutorial Builds An Elite Daily Memory Engine
Transcript
- 0:00You spend 20 minutes perfectly explaining a task.
- 0:04You get the details just right. The formatting
- 0:06is flawless. Then you wake up the next morning,
- 0:09you open a new chat window, and it has completely
- 0:13forgotten who you are. Yeah, it's the classic
- 0:15amnesiac chat bot problem. I mean, we've all
- 0:17been there. It's incredibly frustrating, honestly.
- 0:20You're basically training a new employee from
- 0:22scratch every single day. Well, welcome to today's
- 0:26deep dive. We're shifting away from those temporary
- 0:29AI chats. The goal today is building a permanent
- 0:32living environment. We're calling it a co -work
- 0:35operating system, and we're doing it using Claude
- 0:37projects. So we have a really fascinating journey
- 0:40mapped out for you today. First, we'll clearly
- 0:42define the difference between standard Claude
- 0:44chat and this new co -work concept. Then we'll
- 0:46unpack the underlying architecture. It relies
- 0:49on three surprisingly simple text files that
- 0:51give the AI persistent memory. After that, we're
- 0:54mapping out structured workstations. And finally,
- 0:57we'll look at the advanced workflows to keep
- 0:58this whole system incredibly fast and cheap.
- 1:01Before we jump into the mechanics, I have to
- 1:03confess something. I still catch myself pasting
- 1:06the exact same five paragraph instruction prompt
- 1:09every morning. I'm just so conditioned to the
- 1:12old way of working. I still wrestle with prompt
- 1:14drift myself. Listen, it makes complete sense
- 1:17that you do. We've all been trained over the
- 1:20last few years to treat AI like a disposable
- 1:22tool. You use it, you close the tab, and it vanishes.
- 1:26But to build a true AI operating system, you
- 1:29have to completely change the environment. We
- 1:32need to start by looking at the three distinct
- 1:34flavors of Claude that currently exist. Right.
- 1:36Let's break down that baseline. We need to understand
- 1:38the environment before we can build the system.
- 1:41Exactly. So first, you have Cloud Chat. This
- 1:43is the web browser version. Everybody knows.
- 1:45It's great for doing some quick market research.
- 1:47It's helpful if you just need to draft a quick
- 1:49apology email to a client. But it starts fresh
- 1:52every single session. It has absolutely zero
- 1:55memory of what you did yesterday. It's essentially
- 1:57a blank slate every time you hit Enter. Then
- 2:00you have Cloud Code. Now, this is a very different
- 2:02beast. It's a terminal -based tool built specifically
- 2:04for developers. Meaning a text -only screen where
- 2:08programmers type direct system commands. Precisely.
- 2:10It connects directly to a software code base.
- 2:12It can actually autonomously edit files, run
- 2:15tests, and execute commands. But it's highly
- 2:17specialized for software engineering. It's definitely
- 2:20not built for your general day -to -day workflow.
- 2:23So if chat is entirely temporary and code is
- 2:25strictly for programmers, there has to be a middle
- 2:27ground, something built for persistent, everyday
- 2:30knowledge work. And that brings us to the third
- 2:32flavor, Claude Cowork. This is the actual desktop
- 2:35application. It's engineered specifically for
- 2:38long -term compounding workflows. The true magic
- 2:41here isn't just the AI itself. It's the ability
- 2:44to natively mount local folders from your computer.
- 2:47It maintains a persistent, evolving memory. And
- 2:50it uses external connectors to link directly
- 2:52to your gisting tools, things like Gmail, your
- 2:55Google Calendar, or your Notion databases. OK,
- 2:58let me try to contextualize this. Standard Cloud
- 3:01Chat is basically a temp worker. They show up,
- 3:04do a decent job. but they get a complete memory
- 3:06wipe every single night. It's literally like
- 3:09that movie, 50 First Dates. Beat. But co -work
- 3:13is fundamentally different. It's like hiring
- 3:15a full -time partner. You're giving them a dedicated
- 3:18desk. You're giving them a permanent filing cabinet.
- 3:21That desk analogy is perfect. You're giving the
- 3:23AI a physical home on your machine. It doesn't
- 3:26exist in a vacuum anymore. It has a spatial awareness
- 3:29of your digital life. But that raises a pretty
- 3:32massive privacy concern for me. Does mounting
- 3:35a local folder mean the AI is just freely reading
- 3:37everything on my computer? Is it scanning my
- 3:39personal photos while I sleep? No, absolutely
- 3:42not. The permissions are strictly ring -fenced.
- 3:45It only reads the specific folder you explicitly
- 3:47mount and allow inside the application. It has
- 3:50zero access to the rest of your hard drive. So
- 3:52it's a fenced -in playground, not a master key
- 3:54to my hard drive. Right. You dictate the boundaries.
- 3:56It only sees what you place inside the fence.
- 3:59Now that we understand Co -Work acts like a dedicated
- 4:01filing cabinet, we need to know what actually
- 4:03goes inside those folders. How does it physically
- 4:06remember us? This is where it gets genuinely
- 4:08fascinating. You would think an advanced AI operating
- 4:11system requires massive, complex databases. But
- 4:15the whole system runs on a surprisingly simple
- 4:17architecture. You literally just need three plain
- 4:21text files sitting inside a root folder. We'll
- 4:24call that master folder your Co -Work OS. Just
- 4:26three files. I mean, that feels almost too simple.
- 4:29Yeah, but they're just markdown files. Meaning
- 4:31basic text files with simple formatting headers
- 4:34and bullet points. Right. There's no complex
- 4:36programming logic. No Python scripts. The first
- 4:39file is called Claude .md. Think of this file
- 4:43as the master operating manual or the global
- 4:46routing map. When you open a new session, this
- 4:48file tells Claude what to load and where to look.
- 4:51It establishes the fundamental ground rules for
- 4:53how it should behave. Yes, and it has three core
- 4:56components. First is the memory system itself.
- 4:59It instructs Claude to always check your pass
- 5:01context before answering. Second is the routing
- 5:04map. This prevents the AI from guessing. If you
- 5:07ask a question about your taxes, the map routes
- 5:09it straight to your financial folder. Third,
- 5:12you have references. These are pointers that
- 5:14tell Claude where to find extra context, but
- 5:17only when it actually needs it. Two secs silence.
- 5:19I really have to push back here. It feels wild
- 5:21to me. The absolute cutting edge of artificial
- 5:24intelligence relies on basic text files from
- 5:26the 1990s to function. And the sources are very
- 5:29strict about this. They say we absolutely must
- 5:31cap this Claude .md file at 300 lines. Why cap
- 5:35it? If this is my master operating manual, why
- 5:37not dump every single rule, preference, and workflow
- 5:40I have in there? It's a really common trap people
- 5:42f— fall into to understand the cap, you have
- 5:45to understand how token limits actually work
- 5:46under the hood. Meaning the maximum amount of
- 5:49text the AI can process simultaneously. Think
- 5:51of it like the AI's cognitive battery power for
- 5:54a single conversation. Your large root files
- 5:57load every single time you start a new session.
- 6:01Every time you say hello, Claude has to silently
- 6:03read that entire Claude .md file first. If your
- 6:07manual is 20 pages long, it slowly drains your
- 6:10token limits before you even start working. It
- 6:12costs more money. It severely slows down the
- 6:15response time. Leaving barely any processing
- 6:17power to actually complete the task I asked for.
- 6:20Exactly. You're giving it cognitive overload.
- 6:22That's why you keep that main routing file under
- 6:24300 lines. It should be a traffic cop, not encyclopedia.
- 6:27Well, that makes a lot of sense. So if Claude
- 6:29.md is the traffic cop, what's the second file
- 6:31handling? The second file is memory .md. This
- 6:35functions as your running ledger. It holds your
- 6:37active, ongoing projects and your saved personal
- 6:40preferences. This specific file is the magic
- 6:43ingredient that stops your conversations from
- 6:45starting at zero every day. It's the actual continuity
- 6:48mechanism. Yeah, it tracks the reality of your
- 6:50current work. We usually split it into two sections.
- 6:53One section tracks active projects so it knows
- 6:56your upcoming newsletter is currently in the
- 6:57rough draft stage and you're struggling with
- 6:59the conclusion. The second section tracks granular
- 7:03preferences. Like knowing a specific client absolutely
- 7:07hastes bullet points and prefers numbered lists.
- 7:10Okay, and the third file in this architecture?
- 7:12That one is called VoicePrinciples .md. We usually
- 7:15drop this into a subfolder called Resources.
- 7:18This file's entire job is teaching the AI exactly
- 7:21how you write and communicate. We'll definitely
- 7:24dig into that voice file in a minute because
- 7:26sounding human is huge. But first, just managing
- 7:29these individual markdown files sounds potentially
- 7:32messy. If I'm manually editing text files all
- 7:34day, I feel like I'm stepping backward in productivity.
- 7:37That's a totally valid concern, and it's why
- 7:39a lot of power users rely on an app called Obsidian.
- 7:42It's a completely free markdown editor. You don't
- 7:45have to migrate anything. You just open your
- 7:47existing Cowork OS folder as a vault inside Obsidian.
- 7:52It instantly gives you a beautiful visual interface.
- 7:55You can easily view, link, and manage all these
- 7:58text files in one clean dashboard. But practically
- 8:01speaking, do I have to manually open that dashboard
- 8:04and type into memory .md every single time I
- 8:07finish a task? That feels like a lot of friction.
- 8:10No, not at all. That's the beauty of the system.
- 8:12You just tell Claude in your normal chat window
- 8:14to remember this. The AI takes your natural language
- 8:17instruction. reformats it, and silently writes
- 8:20it to the Markdown file automatically. You just
- 8:22talk to it and it updates its own memory notes.
- 8:24It completely removes the administrative burden.
- 8:26You just do your work and the system documents
- 8:29itself. Okay, so the AI can take its own notes.
- 8:31But let's circle back to that third file. How
- 8:33do we actually stop it from sounding like a generic,
- 8:36overly enthusiastic robot? And as we scale this
- 8:39up, how do we keep it from mixing up completely
- 8:41different parts of your life? Those are two separate
- 8:43but incredibly important challenges. That's where
- 8:46the voice system and the concept of workstations
- 8:48come into play. Let's tackle the voice system
- 8:50first. How do we actually build that voice principles
- 8:53file without it taking weeks of training? Well,
- 8:56you have to feed the AI real historical examples
- 8:59of your authentic writing. If your Gmail is already
- 9:02connected to your cowork app, you can literally
- 9:05just have it analyze your last 50 cent emails.
- 9:08If you don't want to connect your email, just
- 9:09copy and paste five to 10 past articles or reports
- 9:12you've written right into the chat. You ask it
- 9:15to reverse engineer your tone. Yes. But here's
- 9:18the critical failure point for most people. Vague
- 9:21rules fail miserably. If you tell Claude to be
- 9:25warm and professional, you're going to get terrible
- 9:27generic AI output. It'll sound like every corporate
- 9:30memo ever written. You have to extract highly
- 9:33specific mechanical rules. You tell it, keep
- 9:36paragraphs under three sentences. Or you instruct
- 9:39it, never use passive voice and avoid exclamation
- 9:41points. Because specificity is what creates repeatable
- 9:44behavior. A machine doesn't understand warmth,
- 9:46but it perfectly understands a maximum word count.
- 9:49And it's an iterative process. Over the first
- 9:52few weeks, you'll catch it making mistakes. You
- 9:55correct the output. You tell it to remember that
- 9:58specific correction. Over time, that voice principles
- 10:01file evolves into a living, incredibly accurate
- 10:05record of your actual voice. So that solves the
- 10:07robot tone. What about keeping my life organized?
- 10:11That requires building out workstations. Think
- 10:14of workstations as dedicated subfolders for specific
- 10:17domains of your life. I really like comparing
- 10:19workstations to the rooms in a physical house.
- 10:22You wouldn't keep your kitchen blender sitting
- 10:24on your bed. It makes no sense. Similarly, your
- 10:27highly confidential mortgage refinancing notes
- 10:29absolutely shouldn't bleed into your weekly creative
- 10:32writing drafts. That's a brilliant way to conceptualize
- 10:34it. You need structural boundaries. And we generally
- 10:37see two distinct types of workstations you can
- 10:40build within this system. So if one type handles
- 10:42broad categories, what does a universal workstation
- 10:45look like? Universal workstations cover broad
- 10:47functional areas that touch multiple parts of
- 10:50your life. An email HQ is the perfect example.
- 10:53You send emails for your day job, for your personal
- 10:55life, maybe for side hustle clients. The core
- 10:57rules of what makes a good email for you, like
- 10:59your sign off and formatting, apply everywhere.
- 11:02So that lives in a universal folder. And the
- 11:04second type, I imagine it's something highly
- 11:06siloed. Right. Dedicated workstations. These
- 11:09cover one highly specific contained domain. a
- 11:12folder strictly for personal finances, a folder
- 11:15dedicated to a complex upcoming international
- 11:17trip, or a folder locked down for one specific
- 11:21client launch. And each one of these individual
- 11:24workstations gets its own isolated setup. Yes.
- 11:27Each dedicated workstation gets its own local
- 11:29CLAW .md file and its own localized memory .md
- 11:33file. The structure perfectly mirrors your main
- 11:35root folder, but the rules inside only apply
- 11:37when you're actively working inside that specific
- 11:40space. Let me test the logic here. If I create
- 11:42a really fantastic rule for formatting emails
- 11:45inside my finance workstation, should I copy
- 11:48paste that rule up into my universal email HQ?
- 11:51No, never duplicate. You should never repeat
- 11:54global rules in multiple places across the system.
- 11:57It's a core design principle. Duplicate rules
- 11:59waste your token limits, and worse, they cause
- 12:02inconsistent AI behavior when the system tries
- 12:05to reconcile two slightly different versions
- 12:07of the same rule. Write the rule once at the
- 12:10highest level. Never duplicate it. Exactly. It
- 12:12keeps the entire architecture clean, efficient,
- 12:15and predictable. Sponsor's hearing. OK, so we've
- 12:19built the house. We furnished the individual
- 12:21rooms with workstations. What does actually living
- 12:24and working inside this system look like on a
- 12:27random Tuesday afternoon? And crucially, how
- 12:30do we pay the electric bill? Meaning, how do
- 12:33we manage the token costs so we don't go broke
- 12:35running this thing? Let's look at some advanced
- 12:37daily workflows. When you integrate those external
- 12:39connectors, the capability styrockets. The first
- 12:43amazing use case is automatic meeting follow
- 12:45-ups. Walk me through the actual mechanics of
- 12:47that. How does it happen? Imagine your Google
- 12:49Calendar is securely connected to Cowork. You
- 12:52also use a standard tool that records and transcribes
- 12:55your Zoom meetings. You simply open your chat
- 12:57and ask Claude to handle the follow -up for your
- 13:002 p .m. meeting. Just one simple natural language
- 13:03request. That's it. Claude pings the calendar
- 13:05API to see who is in the 2 p .m. slot. It pulls
- 13:08the WAW transcript from your local folder. It
- 13:11reads the entire conversation. identifies the
- 13:13agreed -upon next steps, and figures out who
- 13:16is responsible for what. It then dynamically
- 13:18routes to your email HQ workstation to grab your
- 13:21communication rules. Finally, it drafts a perfectly
- 13:24formatted follow -up email using your exact tone,
- 13:27all from one single prompt. That is incredible
- 13:29leverage. What about applying this to project
- 13:31management? The sources mention Notion. Notion
- 13:34auto projects are a huge unlock. Let's say you
- 13:37tell Claude you're traveling to Boston for a
- 13:39tech conference in June. Claude automatically
- 13:41connects to your Notion workspace. Because it
- 13:44checks your memory file, it already knows exactly
- 13:47how you like your travel properties formatted.
- 13:50It knows you like a column for confirmation numbers
- 13:52and a separate view for daily itineraries. It
- 13:55builds the complete customized travel board autonomously.
- 13:59It's not just generating text anymore. It's generating
- 14:01customized software environments based on my
- 14:04historical preferences. Precisely. But if you
- 14:06want to see the true power of persistent memory,
- 14:08the finance tracking is where it gets really
- 14:10wild. Tell me about the finance workstation.
- 14:12How is that different from just using a spreadsheet
- 14:14macro? Whoa. Beat. Imagine it just silently watching
- 14:18a messy folder of bank statements. You literally
- 14:21just drop raw PDF statements into the mounted
- 14:23folder. It instantly reads them. It categorizes
- 14:26an entire year of transactions. It groups your
- 14:28software subscriptions, your travel expenses,
- 14:30your dining out. Then it generates a clean Excel
- 14:32summary. It's doing hours of tedious manual data
- 14:35entry in seconds. But AI hallucinates. What happens
- 14:39when it inevitably makes a mistake? What if it
- 14:42categorizes a business lunch as a personal grocery
- 14:45expense? You just correct it once in the chat.
- 14:47You say, hey, anything from this specific restaurant
- 14:48is a business expense. Remember this category
- 14:51rule. It instantly writes that logic to the local
- 14:54memory .md file inside the finance workstation.
- 14:57It will never make that specific mistake again.
- 15:00The system actually learns. Two -sex silence.
- 15:03That is genuine compounding leverage. But processing
- 15:07dozens of PDFs, reading calendar APIs, managing
- 15:11Notion databases, doing all of this continuously
- 15:13must cost an absolute fortune in API credits.
- 15:16It definitely can if you aren't careful with
- 15:18your architecture. You need strict cost control
- 15:20measures in place. And the biggest secret to
- 15:22keeping the bill low is intelligent model selection.
- 15:25Explain how we manage that. Are we restricting
- 15:26what it can do? Not restricting, just optimizing.
- 15:29You should use Claude 3 .5 Sonnet for 80 % of
- 15:32your routine tasks. Sonnet is incredibly capable.
- 15:34It's very fast. and it is significantly cheaper
- 15:37to run. You save the heavier model, Claude Erpus,
- 15:40exclusively for highly complex multi -step reasoning
- 15:43tasks. So, Sonnet is for the daily grind. Sorting
- 15:47the emails, summarizing the transcripts, Obis
- 15:50is for the heavy lifting. like analyzing a complex
- 15:53legal contract. Exactly. You don't need a supercomputer
- 15:56to draft a calendar invite. And the other cost
- 15:59-saving measure brings us back to those text
- 16:01files. Keep your route instructions ruthlessly
- 16:04short. Let the routing map pull detailed contests
- 16:07from deeper subfolders only when it is actually
- 16:10needed for the task at hand. Before we wrap up,
- 16:12the sources mentioned a specific habit to build.
- 16:15What exactly is a session audit before closing
- 16:18a chat window? It's a really powerful, simple
- 16:21prompt to use at the end of your day. Before
- 16:23you hit the close button and wipe that temporary
- 16:26context window, you just ask Claude to scan our
- 16:28current conversation. You ask it to find anything
- 16:30worth saving to the memory files and update them.
- 16:33It's a quick debrief to ensure nothing valuable
- 16:35gets forgotten. Exactly. It ensures the system
- 16:38keeps learning and compounding without you having
- 16:41to sit there doing manual data entry or trying
- 16:43to remember what you discussed three hours ago.
- 16:45This whole conversation really highlights a massive
- 16:48shift in how we interact with machines. We're
- 16:51looking at a profound paradigm shift. For the
- 16:54last two years, everyone has been obsessed with
- 16:56finding the perfect combination of words. The
- 16:58age of obsessing over individual prompt engineering
- 17:01is ending. It really is. The focus on the magic
- 17:04prompt was a symptom of temporary AI. We're entering
- 17:07a completely new era. The new era is context
- 17:10engineering. The ultimate utility of artificial
- 17:13intelligence no longer depends on perfectly wording
- 17:17a single prompt in an empty chat box. It depends
- 17:20entirely on building an organized, logical folder
- 17:23infrastructure around the AI. It's about building
- 17:26a robust foundation. If the foundation is solid,
- 17:29the AI does the heavy lifting effortlessly. Your
- 17:31Co -Work OS is essentially a living database
- 17:34of your professional life, and it compounds in
- 17:36value over time. every single session, every
- 17:39single correction, teaches the system a little
- 17:41more about how you operate and how you think.
- 17:43And that's the most beautiful part of this entire
- 17:45framework. You don't need a perfect, massive
- 17:47system on day one. In fact, you shouldn't try
- 17:49to build one. You just need a working system.
- 17:51So here is your call to action for today. Start
- 17:54incredibly small. Build just your master root
- 17:57folder. Set up your claw .md file. Maybe add
- 18:00two simple workstations, perhaps an email HQ
- 18:03and a project folder. Get comfortable with how
- 18:06the active memory talks to the routing file.
- 18:08Yeah, let your real actual daily needs dictate
- 18:11the structure. Don't build elaborate workstations
- 18:13you don't actually use just because they look
- 18:15nice in a folder tree. It all comes back to that
- 18:17initial frustration we talked about at the start,
- 18:20the amnesiac chatbot. You don't have to live
- 18:23with a tool that forgets you every single morning
- 18:25anymore. You can build a system that actually
- 18:27remembers. Right. You're building a permanent
- 18:29partner. But stepping back, it leaves us with
- 18:31something much deeper to consider. If this persistent
- 18:34compounding system perfectly mimics your writing
- 18:37style over time, if it learns your organizational
- 18:41habits and begins anticipating your decision
- 18:43-making processes, will there come a day where
- 18:46you actually forget where your natural workflow
- 18:48ends and the AI's structural influence begins?
- 18:52Beat. What happens when your artificial coworker
- 18:54knows your own habits and preferences better
- 18:56than you do?