Latest / AI Fire Daily / #86 Robin: The "God Mode" Setup - Syncing Claude Cowork OS Across Desktop, Web, and Mobile Using Dropbox

Transcript

#86 Robin: The "God Mode" Setup - Syncing Claude Cowork OS Across Desktop, Web, and Mobile Using Dropbox · AI Fire Daily

• 0:00
  You carry a genius AI on your phone today. It
• 0:03
  is incredibly powerful and endlessly knowledgeable.
• 0:07
  Yeah, it really is. But then you just close your
• 0:09
  laptop screen. Right. Suddenly, that pocket genius
• 0:13
  develops complete amnesia. It is a massive modern
• 0:17
  problem. It cannot see your closed local files
• 0:19
  anymore. No, it cannot. It completely forgets
• 0:22
  exactly who you actually are. Yeah. This sync
• 0:25
  gap breaks everything. We expect our AI to follow
• 0:27
  us everywhere effortlessly. Welcome to today's
• 0:31
  deep dive. We are exploring a fascinating architectural
• 0:34
  challenge today. Exactly. We need our digital
• 0:37
  assistants to be truly Amida present. Our mission
• 0:40
  today is bridging that frustrating sync gap.
• 0:43
  We want to run one unified Claude co -work system.
• 0:46
  It needs to work seamlessly across desktop. web,
• 0:49
  and mobile. So we are building a completely new
• 0:52
  architecture. Yeah, a clever two -layer architecture.
• 0:54
  We will use skills and cloud storage together.
• 0:56
  Right. This ensures the AI shares your exact
• 0:58
  memory everywhere. Your business context remains
• 1:01
  perfectly aligned across all your devices. OK,
• 1:04
  let's unpack this setup carefully. Let's do it.
• 1:06
  Claude currently lives in three distinct computing
• 1:09
  environments. Yeah, that is a huge shift. First,
• 1:12
  you have the completely revamped desktop app.
• 1:15
  Desktop now gives you a dedicated code tab. You
• 1:18
  also get a home tab for standard chats. Right.
• 1:21
  You can filter your reasons by chat or task.
• 1:24
  Finding your previous work is significantly easier
• 1:27
  now. Then you have your standard web browser
• 1:29
  access. Yeah. Finally, you have the mobile application
• 1:32
  in your pocket. Right. And those three environments
• 1:35
  are only halfway connected today. I mean, the
• 1:37
  biggest recent shift involves scheduled cloud
• 1:40
  tasks. Oh, absolutely. This fundamentally changes
• 1:44
  how we interact with artificial intelligence.
• 1:47
  Your laptop can be completely closed and powered
• 1:49
  down. Yeah. A scheduled task will still execute
• 1:52
  perfectly in the cloud. These tasks appear with
• 1:55
  a small cloud icon next to them. You just open
• 1:58
  the run to review the final output. But there
• 2:01
  are a few crucial operational details here. Right,
• 2:03
  there are strict rules. Scheduled tasks absolutely
• 2:06
  must use the auto -approval setting. Yeah, if
• 2:09
  they require manual permission, the AI just waits.
• 2:11
  It waits indefinitely. Exactly. Local tasks operate
• 2:16
  differently though. Right. They only run when
• 2:18
  your physical computer is awake. You also need
• 2:20
  to post those outputs as chat links. Mm -hmm.
• 2:24
  Posting links makes them visible inside your
• 2:26
  mobile artifacts. And an artifact basically just
• 2:28
  gives you a clean reading window. Right, and
• 2:30
  that brings up the primary bottleneck. Yeah,
• 2:33
  it does. The mobile app cannot read your local
• 2:36
  computer files directly. Not at all. Your desktop
• 2:38
  machine absolutely must be awake. So it is like
• 2:41
  having two entirely different assistants. Exactly.
• 2:44
  Your office assistant knows all your highly specific
• 2:47
  project files. Mm -hmm. Your travel assistant
• 2:50
  knows your general personality and tone. Right.
• 2:53
  But they are strictly forbidden from talking
• 2:55
  to each other. They operate in completely isolated
• 2:58
  knowledge silos. Two -sex silence. Why does a
• 3:02
  local Mac folder create a fundamentally weak
• 3:05
  setup? Because Mobile Cloud only understands
• 3:08
  your general account settings. It completely
• 3:10
  lacks the deep context of your local files. If
• 3:14
  your entire system lives locally, your phone
• 3:16
  flies blind. It cannot access the reasoning behind
• 3:20
  your previous working sessions. Local -only folders
• 3:23
  basically trap your AI's brain at home. Precisely.
• 3:26
  That is exactly why we need a totally new architecture.
• 3:29
  We need a system that reaches everywhere reliably.
• 3:32
  Yeah, we really do. The smartest approach involves
• 3:35
  splitting the AI's brain apart. We're gonna build
• 3:38
  a specific to layer fix. Let's look at how we
• 3:40
  divide this information Okay, so layer one involves
• 3:43
  treating your stable identity as skills. Mm -hmm
• 3:46
  You take those foundational files that rarely
• 3:48
  change day to day like your personal About Me
• 3:51
  document. That is the perfect example here. It
• 3:53
  gives Claude a baseline understanding of who
• 3:55
  you are. Right. Your voice profile ensures Claude
• 3:58
  always writes in your style. Your core writing
• 4:00
  rules keep the generated outputs perfectly formatted.
• 4:03
  And brand values keep your underlying business
• 4:06
  context front and center. So how do you actually
• 4:08
  build one? You just prompt Claude to turn a file
• 4:11
  into a skill. You download that newly generated
• 4:15
  Markdown file to your machine. Right. Then you
• 4:17
  upload it into your Claude customization settings.
• 4:20
  These skills attach directly to your account,
• 4:23
  not your computer. Exactly. They travel with
• 4:25
  you seamlessly across desktop, web, and mobile.
• 4:28
  I still wrestle with prompt drift myself. Oh,
• 4:30
  we all do. You know when the AI slowly forgets
• 4:33
  your original instructions? Yes. It is incredibly
• 4:36
  frustrating. Establishing a rock -solid, about
• 4:38
  -me skill fixes this drift entirely. It really
• 4:42
  does. It forces Claude to remember your core
• 4:44
  identity constantly. Every new chat starts with
• 4:47
  a perfectly anchored baseline. Right. Now Layer
• 4:50
  2 handles all of your fast -changing system files.
• 4:52
  We are talking about active project memory and
• 4:55
  working briefs. Yeah, your daily context maps
• 4:58
  and work logs change entirely too fast. They
• 5:01
  simply cannot live as skills due to update friction.
• 5:04
  They need a dedicated folder that syncs across
• 5:06
  your devices. The source material evaluated a
• 5:09
  few different cloud storage options. Yeah, let's
• 5:11
  break those down. GitHub is incredible for serious
• 5:14
  version control and team coding. But it requires
• 5:17
  manual pushes for every single daily session.
• 5:20
  That introduces far too much friction. for a
• 5:23
  solo user. Absolutely. Obsidian Sync is popular,
• 5:26
  but adds an unnecessary paid layer. Sync theme
• 5:28
  mirrors folders beautifully across desktops.
• 5:31
  But it fails on mobile entirely. Right. That
• 5:34
  is why Dropbox emerges as the clear winner here.
• 5:37
  It provides seamless, invisible folders syncing
• 5:39
  across all your devices. Crucially, it offers
• 5:41
  a make available offline feature for Mac. This
• 5:45
  ensures your local desktop always has the physical
• 5:47
  files. It also provides excellent access control
• 5:50
  and two -factor authentication. Most importantly,
• 5:53
  Claude has a native Dropbox connector built right
• 5:55
  in. The AI reads the synced folder locally while
• 5:58
  on desktop. And it seamlessly switches to the
• 6:01
  mobile connector when local fails. Active memory
• 6:08
  files change daily and quickly exceed strict
• 6:11
  token limits. Tokens are basically the basic
• 6:14
  units of memory in AI uses. You would spend all
• 6:17
  your time manually rebuilding that skill. Constant
• 6:20
  updates make the skill system totally unmanageable
• 6:23
  for daily work. Skills require static stability,
• 6:27
  while active memory demands dynamic flexibility.
• 6:30
  Exactly. Separating those two functions makes
• 6:32
  the entire ecosystem thrive. Let's walk through
• 6:35
  the actual Dropbox migration process together.
• 6:37
  Yeah, you want to move your existing setup over
• 6:40
  very carefully. First, You open Dropbox on your
• 6:43
  computer. You create a brand new Co -Work OS
• 6:46
  folder inside Dropbox. Right. Then you right
• 6:49
  -click that folder. And you choose Make Available
• 6:51
  Offline. This is critical, so your local machine
• 6:53
  always retains access. Then you carefully copy
• 6:56
  your existing files into that new folder. Next,
• 6:58
  you open up Claude and navigate to your customization
• 7:01
  menu. You click on Connectors and add the native
• 7:04
  Dropbox integration. I highly recommend only
• 7:07
  moving your core Co -Work system here. Yeah,
• 7:09
  you should keep your active coding project strictly
• 7:10
  local for now. GitHub genuinely handles those
• 7:13
  complex coding environments much better anyway.
• 7:15
  For sure. Once your files are copied, you use
• 7:18
  Claude to clean house. Claude needs to scan and
• 7:21
  update all your old file paths. It has to point
• 7:24
  everything toward the new Dropbox destination.
• 7:26
  You just provide a prompt comparing the old and
• 7:28
  new locations. Claude will meticulously find
• 7:31
  every reference to the old local path. It safely
• 7:34
  updates those references to the new Dropbox structure.
• 7:37
  You should definitely use a very capable AI model
• 7:40
  for this task. Right. The AI has to inspect numerous
• 7:44
  files and compare directory trees. You also need
• 7:47
  to watch out for red access warnings. Claude
• 7:49
  will frequently ask for permission to read or
• 7:52
  edit files. You must approve those folder updates
• 7:54
  step -by -step as they happen. Now we need to
• 7:57
  configure the underlying routing instructions.
• 7:59
  This is the really fun part. We have to edit
• 8:01
  a global file inside a hidden folder. Right.
• 8:04
  You press Command -Shift and Dot on your Mac
• 8:07
  keyboard. That combination reveals the hidden
• 8:09
  system folders on your local drive. You open
• 8:11
  the DotClaude folder to find a specific configuration
• 8:14
  file. It is called ClaudeEMD and it acts as the
• 8:18
  master brain. You open it and add a rule to read
• 8:21
  Dropbox first. If local access fails, it defaults
• 8:24
  to the Dropbox mobile connector. It gives the
• 8:26
  AI one clear routing map across every device.
• 8:30
  Hold on. I have to push back on this friction.
• 8:33
  OK, what is bothering you? Revealing hidden folders
• 8:36
  and editing global system files feels incredibly
• 8:40
  tedious. Yeah, I hear that. We're supposedly
• 8:42
  building an intuitive seamless operating system
• 8:45
  here. Right. Does setting this up really need
• 8:48
  to be this deeply technical? I completely understand
• 8:50
  the frustration with hidden system folders. It
• 8:52
  definitely feels a bit like retro computer hacking.
• 8:55
  It really does. But that configuration file is
• 8:58
  the absolute bedrock of this architecture. It
• 9:01
  prevents the desktop and mobile apps from getting
• 9:03
  totally confused. It explicitly tells the AI
• 9:06
  where its real memories actually live. You also
• 9:09
  have to create a very specific sync rule here.
• 9:12
  Yeah, this rule is vital. You tell Claude to
• 9:15
  load identity files before every new session.
• 9:17
  If those local files are unreachable, it safely
• 9:20
  falls back to skills. You only load the heavy
• 9:23
  voice profile when actually generating text.
• 9:26
  This brilliantly saves your context window and
• 9:28
  your processing budget. That sync rule is what
• 9:31
  prevents your digital identity from drifting.
• 9:34
  If you update a local file that also exists as
• 9:37
  a skill, the configuration file forces Claude
• 9:39
  to flag a quick reminder. Right. It reminds you
• 9:42
  to rebuild that matching skill in your settings.
• 9:45
  This ensures your Dropbox files and cloud skills
• 9:48
  never drift apart. They will not slowly evolve
• 9:51
  into two completely different versions of you.
• 9:54
  Two -sec silence. What is the exact danger if
• 9:57
  a user skips this sync rule? Your local files
• 9:59
  and cloud skills will eventually become totally
• 10:02
  mismatched. The AI will start using outdated
• 10:04
  versions of your core identity. You immediately
• 10:07
  lose that unified intelligence across your multiple
• 10:09
  devices. Skipping it fractures your AI into completely
• 10:12
  conflicting identities. Precisely. The master
• 10:14
  configuration file keeps everything perfectly
• 10:17
  harmonized globally. We will be right back after
• 10:20
  this short break. Sponsor. Okay, let's look at
• 10:23
  setting up a second Mac. Once your co -work system
• 10:26
  lives in Dropbox, this becomes brilliantly simple.
• 10:29
  You just install the Claw desktop app locally
• 10:31
  on the new machine. You sign into your specific
• 10:34
  account to sync your basic settings. You verify
• 10:38
  that your scheduled cloud tasks appear in the
• 10:40
  interface automatically. Then you install the
• 10:42
  Dropbox application. Right. and you selectively
• 10:45
  sync your folders. You only need to sync that
• 10:47
  one specific Cowork OS folder. You right -click
• 10:50
  it to make it available offline on this machine,
• 10:52
  too. Finally, you point your Cowork system toward
• 10:54
  that newly synced folder. The second device now
• 10:57
  reads the exact same foundational system. But
• 10:59
  there is a massive warning light flashing for
• 11:01
  multi -Mac users. Oh yeah, pay attention to this
• 11:04
  part. Your Mac... Top -level user folder names
• 11:07
  absolutely must match perfectly. A file path
• 11:10
  pointing to user slash Neil breaks everything
• 11:13
  instantly. Right. It will not resolve on a laptop
• 11:16
  where the user is Simon. If those hard -coded
• 11:18
  file paths break, the AI cannot find your memory.
• 11:22
  If you use multiple Macs, keep those core user
• 11:24
  names identical. It makes path references significantly
• 11:27
  cleaner for the entire routing system. We should
• 11:30
  also discuss linking specific co -work projects
• 11:33
  to your local folders. Co -work projects are
• 11:36
  incredibly powerful, but they have distinct syncing
• 11:38
  limitations. If you link a project directly to
• 11:42
  a local desktop folder, that project stays permanently
• 11:45
  locked to that specific physical computer. Yeah,
• 11:48
  it will not sync cleanly across your other hardware
• 11:51
  devices. For local -only development work, linking
• 11:54
  a local folder is totally fine. But for cross
• 11:56
  -device workflows, do not link the folder directly.
• 11:59
  Create the project in Claude without any direct
• 12:02
  local folder link. Instead, tell the project
• 12:04
  instructions to load the Dropbox folder manually.
• 12:07
  You keep the project instructions extremely simple
• 12:09
  and highly targeted. Right, and let the master
• 12:12
  configuration file do the heavy lifting. The
• 12:14
  system always reads the root folder directly
• 12:16
  from Dropbox first. It seamlessly fails over
• 12:19
  to the mobile connector when you leave your desk.
• 12:21
  You can use projects primarily for organizing
• 12:24
  and grouping your various chats. You star the
• 12:26
  top -level folder and ask Claude to navigate
• 12:28
  there. Each subfolder can maintain its own specific
• 12:32
  active memory files perfectly. You absolutely
• 12:34
  must test this setup on your mobile device afterward.
• 12:37
  The mobile test proves whether this entire architecture
• 12:39
  actually works correctly. Yeah. The mobile app
• 12:42
  obviously does not have local Mac folders mounted.
• 12:44
  Start a brand new co -work task directly on your
• 12:48
  mobile phone. Ask it something that requires
• 12:50
  deep context, not just generic knowledge. Right,
• 12:54
  ask it to review your current daily productivity
• 12:56
  system in detail. Ask it to check your specific
• 13:00
  strategic priorities for today. Claude should
• 13:03
  automatically read those route instructions and
• 13:05
  load your identity. It seamlessly finds the Dropbox
• 13:08
  folder through the native mobile connector. It
• 13:11
  uses your actual personalized knowledge base
• 13:14
  to formulate the answer. It might run just a
• 13:17
  little bit slower on a mobile network. Yeah,
• 13:19
  that is expected. Whoa, imagine scaling to a
• 13:22
  billion queries. It is wild to think about. Millions
• 13:25
  of people instantly syncing their entire digital
• 13:28
  brains. Right. It is a profound shift in global
• 13:31
  data architecture. The mobile application processing
• 13:34
  this A bit slower. Makes total sense. But that
• 13:36
  slight delay is completely acceptable for this
• 13:38
  specific workflow. The ultimate goal on mobile
• 13:42
  is not pure blistering processing speed. Right.
• 13:45
  The goal is retrieving the absolute perfect memory
• 13:48
  context every time. That is the true magic of
• 13:52
  this unified architecture. Your highly personalized
• 13:55
  system finally follows you away from your desk.
• 13:58
  Two secs silence. Why must we accept slower processing
• 14:01
  speed on mobile? Perfect context requires securely
• 14:05
  fetching remote cloud data through third -party
• 14:07
  connectors. Achieving perfect accuracy simply
• 14:10
  justifies that extra moment of network latency.
• 14:13
  Perfect context easily justifies a slight mobile
• 14:16
  processing delay. Exactly. Fast and wrong is
• 14:19
  completely useless for a personalized assistant.
• 14:21
  We need to honestly discuss the caveats and team
• 14:24
  limits here. There are definitely constraints
• 14:26
  to this unified architecture right now. The biggest
• 14:28
  constraint involves directly editing files on
• 14:31
  your mobile device. Yeah. Through the Dropbox
• 14:33
  connector, Claude can read and create files perfectly.
• 14:36
  But it seriously struggles to make clean, block
• 14:39
  -by -block file edits. It cannot surgically edit
• 14:43
  a few lines within a massive document yet. It
• 14:45
  usually has to completely replace the entire
• 14:47
  file from scratch. Sometimes it just creates
• 14:49
  a brand new version of the file instead. That
• 14:51
  matters immensely for your highly dynamic active
• 14:54
  memory logs. If you try updating memory files
• 14:56
  from your phone, be incredibly careful. Claude
• 15:00
  might accidentally corrupt the formatting during
• 15:02
  that massive overwrite process. A shared Dropbox
• 15:05
  folder works beautifully for solo founders and
• 15:07
  freelancers. Right. But for larger teams, two
• 15:11
  people editing simultaneously causes major version
• 15:14
  conflicts. Dropbox simply isn't designed to merge
• 15:17
  simultaneous text edits cleanly together. GitHub
• 15:20
  remains the vastly superior fit for technical
• 15:23
  teams collaborating together. Its entire purpose
• 15:25
  is handling simultaneous team changes safely
• 15:28
  and predictably. Exactly. Bede, I have to push
• 15:31
  back on this architecture just slightly. Sure.
• 15:33
  Is this truly a reliable operating system if
• 15:36
  it breaks so easily? That is a fair question.
• 15:38
  You risk corrupting your core memory files just
• 15:41
  riding on the train. Yeah. Updating a simple
• 15:43
  daily task list could theoretically break your
• 15:46
  system. It is a totally valid concern for anyone
• 15:48
  adopting this early. It is a functioning operating
• 15:51
  system, but one with very strict boundaries.
• 15:54
  Right. You should use mobile primarily for reading,
• 15:57
  planning, and capturing raw ideas. You simply
• 16:00
  avoid complex structural edits while you are
• 16:03
  away from your desk. Dropbox remains the easiest,
• 16:06
  most accessible starting point for solo users.
• 16:09
  It provides simple syncing and one beautifully
• 16:11
  unified source of truth. Two secs silence. When
• 16:15
  is the definitive time to avoid mobile clod?
• 16:18
  Never use mobile clod to edit your complex structural
• 16:21
  memory files directly. Stick exclusively to queries,
• 16:25
  daily reviews, and generating entirely new text
• 16:27
  outputs. Avoid mobile for editing complex system
• 16:30
  memory files entirely. Precisely. Read and review
• 16:33
  on your phone, build and edit at your desk. Let's
• 16:36
  slowly synthesize the core philosophy we uncovered
• 16:39
  today. It is a really beautiful concept once
• 16:42
  it clicks. The real magic here is not the mobile
• 16:44
  app itself. No, not at all. The magic. is this
• 16:47
  resilient two -layer architecture we build together.
• 16:50
  Your stable, enduring identity lives permanently
• 16:54
  in the cloud as skills. Your rapidly changing
• 16:57
  reality lives dynamically inside your synced
• 17:00
  Dropbox folders. And they are bound tightly together
• 17:03
  by that master configuration file. It creates
• 17:06
  one single unbroken intelligence across every
• 17:10
  single device. Every piece of glass you own now
• 17:12
  shares one unified brain. You can start a complex
• 17:15
  strategic task on your phone outside. You seamlessly
• 17:18
  continue that exact same task on your laptop
• 17:20
  later. It is a genuinely fascinating evolution
• 17:23
  in personal computing history. It really is.
• 17:25
  It changes how we think about digital environments
• 17:28
  entirely. We're rapidly moving away from AI as
• 17:31
  a specific destination. Right. It is no longer
• 17:33
  just a smart desk you occasionally visit. It
• 17:35
  is becoming an always -on unified digital shadow
• 17:38
  that follows you. Yeah. If your AI shares your
• 17:41
  exact context, memory, and rules everywhere,
• 17:44
  no matter where you happen to be in the physical
• 17:46
  world, what does that actually mean for the future
• 17:48
  of our own minds? That is a massive question.
• 17:52
  Will we eventually stop needing to form our own
• 17:55
  human memories entirely?