Latest / AI Fire Daily / #595 Neil: Claude Cowork Setup Makes Daily Business Work Less Manual

Transcript

#595 Neil: Claude Cowork Setup Makes Daily Business Work Less Manual · AI Fire Daily

• 0:00
  Most people use AI like a smart typewriter. They
• 0:03
  read files, they write emails, they do isolated
• 0:05
  tasks. But that only scratches the surface. What
• 0:09
  happens when the AI actually remembers how your
• 0:12
  business works? Oh, it changes everything. Welcome
• 0:15
  to the Deep Dive. Hey there. Our mission today
• 0:18
  is exploring a real master class. We are looking
• 0:22
  at transforming Claude Cowork. We want to move
• 0:25
  it from a basic assistant to an automated system.
• 0:28
  We want a reliable engine for everyday business.
• 0:31
  It is a complete architectural shift. It really
• 0:34
  is. We are going to trace that exact framework
• 0:36
  today. We start by building a really clean data
• 0:40
  foundation, then we create the core context files.
• 0:42
  Right. After that, we examine how to route those
• 0:45
  files efficiently. Finally, we set up automated
• 0:48
  workflows in the cloud. I am really looking forward
• 0:51
  to unpacking this. because I still wrestle with
• 0:54
  prompt drift myself. Oh, it happens to absolutely
• 0:56
  everyone. It is incredibly frustrating. For anyone
• 0:59
  unfamiliar, that is when the AI slowly forgets
• 1:02
  your original instructions over time. You start
• 1:04
  a chat. and it works perfectly. You establish
• 1:07
  the tone, you set the rules, then 50 messages
• 1:10
  later, it gets totally confused. It completely
• 1:12
  loses the thread of the conversation. It hallucinates
• 1:15
  policies you never set. Exactly. It just reverts
• 1:18
  to that generic robotic voice. It defeats the
• 1:21
  entire purpose of automation. You cannot build
• 1:24
  a business on a system that randomly forgets
• 1:26
  your company name. Definitely not. So to stop
• 1:29
  that drift, We must look at where the AI actually
• 1:32
  lives. We have to fix the workspace itself Yeah,
• 1:35
  people tend to treat AI chats like temporary
• 1:38
  scratch pads, right? They dump information in
• 1:40
  they get an answer and they close the window
• 1:42
  Which means next time you have to explain it
• 1:44
  all again exactly the framework insists. We stop
• 1:47
  explaining things repeatedly We need one main
• 1:51
  permanent folder per business. Or if you run
• 1:55
  an agency, one main folder per client. That makes
• 1:58
  perfect sense logically. If you mix clients in
• 2:00
  one space, the context window gets contaminated.
• 2:02
  Highly contaminated. The AI tries to draw connections
• 2:04
  where none exist. It pulls a legal disclaimer
• 2:08
  from client A, then it tries to apply it to a
• 2:11
  casual tweet for client B. Oh, wow. That's bad.
• 2:14
  It gets very confused. So the structure of this
• 2:17
  workspace needs to be remarkably simple. You
• 2:19
  organize it by fundamental business pillars.
• 2:21
  You do. You set up folders for brand and customers.
• 2:25
  Right. You add products, projects, assets, and
• 2:28
  processes. That is it. Keep the top level clean.
• 2:31
  Keep it super clean. The source gives a very
• 2:34
  rigid rule of thumb here. Have you explained
• 2:37
  a concept to Claude twice? Mm -hmm. If yes, it
• 2:40
  needs to become a saved file. That is a powerful
• 2:42
  rule. It forces you to treat the AI like a new
• 2:45
  employee. Exactly. It is kind of like stacking
• 2:48
  Lego blocks of data. You would not verbally explain
• 2:51
  the company mission every morning. No. You would
• 2:54
  hand them an onboarding document. You build a
• 2:56
  permanent library of truth. It saves so much
• 2:58
  manual effort down the line. You stop rebuilding
• 3:01
  context from scratch every single time you open
• 3:04
  a chat. I see the value there. But let me push
• 3:06
  back a little bit. I am thinking about my own
• 3:09
  hard drive right now. It is a mess, I assume.
• 3:12
  It is a complete disaster. What if I have a really
• 3:15
  messy existing folder structure? Years of overlapping
• 3:19
  client files. Do I have to manually organize
• 3:21
  everything before I can even start this process?
• 3:24
  That sounds like a nightmare. It does sound awful,
• 3:26
  but you don't have to do that. You can actually
• 3:28
  prompt Claude to design a clean structure first.
• 3:31
  How did that work? You feed it your messy directory
• 3:33
  list. It analyzes the chaos. OK. Then it maps
• 3:37
  out a clean, logical folder hierarchy without
• 3:40
  deleting anything. So ask the AI to map the mess
• 3:43
  before moving any files. Precisely. It is entirely
• 3:47
  risk free. You just review the proposed structure.
• 3:50
  You check if the categories make sense. Then
• 3:53
  you approve it before making it any actual structural
• 3:55
  changes. OK, so we have organized the workspace.
• 3:57
  The foundation is clean. Now we need to furnish
• 3:59
  it. We have to give the AI permanent memory.
• 4:02
  We need to give it files it can reliably read.
• 4:04
  We do this by creating specific .n -d files.
• 4:07
  These are Markdown files. Yes, Markdown files.
• 4:10
  Let us pause on that detail for a second. Why
• 4:12
  Markdown? Why not just upload massive PDF files
• 4:15
  or Word documents? Well, that comes down to how
• 4:18
  language models actually parse data. Word documents
• 4:22
  are full of invisible formatting code. What do
• 4:24
  you mean by invisible code? Underneath the text
• 4:27
  you see, there are hidden tags. Tags for fonts
• 4:30
  and margins. Oh, I see. The AI has to read all
• 4:33
  of that hidden junk. It wastes valuable context
• 4:35
  tokens on formatting data it doesn't need. It
• 4:38
  eats up the context window rapidly. It does.
• 4:40
  Markdown strips all that away. It is incredibly
• 4:43
  lightweight, plain text. So it's just pure data.
• 4:46
  Exactly. It uses simple hashes and asterisks
• 4:48
  for structure. It is the native language of these
• 4:51
  models. OK. So we build our core files in Markdown.
• 4:54
  The most critical one... is the brandvoice .mday
• 4:57
  file. This captures the subtle elements, the
• 5:00
  tone, the phrasing, and the formatting rules.
• 5:02
  It also captures the phrases you actively want
• 5:04
  to avoid. If your brand never uses the word synergy,
• 5:08
  you put that in the file, it gives the AI a deeply
• 5:10
  stable personality. That is brilliant. Next,
• 5:13
  you build the visualidentity .mday file. I imagine
• 5:16
  that is vital for design -heavy tasks. It holds
• 5:19
  your stable visual rules. It defines your exact
• 5:22
  hex codes for colors. It lists your primary and
• 5:25
  secondary fonts. It details the logo placement
• 5:28
  rules. It even covers spacing and a general photography
• 5:31
  direction. Then we move to the hard business
• 5:33
  facts. Right. You create a company overview .m
• 5:36
  -defile. You add a products .m -file. Finally,
• 5:41
  a CustomerProfile .m -d file. This gives the
• 5:44
  AI the literal facts of your operation. There
• 5:47
  is a crucial distinction we need to highlight
• 5:49
  here, though. We have to keep temporary things
• 5:51
  out of these files. It is a very common mistake.
• 5:54
  People put their current summer sale details
• 5:56
  into the main product file. Seasonal campaigns
• 5:59
  and current inventory fluctuate. They do not
• 6:01
  belong in permanent memory files. Temporary data
• 6:04
  becomes outdated very quickly. If you leave the
• 6:06
  summer sale in there, the AI will reference it
• 6:09
  in December. It would be a disaster. It degrades
• 6:10
  the reliability of the system. Keep permanent
• 6:13
  context in permanent files. Put temporary data
• 6:15
  in temporary project folders. What happens when
• 6:18
  the AI is drafting something, but a specific
• 6:20
  piece of data is missing? The source outlines
• 6:23
  a brilliant safeguard for this. You build a needs
• 6:26
  input section into your templates. How does that
• 6:29
  work in practice? If the AI cannot find the price
• 6:32
  of a new service, it does not guess. It leaves
• 6:35
  a placeholder in the draft. Okay. Then it flags
• 6:38
  the missing variable in that needs input section.
• 6:40
  It stops the AI from hallucinating details just
• 6:43
  to please you. That is very clever. It really
• 6:45
  is. But I do have a question about that brand
• 6:47
  voice file. How does the AI handle different
• 6:50
  voices across different mediums? What do you
• 6:52
  mean? Well, a formal investor update sounds very
• 6:56
  different than a snappy Instagram caption. If
• 6:59
  we only have one brand voice file, does everything
• 7:01
  sound the same? Oh, no. Claude records channel
• 7:05
  -specific differences in the brand voice file
• 7:07
  itself. You create subsections within that single
• 7:10
  markdown document. One section defines email
• 7:13
  tone. Another defines social media tone. It cross
• 7:16
  -references the task with the specific channel
• 7:18
  rules. Got it. It maps different tones for different
• 7:21
  platforms in one file. It handles the nuance
• 7:23
  beautifully. It reads the core personality, then
• 7:26
  applies the specific platform filter. So, we
• 7:30
  have built all these great memory files. But
• 7:33
  this creates a new problem. What is that? If
• 7:36
  I ask the AI to write a simple internal slack
• 7:39
  message, I don't want it reading the visual identity
• 7:42
  file. No, you definitely do not. It would completely
• 7:45
  bloat the system. The AI would crash or slow
• 7:48
  to a crawl if it read every file every time.
• 7:50
  It needs a traffic cop. Which brings us to the
• 7:53
  routing layer. We introduce a file literally
• 7:56
  called CLAWUDE .md. Okay. This sits at the very
• 8:00
  top of your workspace. It acts as the master
• 8:02
  router. Think of it less like a librarian and
• 8:05
  more like a triage nurse at a hospital. That
• 8:07
  is a much better analogy. If you walk in with
• 8:09
  a sprained ankle, the triage nurse does not page
• 8:11
  the top neurosurgeon. They route you to the x
• 8:14
  -ray tech. Right. The CLAWUDE .md file looks
• 8:17
  at your prompt in the same way. It says this
• 8:19
  is an email task. Give you the brand voice file.
• 8:22
  Exactly. Keep the visual design file in the filing
• 8:24
  cabinet. It explicitly tells the AI which context
• 8:28
  files to reference for specific categories of
• 8:31
  work. It separates content tasks from design
• 8:34
  tasks. It isolates business research. It completely
• 8:38
  prevents data duplication in the context window.
• 8:41
  It makes everything faster. It keeps the output
• 8:43
  highly accurate. It is the critical piece. I
• 8:45
  am curious about overlapping data though. What
• 8:48
  happens if two context files actually have conflicting
• 8:51
  information? That happens a lot. Right. Say the
• 8:54
  product file lists the price as $50, but a recent
• 8:57
  project brief lists it as $60. How does the triage
• 9:01
  nurse handle that? That is where the router really
• 9:04
  proves its worth. The cleavd .md file establishes
• 9:07
  strict priority rules for conflicts. Oh. You
• 9:10
  program it to always trust the project brief
• 9:12
  over the general product file if dates overlap.
• 9:15
  The router acts as the ultimate tiebreaker for
• 9:17
  conflicting data. It prevents the AI from freezing
• 9:19
  up or combining the numbers. It gives it a clear
• 9:22
  hierarchy of truth. We have the permanent memory.
• 9:24
  We have the routing mechanism. Now we have to
• 9:26
  make it actually execute the work. We turn recurring
• 9:29
  business processes into clawed skills. Right.
• 9:32
  You take a manual workflow you already do every
• 9:34
  week, then you map it out structurally. You cannot
• 9:37
  just tell the AI to do my job. You have to outline
• 9:40
  the goal. You define the exact trigger. You list
• 9:43
  the step -by -step logic. You specify which apps
• 9:46
  it is allowed to touch. Let us walk through a
• 9:48
  very specific example from the source. The weekly
• 9:52
  Friday supplier risk check. It is a tedious manual
• 9:56
  task. It is a perfect use case for a Claude scale.
• 9:59
  It really is. I am trying to picture the reality
• 10:02
  of this. Real business data is incredibly messy.
• 10:06
  It is usually a complete disaster. Right. You
• 10:09
  have a Google Sheet for purchase orders. It probably
• 10:11
  has 20 columns. And half of them are missing
• 10:13
  data. Exactly. Half of them are formatted inconsistently.
• 10:16
  Someone typed a date in European format. Someone
• 10:19
  else just wrote next Tuesday. Then you factor
• 10:21
  in the email threads. Suppliers do not write
• 10:24
  neat, structured updates. No, they send a three
• 10:26
  -word reply from a warehouse floor. Shipment
• 10:28
  delayed, sorry. That is all you get. So how does
• 10:30
  this skill actually parse that chaos? The trigger
• 10:33
  hits on Friday morning. The skill directs Claude
• 10:36
  to read the shared Google Sheet. It uses its
• 10:39
  natural language processing to understand the
• 10:41
  messy date columns. It deciphers the next Tuesday
• 10:44
  comment. Yes. Then it cross -references those
• 10:47
  PO numbers with the messy Gmail threads. It reads
• 10:50
  the unstructured thunder email. It identifies
• 10:52
  the PO number. It spots the delay. Right. And
• 10:56
  then it compares that delayed date against your
• 10:58
  upcoming product launches. It does all of this
• 11:00
  analytical heavy lifting in seconds. It flags
• 11:03
  any delays actively threatening a launch. Then
• 11:06
  it drafts a short risk summary. It lists the
• 11:09
  core issue. Yeah. It notes the expected financial
• 11:12
  impact. It recommends a specific action to take.
• 11:15
  And it does not just fire off angry emails to
• 11:18
  your suppliers. There is a crucial safeguard
• 11:20
  built into the skill architecture. Great. It
• 11:22
  mandates a hard stop for human approval. It drafts
• 11:25
  the warning email, but it waits for you to click
• 11:27
  send. It leaves the draft in your outbox. You
• 11:30
  review it, tweak it if necessary, and approve
• 11:32
  it. Whoa. Imagine scaling to a billion queries.
• 11:36
  or even just having it perfectly manage a massive
• 11:39
  global supply chain. The efficiency gains are
• 11:42
  staggering. You do the hard structural thinking
• 11:45
  once. You map the messy reality of the workflow,
• 11:49
  then the AI manages that chaos forever. But let
• 11:52
  me ask about technical reliability. What happens
• 11:54
  if the Google Sheet API is down? Yeah, that happens.
• 11:58
  Or what if it simply cannot access the Gmail
• 12:01
  thread? during the Friday check. The framework
• 12:03
  is very strict on this point. You explicitly
• 12:07
  instruct the skill logic to handle blind spots.
• 12:10
  If a data source is offline, it must flag the
• 12:13
  unavailable source clearly. It cannot try to
• 12:16
  guess the vendor's status based on last week's
• 12:18
  data. So it admits it can't see the data instead
• 12:20
  of hallucinating answers. It prioritizes absolute
• 12:23
  accuracy over just giving you a completed summary.
• 12:25
  It tells you exactly where the chain broke. Sponsor.
• 12:28
  Okay, so building a skill is incredible powerful,
• 12:31
  but you still have to manually click start every
• 12:33
  Friday morning. Right, which is still manual.
• 12:35
  How do we take this to the final level? How do
• 12:37
  we make it truly autonomous? We move into scheduled
• 12:41
  tasks, we take that supplier risk check skill,
• 12:43
  and we move it to the cloud. Let us define what
• 12:46
  in the cloud actually means here. We're not talking
• 12:49
  about literal magic. We're talking about API
• 12:52
  bridges. Tools like Zapier or Make .com, they
• 12:56
  act as the connectors. They sit on a server waiting
• 12:59
  for a specific time. Your permanent context files
• 13:02
  just need to live somewhere accessible to those
• 13:04
  APIs. They need to be in a shared Google Drive
• 13:07
  or a Notion database. If the workspace is online,
• 13:11
  the automation tool can read it. Right. It triggers
• 13:14
  the skill at 9 a .m. every Friday. It pulls the
• 13:17
  data, runs the analysis, and drafts the email.
• 13:20
  It does all of this while your laptop is totally
• 13:23
  closed. You could be on a flight. You could be
• 13:25
  asleep. The system runs on external servers.
• 13:27
  You wake up and the analytical work is already
• 13:30
  finished. The system only pings you when it hits
• 13:32
  that required human approval checkpoint. You
• 13:34
  just get a Slack notification. Yes. You check
• 13:37
  the drafted email on your phone, you hit approve,
• 13:40
  and the crisis is averted before you even get
• 13:42
  to the office. I do have one final technical
• 13:45
  question, though. If I move this process to Zapier,
• 13:49
  do I have to rewrite all the logic? Do we have
• 13:52
  to rebuild the skill for the scheduled task to
• 13:54
  understand it? Not at all. The schedule task
• 13:57
  is just an API call. It simply points to the
• 14:00
  exact Claude skill and the existing context files
• 14:03
  you already perfected. It just presses play on
• 14:06
  the system you already built. It leverages the
• 14:08
  exact same triage nurse and memory files. It
• 14:10
  just operates on a timer. Okay, let's pull this
• 14:13
  massive architectural shift together. Building
• 14:16
  a highly reliable AI setup doesn't have to be
• 14:19
  overwhelmingly complicated. It requires discipline,
• 14:21
  not necessarily complex coding. It relies on
• 14:24
  a very simple progression of truths. You organize
• 14:27
  the workspace. You save permanent context once.
• 14:30
  You use a router to manage the token limits.
• 14:33
  You build a reproducible skill. Finally, you
• 14:35
  attach it to a cloud schedule. You stop treating
• 14:37
  the AI like a disposable typewriter. You start
• 14:39
  treating it like a structural component of your
• 14:41
  business. It becomes a system that actually remembers
• 14:44
  how your company operates. It reduces repetitive
• 14:47
  manual work in a way that feels almost alarming.
• 14:50
  It shifts your daily focus entirely. It really
• 14:53
  does. And that leaves us with a final thought
• 14:55
  to mull over. If the AI can context switch seamlessly,
• 14:59
  if it can route its own data to manage token
• 15:01
  limits, and if it can parse messy business spreadsheets
• 15:05
  while our laptops are completely closed, what
• 15:07
  does our daily work actually look like when our
• 15:09
  main job is simply acting as the final human
• 15:12
  approval note?