Latest / AI Fire Daily / #07 Alex: [Claude AI Mastery Playbook] Lesson 5: 3 Practical Claude Cowork Use Cases & A Bonus Skill
Transcript
- 0:00Imagine an AI that doesn't just talk, but actually,
- 0:03you know, works while you sleep. Oh, yeah. That
- 0:05is the dream, right? Think about that for a second.
- 0:07It's the fundamental difference between having
- 0:10a simple chat with an AI and building a functioning
- 0:13independent system. If you want a digital employee,
- 0:17you basically have to build it a workspace. Precisely.
- 0:19And that's exactly what we are doing today with
- 0:22Claude Cowork. Right. Welcome back to the Deep
- 0:25Dive. Today, we are exploring building your first
- 0:29Claude Cowork Agent. We're going to cover the
- 0:32setup, the folder logic, and the automation steps.
- 0:34Yeah, you probably use Claude every day. But
- 0:36honestly, you're likely using it wrong, or at
- 0:38least you're only scratching the surface. Totally.
- 0:42Today is a literal roadmap for moving beyond
- 0:45that traditional chat box. We're going to explore
- 0:48how you can use Cowork to build a real, functional,
- 0:52daily AI news dashboard. OK, let's unpack this.
- 0:55Because when I first heard about Cowork, I honestly
- 0:57thought it was just I don't know, another UI
- 0:59update. Right, like just a visual thing. Yep,
- 1:01exactly. Another place to tap a question. But
- 1:04Cowork is totally different. It's built for repeatable
- 1:07file -based jobs. It lives inside your folders,
- 1:11it reads, it writes, and it saves documents rather
- 1:15than just existing in a fleeting conversation
- 1:17window where you have to constantly copy and
- 1:20paste everything. What's fascinating here is
- 1:22the shift in computing behavior. We are all used
- 1:25to standard cloud chat. Which is great, obviously.
- 1:28Oh, it's fantastic for quick brainstorming, drafting
- 1:31a single email, or just answering a one -off
- 1:33question. To transition from getting one -time
- 1:37help to building automated business workflows,
- 1:39you are no longer a conversational partner. Right.
- 1:42You are a systems architect. I love that. A systems
- 1:45architect. So if I'm playing that role, what
- 1:47are my actual building materials? Let's talk
- 1:49prerequisites for you guys listening. First,
- 1:52you need the cloud desktop app installed. Yes.
- 1:54This isn't running in a browser tab. Second,
- 1:56you need a paid plan. Pro is the bare minimum
- 1:59here. And we should probably explain exactly
- 2:01why that free tier simply won't cut it. Yeah,
- 2:03please do. Core workflows inherently require
- 2:07much higher usage limits and significantly more
- 2:11processing context. Right. When you run a dashboard
- 2:14workflow, the AI is reading multiple files. It's
- 2:18cross -referencing your instructions against
- 2:20several external articles, and it's generating
- 2:23heavily formatted outputs. It's doing a lot at
- 2:26once. Exactly. It needs a massive working memory
- 2:29to hold all that data simultaneously. That makes
- 2:31total sense. And that actually brings up what
- 2:33I like to call the work Golden Roll, you need
- 2:36a dedicated workspace folder on your actual computer.
- 2:39Yes, absolutely. You might name it something
- 2:41like Claude Cookework News Dashboard, but I want
- 2:44to make a hard stop right here because the temptation
- 2:46is just to give the AI access to everything.
- 2:49Oh, big mistake. Do not just point Claude at
- 2:52your whole computer or your entire documents
- 2:54folder. That is like letting a brand new intern
- 2:56rifle through all your personal filing cabinets,
- 2:58your tax returns, your family photos, just to
- 3:01find a single press release. Yeah, that is an
- 3:03incredibly dangerous trap for new users. It really
- 3:06is. If we connect this to the bigger picture
- 3:08of how language models operate, giving the AI
- 3:11too much access just creates massive confusion.
- 3:14Right, because it's reading everything. Exactly.
- 3:16Every time it reads an irrelevant file, it eats
- 3:19up its processing context, its working memory.
- 3:22It has to sift through your grocery lists and
- 3:24old spreadsheets just to figure out what a news
- 3:27update means. Wow. That dilutes its focus. You
- 3:30want to isolate the work space so the AI only
- 3:33sees exactly what it needs to execute the task.
- 3:36Nothing more, nothing less. And speaking of focus,
- 3:39you also have to dictate which underlying model
- 3:41is actually doing the processing. The sources
- 3:43are very clear about using the strongest model
- 3:46available for research tasks. So we are talking
- 3:49about using something like Claude 3 .5 Opus or
- 3:52Sonnet rather than the faster, lighter Haiku
- 3:55model, right? Yes. Precision matters immensely
- 3:57here. When you ask an AI to analyze multiple
- 4:00sources, rank them, and synthesize the information,
- 4:03you are fighting against a phenomenon called
- 4:05prompt drift. Wait, prompt drift. Explain prompt
- 4:08drift for me. Sure. Think of it like a game of
- 4:10telephone happening inside the AI's own head.
- 4:12Over the course of a complex multi -step task,
- 4:16a weaker model slowly forgets or misinterprets
- 4:19the original instructions. Oh, I've seen that
- 4:22happen. Right. It might start out strong on the
- 4:24first two articles, perfectly applying your formatting.
- 4:27But by the time it reads the 10th article, its
- 4:30working memory is stretched, and it starts hallucinating,
- 4:33summarizing things in the wrong tone, or just
- 4:36completely ignoring your formatting rules. That's
- 4:39so frustrating. It is. So using the strongest
- 4:41model, provides the immense cognitive stamina
- 4:44required to hold onto your rules from start to
- 4:47finish. Got it. So you've restricted the AI's
- 4:50access to your tax returns, and you've selected
- 4:52a model with enough stamina to avoid drifting
- 4:55off course. Right. Now you actually have to build
- 4:57the physical structure inside that isolated folder.
- 5:00Exactly. Think of the folder structure like an
- 5:02assembly line in a factory. To get this digital
- 5:05employee working smoothly, you need three main
- 5:08folders on that assembly line. Okay, what are
- 5:10they? about, outputs, and templates. I have to
- 5:14admit, creating three separate folders and a
- 5:17bunch of markdown files just to get a new summary
- 5:19feels incredibly tedious. It sounds like a lot
- 5:22of work. It really does. Why can't I just paste
- 5:25a massive three page prompt into the regular
- 5:28clawed chat window every morning and get the
- 5:31exact same result? Well, because copying and
- 5:33pasting a massive prompt every day isn't a system,
- 5:35it's a chore. Fair point. If you put all your
- 5:38rules into a chat prompt, that knowledge vanishes
- 5:40the moment you close the window. Yeah. By building
- 5:43the structure on your hard drive, you are creating
- 5:45permanence. The about folder serves as the permanent
- 5:49blueprint for the assembly line. Yeah. The templates
- 5:52folder is the mold. If you don't build the track
- 5:55correctly and permanently, the robotic arm claw
- 5:58just flails around in the dark every single time
- 6:00you start a new session. OK, that clarifies it.
- 6:03We are trading upfront effort for long term automation.
- 6:06I can get behind that. Exactly. So let's look
- 6:08at that blueprint, the ABOUT folder. Our sources
- 6:11say this is where you store the high signal rules
- 6:13that filter out the noise before the AI even
- 6:16begins to work. Yes. The internet is filled with
- 6:19garbage data. So much garbage. Left to its own
- 6:22devices, and AI will try to summarize all of
- 6:24it. The files in your about folder are your hard
- 6:28filters to prevent that. Let's walk through those
- 6:30filters. First up is your source list dot dash
- 6:33MD file. And in this file, you explicitly limit
- 6:36the AI to five to eight trusted sources. You
- 6:39organize them by priority. Priority 1 is official
- 6:42company blogs like Antropic or OpenAI. Priority
- 6:452 is research repositories like Hugging Face.
- 6:48Priority 3 is general business and tech news.
- 6:51Right. But I'm looking at this limitation and
- 6:53wondering why handcuff the AI? Why not just say,
- 6:56search the whole internet for AI news today?
- 6:58Because a highly curated source list produces
- 7:01dramatically cleaner reports. If you tell an
- 7:03AI to search the entire web, it will pull from
- 7:06SEO, spam sites, unverified Reddit threads, and
- 7:09out -of -date opinion blogs. That's true. By
- 7:12restricting it to, say, five official URLs, you
- 7:14are guaranteeing the data quality. You are also
- 7:17explicitly instructing it, do not treat rumors
- 7:19as facts. OK, so we've restricted the sources.
- 7:22But even within five trusted official sites,
- 7:24there's a ton of garbage. Oh, absolutely. Companies
- 7:27published market. fluff all the time. How do
- 7:30we stop the AI from summarizing a minor interface
- 7:33tweak like it's a major technological breakthrough?
- 7:36That brings us to the second file in your About
- 7:39folder, the DashboardRules .md file. Okay. This
- 7:43is where you define consequence. You instruct
- 7:45the AI to rank every single news item on a strict
- 7:48scale of one to five based on practical impact
- 7:51and urgency. So how does that play out in a real
- 7:54scenario? Let's say OpenAI drops a massive new
- 7:57model upgrade, but on the exact same day, they
- 8:00also release a minor bug fix for their billing
- 8:02page. In a standard chat prompt, the AI might
- 8:05just give you two bullet points of equal length,
- 8:07treating both events as equal news. Right. But
- 8:09with your dashboard rules file, the AI is forced
- 8:12to pause and evaluate. It looks at the massive
- 8:14model upgrade, sees that it alters capabilities,
- 8:17and scores it a 5. OK. It looks at the billing
- 8:20bug fix, realizes it has zero impact on output
- 8:22quality, and scores it a 1. Your rules can then
- 8:25dictate only include items that score a three
- 8:28or higher in the final report. You've just built
- 8:31an automated editorial filter. You're forcing
- 8:33the AI to have editorial judgment based on your
- 8:36criteria. I see. But even if it picks the right
- 8:40stories, there's still the problem of how it
- 8:42actually talks. Yes, the tone. Which leads to
- 8:46the third file in the blueprint, the writing
- 8:48style .md file. Yes, and this might be the most
- 8:52crucial file for long -term readability. You
- 8:55use this strict writing style file to enforce
- 8:58clear, direct, human -sounding English. Right.
- 9:01You demand short sentences. You demand factual
- 9:03reporting. And the absolute most fascinating
- 9:05part of this file is the banned word list. Oh,
- 9:08it's so good. You explicitly ban those cliché,
- 9:11exhausting AI buzzwords. You tell it under no
- 9:13circumstances are you allowed to use the words
- 9:15revolutionary, game -changing, unlock, unleash,
- 9:18seamless, or dive into. This is a brilliant structural
- 9:21hack. AI models naturally default to marketing
- 9:23speak. They really do. Think about their training
- 9:26data. They have ingested millions of press releases
- 9:29and marketing websites. When an AI sees a product
- 9:32update, its highest probability predictive text
- 9:35pathway is to describe it as game changing. So
- 9:38what does this all mean when we ban those words?
- 9:41Were you just making it sound less annoying?
- 9:42It goes much deeper than just tone. Banning those
- 9:45words physically forces the model's neural network
- 9:48to access different, more analytical pathways
- 9:51to complete its sentence. If the AI is literally
- 9:55prohibited from saying, a new tool will unleash
- 9:57your potential, it hits a roadblock. It is forced
- 10:01to look closer at the data and explain the actual
- 10:03mechanism. That makes a lot of sense. It has
- 10:05to say, this tool reduces rendering time by 40
- 10:08percent. It completely shifts the output from
- 10:10passive marketing fluff to active practical analysis.
- 10:13So you're saying we have to artificially constrain
- 10:15the AI to make it smarter. Yes. We take away
- 10:18its favorite adjectives. to rely on strong verbs
- 10:20and hard facts. That is a massive paradigm shift.
- 10:23OK. So we have our assembly line blueprint built.
- 10:26The About folder is locked. But how does the
- 10:28AI actually know where to find the blueprint
- 10:31and what to do with it? That is the job of global
- 10:34instructions. You find these in the Settings
- 10:36menu of Cowork. Think of global instructions
- 10:39as the permanent laws of your workspace. They
- 10:42act as the traffic director. Right, so your global
- 10:45instructions would read something like read the
- 10:47source list, the dashboard rules, and the writing
- 10:49style in the about folder. Use the template in
- 10:52the templates folder and save the final output
- 10:55to the ITPU test folder. Exactly. But here's
- 10:58my question. If we went through all the trouble
- 11:00to write these complex rules, why do the sources
- 11:03explicitly warn us to keep the global instructions
- 11:06relatively short? Why not just put the band words
- 11:09right there in the main settings? This raises
- 11:11an important point about system architecture.
- 11:14Global instructions exist only to define the
- 11:16workspace logic. If you shove your one to five
- 11:19ranking system, your band words, and your source
- 11:21URLs into the global instructions, the cognitive
- 11:24load is too heavy at the top level. I get it.
- 11:27The traffic director gets overwhelmed trying
- 11:29to build the car while directing traffic. Great
- 11:31analogy. keep the global instructions short so
- 11:34they act as a stable foundation just pointing
- 11:37to the folders. The complex, specific rules belong
- 11:40safely isolated inside those separate markdown
- 11:43files. The traffic director just points. The
- 11:45blueprint does the heavy lifting. I like that.
- 11:47So the traffic director points the AI to the
- 11:50sources, filters the news through the rules,
- 11:53and then point it to the template. Right. Inside
- 11:56your templates folder, you have a file called
- 11:58inews -template .md. Exactly. And a weak dashboard
- 12:03template just asks for a list of headlines. A
- 12:06powerful, useful dashboard demands structured,
- 12:09actionable intelligence. The sources show this
- 12:12template dictating a very specific final output.
- 12:15It must include an executive summary. It must
- 12:17display the impact scores for each item that
- 12:19we talked about earlier. But the most valuable
- 12:21part is the requirement for content angles. It
- 12:25literally forces the AI to provide a newsletter
- 12:27idea or a LinkedIn post idea based specifically
- 12:30on that news item. What's happening here is the
- 12:32transformation of passive information into active
- 12:34utility. You aren't just reading that Anthropic
- 12:37released a new feature. Your system is handing
- 12:40you immediate, actionable ways to use that news.
- 12:43Which is huge. Right. For example, if you are
- 12:46a developer, your template might demand a code
- 12:49integration angle. If you are a marketer, it
- 12:51demands a campaign angle. You are tailoring the
- 12:54news directly to your daily deliverables. It
- 12:57is brilliant. We have our isolated folder. We
- 12:59have our assembly line built. We have our traffic
- 13:02directing global instructions. Now it is time
- 13:05to flip the switch and watch this digital employee
- 13:08go to work. Let's talk about execution. This
- 13:12is where you see the magic of the architecture
- 13:14you've built. You open a new task in Cowork,
- 13:16ensure it is pointed at your specific Cloud Cowork
- 13:19News dashboard workspace, and give it a simple
- 13:22one -line instruction. Like what? Just create
- 13:25today's daily AI news dashboard. That's it. One
- 13:28sense. That's it. Because of all your setup,
- 13:31Claude knows exactly what to do. It checks the
- 13:33blogs on your approved source list. It evaluates
- 13:35the rumors versus the facts. It ranks the updates
- 13:38one to five. It forces its neural pathways to
- 13:41avoid the banned words. And it formats everything
- 13:44into your template. And along the way, it might...
- 13:47pause and ask you questions, right? Yes, and
- 13:50this is a vital part of the co -work process.
- 13:52It might ask clarifying questions before it starts
- 13:55generating the final report. Give me an example.
- 13:57Well, it might say, I found three updates that
- 14:00score a four today. Do you want me to expand
- 14:02the content angles for all three or just focus
- 14:05on the top story? OK, that's helpful. You should
- 14:07always answer these questions. Providing that
- 14:10context prevents the AI from making weak assumptions
- 14:12and filling in the gaps itself. But I have to
- 14:15push back on this step. If the ultimate goal
- 14:17of this entire deep dive is to have this thing
- 14:20work while I sleep, why am I sitting here answering
- 14:23clarifying questions? I know, it feels counterintuitive.
- 14:26Why is this first manual run so critical? Why
- 14:29can't I just build my folders, set my rules,
- 14:32and immediately automate it to run every morning?
- 14:35It's a very common impulse, but if you automate
- 14:37a flawed system, you just get flawed outputs
- 14:39faster. Oh, that's a good point. The manual testing
- 14:42phase is your safety net. It allows for early
- 14:45corrections before you commit to a fully automated
- 14:47schedule. During this first run, you watch the
- 14:50tool activity window. If the AI grabs a source
- 14:54that isn't on your list, or if it slips up and
- 14:56calls something game -changing, you catch it
- 14:58immediately. And if it does make a mistake, if
- 15:00it hallucinates or breaks a rule, is that just
- 15:04the AI being stubborn? Rarely. And this is a
- 15:07hard truth for a lot of users to accept. If the
- 15:10results feel off, it is almost always a thaw
- 15:13in the system design, not the AI itself. Interesting.
- 15:16It's an error in your instructions. Maybe your
- 15:19definition of a level five impact was too vague.
- 15:22Maybe your source list had a broken URL. The
- 15:24manual run lets you debug your own rules. OK,
- 15:27so we've debugged it. We tweaked the blueprint.
- 15:29We ran it manually. And the output is beautiful.
- 15:31It is short, it is actionable, it is verified,
- 15:34and there's no marketing fluff. Perfect. Now
- 15:37we reach the holy grail, scaling and automation.
- 15:40Once your manual run is perfect, you can utilize
- 15:42the scheduling feature. You literally just type
- 15:44slash schedule in the task bar. Wow. You can
- 15:46tell Claude to run this exact dashboard workflow
- 15:49every single morning at 8 0 a .m. Just typing
- 15:53slash schedule. That feels like magic. Yeah.
- 15:56But. We need to be incredibly clear about the
- 15:59technical reality here because this is where
- 16:00people get frustrated. Yeah, there's a catch.
- 16:03What are the strict requirements for a scheduled
- 16:05task to actually fire? The requirements are physical.
- 16:09This workflow is not running on a server somewhere
- 16:11out in the cloud while you are disconnected.
- 16:14Claude Cowork operates locally. Explain the mechanics
- 16:17of that. Why does my computer need to be involved
- 16:20if the AI lives in the cloud? Because the AI's
- 16:23brain is in the cloud, but your workspace, your
- 16:26about folder, your templates, your outputs lives
- 16:28on your physical hard drive. For the scheduled
- 16:30task to execute, it physically needs to access
- 16:32the files on your hard drive, read them into
- 16:34its memory, process the data, and save a new
- 16:37Markdown file back to your local drive. That
- 16:39makes sense. If your machine is asleep or if
- 16:42the desktop app is closed, the bridge between
- 16:45Claude's cloud brain and your local files is
- 16:47completely subbered. The digital employee can't
- 16:50get into the office, so your claw desktop app
- 16:53must be open and your computer must be awake.
- 16:55So if your laptop is closed in your backpack
- 16:57at 8 a .m., the scheduled task will simply fail.
- 17:02You have to treat it like a real machine that
- 17:03needs power and access. Precisely. But if we
- 17:07pull back and look at what we've accomplished,
- 17:09the core lesson here is profound. Good outputs
- 17:12come from good systems. For the past couple of
- 17:15years, the entire tech world has been obsessed
- 17:17with writing clever, complex prompts. We are
- 17:20finally moving away from simple prompts and moving
- 17:22toward robust processes. You are building an
- 17:25architecture, not just asking a question. That
- 17:27is the perfect summary. Let's distill the actionable
- 17:30advice for you to take away today. You can build
- 17:32this. Start your first workspace folder today.
- 17:35Build your assembly line. Put in the upfront
- 17:37time to create your source list, your ranking
- 17:39rules, and your strict writing style. Keep it
- 17:42incredibly focused. Demand that your outputs
- 17:45are short and verified. And above all, do not
- 17:48grant the AI access to unnecessary files. Keep
- 17:51it strictly contained to its workspace so it
- 17:53doesn't get confused by your personal data. Let
- 17:56the machine work for you, but only on the exact
- 17:59terms you define. And I'll leave you with a final
- 18:01thought to mull over. We just spent this time
- 18:03walking through how to build an AI agent that
- 18:06can independently read industry blogs, rank them
- 18:10by importance based on your custom criteria,
- 18:12and write a human sounding newsletter draft while
- 18:15you are asleep. Yeah. So ask yourself what other
- 18:17repetitive daily digital chores in your life
- 18:20or your business are just waiting to be turned
- 18:22into a scheduled co -work folder. What a question
- 18:24to leave on. The potential is massive. Thank
- 18:27you so much for joining us for this deep dive
- 18:29into building your digital employee. We will
- 18:31see you next time.