Latest / AI Fire Daily / #476 Neil: Claude Guide To 5 Levels From Beginner To Pro Workflow
Transcript
- 0:00You use AI every day. Beat. But that doesn't
- 0:04actually mean you're an advanced user. Right.
- 0:06I mean, it definitely feels advanced, but daily
- 0:08casual use is really just the beginning. I think
- 0:11about this a lot. We type a quick prompt into
- 0:14a chat box. We get a really solid answer. We
- 0:17feel incredibly productive. Oh, yeah. Totally.
- 0:21Beat. But we are barely scratching the surface
- 0:23of what is possible. Welcome to this deep dive.
- 0:28Today, we're charting a very specific path. We
- 0:30are exploring the five levels of cloud proficiency.
- 0:34We're going to move step by step through this
- 0:36entire framework. We'll start with those simple
- 0:38everyday chats that will climb all the way up.
- 0:41We'll explore fully autonomous background workflows.
- 0:44Our goal is to help you pinpoint exactly where
- 0:46you are right now. Right. And more importantly,
- 0:48we want to show you how to level up safely. So
- 0:50let's start at the absolute baseline. Level one.
- 0:53the enthusiast. Yeah, which is honestly where
- 0:55most people tend to live right now. You might
- 0:57assume you're a total power user just because
- 0:59you chat with the AI every single morning. But
- 1:02realistically, this is just the ground floor.
- 1:04Level one is really built for quick daily tasks.
- 1:07You know, you need a fast email draft. You need
- 1:09a short summary of a long article. Right. Or
- 1:12maybe you just need a simple explanation of a
- 1:14complex topic. It just gives you a much faster
- 1:16starting point. You paste in some rough meeting
- 1:18notes. You ask the AI for a clean version. Yeah.
- 1:22A few minutes later, you have something totally
- 1:24useful. It helps you think faster. You create
- 1:26a solid first draft. And then you just finish
- 1:30the actual work yourself. There is a very specific
- 1:33tactic here that I love. One that works incredibly
- 1:36well when you're at level one. What's that? Uploading
- 1:39screenshots directly into the chat. Oh, yeah.
- 1:42That is a fantastic habit to build. You really
- 1:44need to stop typing out every single detail.
- 1:46It just takes way too much time. Exactly. Because
- 1:48you might accidentally miss something important.
- 1:51Just share exactly what you see on your screen.
- 1:53Exactly. So imagine you get a really confusing
- 1:56dashboard error, or you see a weird error message
- 1:59during your work. Right. You panic a little.
- 2:01Yeah. But you don't try to transcribe the code.
- 2:04You just snap a quick screenshot. You upload
- 2:07the image directly. Then you ask it to explain
- 2:10the error in simple English, and you ask for
- 2:13the very next step to fix it. It also works incredibly
- 2:15well for visual content, too. Say you're designing
- 2:18a new newsletter layout, you take a quick screenshot
- 2:21of the draft, you ask the AI to check the title
- 2:24length, or you ask if the call -to -action placement
- 2:27needs clearer wording. It's working straight
- 2:29from the real visual data. But there's a really
- 2:32massive limitation at this stage. a severe lack
- 2:35of ongoing context. Oh, completely. One -off
- 2:38chats are perfectly fine for single isolated
- 2:40tasks, but repeated ongoing work is entirely
- 2:45different. Exactly. Repeated work requires a
- 2:47very specific, consistent tone. It requires deeply
- 2:50knowing the target audience. You have internal
- 2:53company rules. Yeah. You have past business decisions
- 2:56to consider. And in a quick chat, the AI... basically
- 3:00forgets all of that. It has total amnesia. You
- 3:03have to repeat your tone and rules every single
- 3:05time. Honestly, I have a vulnerable admission
- 3:08to make here. Okay, let's hear it. Beep beep.
- 3:10I still wrestle with starting every repeated
- 3:12task in a blank chat. Oh, we all do it sometimes.
- 3:15It just feels faster in the heat of the moment.
- 3:17You just want an answer right now. Yeah, I do.
- 3:19But over a month you waste countless hours re
- 3:22-explaining yourself. It really makes me wonder
- 3:24about the true value here. Is level one basically
- 3:27just using Claude as a glorified short -term
- 3:30search engine? Pretty much. It's a fast starter,
- 3:33but it forgets everything by the next day. Which
- 3:35perfectly highlights the core problem we need
- 3:37to solve. We have to fix that blank chat amnesia.
- 3:42Yes. So let's step up to level two, the beginner.
- 3:45We fix this by building a persistent workspace.
- 3:48Level two is entirely focused on creating continuity.
- 3:52Your important work continues across multiple
- 3:54conversations. It continues across your files
- 3:57and your connected data sources. This level relies
- 3:59heavily on three core features, projects, memory,
- 4:03and past chat search. Right. Memory naturally
- 4:05retains your specific working preferences over
- 4:08time. Past chat search lets you easily find older
- 4:11conversations. If you can just open a chat and
- 4:13say, find the campaign direction from last week.
- 4:16Exactly. Then you ask it to prepare the final
- 4:18outline based on that. It seamlessly maintains
- 4:21the historical thread. But it gets really interesting
- 4:23when we bring in connectors. You can actually
- 4:26connect the AI to the information you already
- 4:28use. Yeah, this is a huge step up. Google Drive,
- 4:32Gmail, Slack, Google Calendar, Microsoft 365.
- 4:36Wow, that's a lot. Your real... actionable context
- 4:40already sits inside those apps. Connectors pull
- 4:43that securely approved information right into
- 4:45your active conversation. So you don't have to
- 4:47manually collect every single detail anymore.
- 4:49You don't have to constantly switch tabs to copy
- 4:52and paste. You can just ask it to review a specific
- 4:55client email thread. Right. Or review the latest
- 4:59strategic brief directly from Google Drive. Let's
- 5:01look at a very specific, practical example here.
- 5:04Creating a highly structured meeting summary.
- 5:07You upload your messy, raw client notes. Yeah,
- 5:10we've all been there. And you want a professional
- 5:12Word document out of it. The prompt for this
- 5:14needs to be very deliberate. You ask it to clearly
- 5:17include the meeting background. You explicitly
- 5:20ask for the confirmed decisions. You ask for
- 5:22a list of questions that still need confirmation.
- 5:26And finally, you demand a next actions table.
- 5:30That table must include the responsible person.
- 5:32and the specific deadline. But here is absolutely
- 5:35the most important part of that workflow. You
- 5:37give the AI a very strict, unbreakable rule.
- 5:41Which is? If any information is missing, Claude
- 5:44must write, not confirmed. It is absolutely not
- 5:47allowed to guess or fill in the blanks. I like
- 5:49to think about it this way. Level one is a public
- 5:52whiteboard that gets wiped clean every night.
- 5:54I like that. Level two is a dedicated office
- 5:57desk where your files stay exactly where you
- 6:00left them. That's a perfect way to picture the
- 6:02difference. It reliably turns messy inputs into
- 6:05clearly structured files of files you can actually
- 6:08review, trust, and use in your job. But I do
- 6:11have a lingering concern here. With all those
- 6:14connected apps, can we trust it not to just invent
- 6:16information? It's safer, but you still must review
- 6:19the structured output before sending it anywhere.
- 6:22Okay, so we've got our files beautifully structured
- 6:23on our metaphorical desk, but I'm still the one
- 6:26opening the folders and doing the actual manual
- 6:28labor. Right. I'm still clicking and dragging.
- 6:30Exactly. And that bottleneck is why you have
- 6:32to level up again. Welcome to level three, the
- 6:35intermediate. We've mastered text outputs. Now
- 6:39we give the AI actual hands. We're talking about
- 6:42a feature called Claude Cowork. This is where
- 6:44it actively starts operating your computer for
- 6:47you. It handles complex multi -step tasks directly.
- 6:51It moves way beyond just reading your text files.
- 6:54It actually starts manipulating the files on
- 6:56your machine. Okay. So imagine you have a wildly
- 6:59messy downloads folder. It's just stuffed with
- 7:02random invoices, notes, and PDFs. You can give
- 7:06Cowork a very specific multi -step directive.
- 7:08You literally tell it to sort that folder by
- 7:11file type. You tell it to rename every single
- 7:13file to a standard format, like year, month,
- 7:16day. Then you ask it to summarize the contents
- 7:19of all those files. And if a file is unreadable
- 7:21or unclear, it doesn't just crash, it neatly
- 7:24moves it to a dedicated folder. Like a needs
- 7:26review folder. Exactly. It does all the tedious
- 7:28heavy lifting for you. But we really need to
- 7:31talk about safe file access here. This is absolutely
- 7:34crucial. When you give an AI direct file access,
- 7:38you must start small. You have to set up isolated,
- 7:41protected folders. Specifically, you should have
- 7:44three distinct folders for this. Right. Source
- 7:46files, working files, and outputs. Let's break
- 7:50that down. Source files are the read -only documents
- 7:53it can look at. Working files are the temporary
- 7:56sandbox where it actively processes data. Outputs
- 7:59are the final finished results for you to manually
- 8:02check. Exactly. You never let it loose on your
- 8:04main hard drive. That's just asking for a disaster.
- 8:07This isolated structure leads to something really
- 8:10powerful, though. Reusable skills. Oh, yeah.
- 8:13When a tedious task repeats every single week,
- 8:16you save it as a skill. Let's use a weekly client
- 8:19report generator as a solid example. You build
- 8:21a custom skill to process marketing data every
- 8:24Friday. OK. You establish strict rules for the
- 8:27output. You require five specific sections every
- 8:29single time it runs. So you'd tell it, I need
- 8:31an executive summary at the very top. It's a
- 8:33short, crisp summary of the week's performance.
- 8:36Right. And then you force it to extract the key
- 8:39results. It pulls the important, verified metrics
- 8:43directly from the uploaded data. Right. And you
- 8:45want to compare those numbers to the previous
- 8:47week. So you add a section for important changes.
- 8:49Yeah, exactly. You're building a reliable, repeatable
- 8:52template. And you'd probably want it to flag
- 8:54anything that's missing, too. Definitely. That's
- 8:56your open questions section. It lists any unclear
- 9:00data points. Then it always ends with a clear
- 9:02next actions table. Action, owner, and deadline.
- 9:05Once you save that specific skill, your weekly
- 9:09workflow changes entirely. How so? You just upload
- 9:12the fresh raw data each week. You hit run on
- 9:14the skill. It automatically follows the exact
- 9:16same five -step process without you managing
- 9:18it. You can even schedule these simple tasks
- 9:21to run automatically. Say you're using a cloud
- 9:24desktop application. Yeah. You could easily set
- 9:26up a customized briefing for 7 .30 p .m. every
- 9:29weekday. It quietly checks your connected sources
- 9:31in the background. It stands Gmail. It checks
- 9:33your calendar. It reads your Slack messages.
- 9:36It pulls all that context together. Then it prepares
- 9:39a short summary of tomorrow's meetings and any
- 9:41pending actions. It feels incredibly powerful.
- 9:44But I really have to ask this. Isn't giving AI
- 9:47direct file access a recipe for accidentally
- 9:50deleting your hard drive? Only if you're reckless.
- 9:53Always use copy files in isolated folders to
- 9:56test first, sponsor. We're back! So we just covered
- 10:00automating desktop chores in Level 3. Now we
- 10:03really scale things up. We do. Welcome to Level
- 10:064, the Advanced Builder. This is a major jump.
- 10:09This is where we introduce Claude code. We're
- 10:11moving away from simple folder sorting tasks.
- 10:14We're getting into complex multi -file project
- 10:16engineering. This requires a real dedicated project
- 10:19structure. And crucially, it requires a Claude
- 10:22.md file. Let's quickly define that jargon, a
- 10:25simple text file that tells the AI your project
- 10:27rules. It basically acts like a strict working
- 10:29guide for the system. It clearly defines the
- 10:32overarching project goal. It explicitly explains
- 10:35what all the main folders are actually used for.
- 10:37Right. It lists the specific files the AI is
- 10:39allowed to edit. And crucially, it lists the
- 10:42files it must absolutely leave unchanged. It
- 10:45also tells the AI exactly how to run your internal
- 10:47tests. And it defines what the final report format
- 10:50should look like. But there is a very critical
- 10:53step here that you cannot skip. You must ask
- 10:55the AI to plan and explain its approach first.
- 10:59Yes. It has to do this before it makes any actual
- 11:02edits. It's like asking a contractor for blueprints
- 11:05before they knock down a wall. You prompt it
- 11:07to read the folder structure. You ask for a detailed
- 11:10step -by -step plan. It needs to explain which
- 11:12files will change and exactly why. You review
- 11:15that plan. You approve it. Only then does the
- 11:18actual coding work begin. As these projects get
- 11:21bigger, things naturally get much more complicated.
- 11:24One single session can quickly become overloaded
- 11:27and confused. Yeah, for sure. So we have to deliberately
- 11:29split the workout. We split big work into parallel
- 11:32sessions. We use specialized tools like sub -agents
- 11:35and work trees. One stream handles the front
- 11:37-end layout. Another stream prepares the backend
- 11:40tests. A third stream carefully reviews the overall
- 11:43logic. This keeps the different work streams
- 11:45entirely separate and clean. They happen at the
- 11:48exact same time. but they don't accidentally
- 11:51interfere with each other's code. And you always
- 11:53have a mandatory non -negotiable review step
- 11:55built in. The AI absolutely must check the result
- 11:59before you accept the files. It must run all
- 12:01the available system checks. If a check fails,
- 12:05it has to explicitly explain the issue. It fixes
- 12:08the error itself. It runs the check again. Finally,
- 12:11it prints out a list of every single changed
- 12:13file. You carefully look at that list. You verify
- 12:16the changes. Then you formally consider the task
- 12:19complete. But I'm genuinely curious about the
- 12:21mechanics of this process. Why bother splitting
- 12:23tasks into parallel sessions instead of just
- 12:26one giant prompt? To keep the AI focused, preventing
- 12:28it from getting confused, and mixing up changes.
- 12:31That makes perfect sense. Clear rules. Clear
- 12:33inputs. Clear safety checks. Which finally brings
- 12:36us to the absolute summit of this framework.
- 12:39Level five, the architect. This is fully autonomous
- 12:43infrastructure. We're completely removing the
- 12:45human from the active day -to -day loop. We're
- 12:48finally entering the realm of true background
- 12:51automation. You've already set the strict rules
- 12:53in Level 4. Now, the complex workflow runs without
- 12:57you even sitting at the screen. But the hardest
- 12:59part of Level 5 isn't actually technical. Beat.
- 13:02It's trust. Trust is incredibly difficult here.
- 13:06You're letting an AI touch real files and critical
- 13:09business data in the background. Giving up that
- 13:12control is scary. You have to build this autonomy
- 13:14very, very slowly. At this advanced stage, tasks
- 13:18start from automatic triggers. a time schedule,
- 13:21a brand new email arriving, a specific GitHub
- 13:24event, or an external API call. We use a system
- 13:27called Cloud Routines for this. These are secure
- 13:29cloud -based workflows. They run entirely on
- 13:32Anthropic's own back -end infrastructure. They
- 13:34don't rely on your local desktop being turned
- 13:36on. So you can set a routine to automatically
- 13:38review changed files when a pull request opens.
- 13:41Or you could have it create a massive weekly
- 13:43project summary every Friday afternoon. But we
- 13:45desperately need safety mechanisms here. This
- 13:48is exactly where hooks come in. Let's define
- 13:50that term clearly. Safety gates that check the
- 13:53AI's actions before and after. Right. There's
- 13:56a pre -tool use hook. It acts just like a strict
- 13:58security gate. Before the AI actually runs a
- 14:01command, the hook checks it. If the proposed
- 14:04action is risky, like deleting important files,
- 14:07it blocks it instantly. It asks the AI to rethink
- 14:11and correct the action. Then there's the post
- 14:13tool use hook. After the tool successfully finishes,
- 14:16this hook handles the necessary cleanup. It might
- 14:19automatically format a text file. Or it might
- 14:22send you a quick Slack notification that the
- 14:24background task is done. The entire workflow
- 14:26is fully automated. But those precise hooks keep
- 14:29it tightly and safely controlled. You still want
- 14:32to start this process very safely though. We
- 14:34highly recommend starting with a daily private
- 14:36morning briefing. Yes. Just let it run quietly
- 14:38every morning at 6 a .m. It checks your upcoming
- 14:41calendar. It safely scans your important emails.
- 14:43Right. It reviews your missed slack threads.
- 14:46It prepares a comprehensive summary. But it sends
- 14:49that summary only to you. It doesn't email your
- 14:52boss. You review it over coffee before taking
- 14:54any action. Yeah, you let it run like this for
- 14:57a full week. You check if it misses nuances or
- 15:00messes up your actual priorities. You carefully
- 15:02adjust the background instructions. Only after
- 15:05definitively proved reliable do you give it a
- 15:07more important workflow. Exactly. Beat. Whoa.
- 15:11I mean, imagine truly scaling this up. Imagine
- 15:15an AI autonomously running your complex workflows
- 15:17in the background while you sleep. It's wild.
- 15:21Perfect checks. Perfect balance. Waking up to
- 15:23finished work. It's an absolutely amazing thought,
- 15:26but it clearly demands immense caution in planning.
- 15:29Which naturally raises a really big question
- 15:31for me. How do we stop an autonomous system from
- 15:34going rogue and sending a bad email? You use
- 15:36those hooks to block risky actions and strictly
- 15:39require human review first. Okay, let's bring
- 15:42this all back down to earth for a minute. We
- 15:43need a big idea recap to tie this together. The
- 15:46core insight from all of this source material
- 15:48is surprisingly clear. Advanced use isn't really
- 15:52about writing better, more clever prompts. It's
- 15:55entirely about clearer workflow design. You don't
- 15:58just magically jump to level five. You don't
- 16:00blindly rush into full automation. You have to
- 16:02build the foundational structure first. Yeah,
- 16:04you add proper folder structure, you add firm
- 16:06safety limits, you add mandatory review steps.
- 16:09That's how this technology becomes truly, reliably
- 16:12useful. So here is your practical takeaway for
- 16:15today. Look closely at your own workflow. Pick
- 16:18just one repeated week task that annoys you.
- 16:21Just one task and upgrade it by exactly one single
- 16:25level. If you're stuck at level one, make a project.
- 16:27If you already have a project, connect one live
- 16:29data source. If you have connected sources, try
- 16:32a basic co -work task. Build a safer, more structured
- 16:35workflow. Add a mandatory review step. One single
- 16:38intentional upgrade will teach you way more than
- 16:41trying to build a massive autonomous system on
- 16:44day one. It forces you to truly understand the
- 16:46underlying mechanics. It forces you to actually
- 16:48build trust with the tool. Step by step, level
- 16:51by level. That's how you actually get advanced.
- 16:53It's been a truly fascinating journey through
- 16:55these five levels today, from the eager enthusiast
- 16:58all the way up to the architect. But I want to
- 17:00leave you with a final slightly provocative thought
- 17:03to mull over. Beat. Let's say you successfully
- 17:06level up. You reach a point where AI flawlessly
- 17:08remembers your context, where it perfectly plans
- 17:11your complex projects, where it autonomously
- 17:14executes your daily routines with perfect hooks
- 17:16and safety checks.