Claude AI Complete Course — Step-by-Step Guide & Transcript

[00:00] What if your AI system could do more than just answer questions? What if it could help you think, create, analyze and even build software with you?

[00:13] Hello everyone and welcome to the Claude course where we explore how Claude has evolved into a powerful AI ecosystem designed for your world work. So in this course you will go beyond the basics of AI chatbots and understand how Claude's different models work,

[00:26] how Anthropic builds safer and more reliable AI systems, and how you can use Cloud effectively for productivity, collaboration, and software development. So whether you're a developer building applications, a marketer creating content,

[00:38] an analyst working with information, or someone exploring the possibilities of AI, then this course will help you understand where Cloud fits and how to get the most value from it. So we will begin by exploring the Cloud model family,

[00:50] understanding the difference between models, and how to choose the right one for different tasks. Next, we will dive into constitutional AI, the unique approach Anthropic uses to train Cloud to be helpful, honest, and safe. Then we will move into practical applications where you will learn how features like memory,

[01:05] vision, and artifacts transform Cloud into a powerful workspace for every data. Finally, we will explore Cloud's development capabilities and see how you can use its coding agent to plan, build, test, and define real-time projects.

[01:17] So by the end of this course, you will have a complete understanding of Cloud from how is built to how you can use it effectively in your personal and professional workflows. Also, if you want to go beyond using AI systems like Cloud and learn how to build advanced

[01:30] generative AI applications, you can explore Syntelron's Applied Generative AI and Agentic AI specialization, delivered in collaboration with Microsoft. So, this 12-week program covers key generative AI concepts, including RAG, Multimodal AI,

[01:42] Agentic AI, Autonomous Agents, Multi-agent Systems, MCP, Deployment, Monitoring, and Responsible AI. You will also work with tools and platforms such as OpenAI, Azure AI, GitHub Copilot, LandGraph and N18 through hands-on projects and capstone.

[01:56] This program is designed to help you move from simply understanding AI concepts to building and deploying production-ready enterprise AI solutions. You also earn a completion certificate from Microsoft and Simply Learn. So if you want to build practical Gen-AI and genetic AI systems for dual-world applications,

[02:11] then make sure to check out this program through the link in the description box below. So now let's get started. Hello everyone, welcome to this course on Cloud AI.

[02:24] This course is structured with one goal in mind, to give you a practical, structured understanding of how Query actually works and how to use it well across different ways it shows up in your day-to-day work. So whether you're a developer, marketer, analyst, or simply curious about what AI

[02:39] assistance actually looks like in practice, Cloud has evolved far beyond the basic chat interface. To date, it includes a range of modules with different capabilities and sizes developed through our distinctive safety-focused training process and integrated into several products

[02:53] designed for different types of work. This course explores that broader cloud ecosystem, how the model differs, how they are built and how the various tools can be used in real-world scenarios. This course walks you through the complete picture of choosing the right model for a

[03:07] task through understanding how cloud is actually trained to be safe and genuinely helpful all the way to use in cloud for daily productivity for delegated collaborator work and to professionals of development so in this course we will cover four lessons lesson one introduces the cloud

[03:22] model family the different sizes and tires anthropic offers and how to think about which one sets a given task lesson two covers constitutional ai and sanctuary how cloud is actually trained to be careful and honest without becoming aggressive or unhealthy lesson

[03:37] three covers daily productivity memory vision and handicaps the three features that are single conversation into an ongoing genuinely useful workspace. This lesson is hands-on. We'll go screen by screen inside cloud AI. Lesson 4 focuses on building a project in cloud using its terminal

[03:53] based development agents to plan, write, test and define software directly within a coding workflow. By the end of this course, you will have a clear understanding of the cloud ecosystem

[04:06] and the different capabilities it offers. We will look at cloud's model family and learn how to to choose the right model depending on the complexity and requirements of a task. You will also understand the ideas behind constitutional AI and how anthropic approaches

[04:19] safety while keeping CLOAR helpful and practical. From there, we will move into hands-on features such as memory, vision, projects and handicaps and see how they can support your day-to-day work. Finally, you will learn how skills and connectors can help you create more consistent, repeatable

[04:34] workflows and extend what CLOAR can do beyond a basic conversation. Welcome to this lesson. In this lesson, we will build a strong foundation in how CLOARD is organized as a product and understand why picking the right model matters for every task you will ever hand it.

[04:53] We will begin by looking at where CLOARD came from and what the name actually means. Next, we will explore the current lineup of models, HICO, SONNET, OFFICE, USMDSR, ABASIM and understand what each one of them is actually built for.

[05:06] We will also discuss how the industry measures whether one model is genuinely better than the other through benchmarks. Finally, we will also look at Claude's science, a reason and specialized products so you can understand where anthropic roadmap is actually heading beyond general chart.

[05:19] So let's get started. So what comes into your mind when I say Claude? Claude is an AI system built by anthropic to support clear thinking, writing, coding, analysis and structural problem solving. Its usefulness comes from more than just generating answers.

[05:33] With strong prompting, Claude can understand the role it should take, the context of a task, the constraints it needs to follow and the format of the expected result. Cloud Code extends that capability into software projects where Cloud can help inspect code, plan changes, debug the issues, and test the solutions.

[05:49] Agenda workflows take this further by allowing Cloud to work with tools, maintain context across multiple steps, and support more complete actions instead of responding to each prompt and isolation. Before you can use Cloud well, you need a working mental model of what you've actually chosen between every time you open it.

[06:05] This lesson builds that model and by the end of it, you will be able to look at almost any task and immediately know which model is the most efficient. Let's start with a question most people never stop to ask. Should I always use the most powerful model available or does it actually waste time and money on tasks and don't need it?

[06:21] Now for our trivial question, it has a genuine answer. By the end of this lesson, you will know how to answer it to yourself, task by task. So why model choice matters? Here's something worth sitting with for a moment. for a moment. Every single cloud reports run through a specific model whether you consciously

[06:36] choose it or not and that choice has real consequences for speed, cost and depth of reasoning you get back. Using a small fast model for a task that genuinely needs deep reasoning it's your shallow answer. Using the largest, most expensive model for a simple reparative task

[06:51] wastes both time and money without improving the result. The understanding the lineup is what lets you make that call deliberately instead of defaulting to whatever happens to be selected. So where Claude came from? Claude is a flagship AI model family from Anthropic, a safety-focused AI company founded in 2001 by former open-air researchers including siblings Daniel and Daniela.

[07:12] The founding team had spent years working on large-language models and specifically to build a company where safety research and capability research happen together under one roof rather than safety being treated as an afterthought.

[07:25] Claude the Chadbot first went public in March 2003. The name itself is deliberate. Claude is named after Claude Shannon, the mathematician and engineer often called the father of information theory, the person who figured out how to mathematically measure and compress information decades before anyone had a personal computer.

[07:42] It's a fitting name for a model built to understand and compress words' information into useful, digestible answers. Since the first release, Anthropic has shipped new model generations at a genuinely fast pace. Multiple releases a year, each one narrowing the gap between expensive and slow and cheap and fast,

[07:59] while simultaneously pushing the very top end of what's possible further out. Those are actually two separate improvement curves happening at once, and understanding what explains why a mid-tire model today

[08:11] can outperform last year's top tire model at a fraction of the cost. You don't need to memorize every version number that ever should. What you need is the current mental model for how Anthropic organizes its lineup because it's what tells you which models to reach for in any given task.

[08:26] So think of the current Cloud lineup in three everyday tires, plus a frontier tire above them. At the everyday tire, you've got Haiku. Right now, that's Cloud Haiku 4.5. It's built to be fast and cost-efficient, quick-replied, lightweight automation, high-volume tasks,

[08:42] where your fighting of thousands or ten-thousands of requests and speed and price matters more than squeezing out the last few points of recent quality. Think customer support trials, content moderation and scale, or summarizing a large batch of short documents,

[08:56] where each individual task is fairly simple. One step up is Flonit. Flonit Format 5 is the current version. This is an anthropic-agentic workhorse. The model most people reach for by default because it balances strong reasoning with real speed in a way neither of the other tyres quite matches.

[09:13] It is a model doing the heavy lifting inside tools like cloud code and cloud co-work where it needs to plan a multi-step task, decide which tools to call and follow instructions across many turns reliably without being painfully slow.

[09:26] If you want to remember one model name from this course, make it this one. It's the same as default for the overwhelming majority of daily work. and it is called Everyday Tyres Opus. This tyre recently changed shape in a way worth walking through carefully.

[09:40] For most of my research, Opus 4.8 was a flagship reasoning model. The one you reached for when a touch genuinely needed the deepest available thinking and you were willing to pay for it. As of late July, Anthropic replaced its Lord Opus 5,

[09:53] a model built specifically to land close to Anthropic's most giveable model, at roughly half the price with an adjustable F4 dial that lets you trade off costs against depth of reasoning on a per-task basis.

[10:05] Obert 5 is now the default model on CloudMax and the strongest option available on CloudPro. CloudFable 5 and CloudMigas 5 both launched in June 2016 and they actually share the same underlying model. Fable is the version released with extra safety measures layered on top for the most sensitive capability areas,

[10:22] specifically biology, cybersecurity, and AI research itself. Midas is reserved for a small number of trusted organizations working directly with Anthropic under a program called Project Glasspring. So if you're a working professional using Flot day-to-day, Tableau is the one you might actually get your hands on.

[10:38] It's Anthropic's most capable, widely-released model, and Anthropic still recommends it for the hardest, longest-running project, even now that Opus 5 exists at a fraction of the price. Worth trying, honestly, because it's a genuinely useful case study.

[10:51] Shortly after launch, US export controls briefly forced Anthropix to suspend access to Fabel and NITOS entirely for about 2.5 weeks in June. That being resolved, the controls were lifted and Anthropix restored the access on July 1st.

[11:04] This is generally fast-moving space shaped by more than just engineering, regulation and export policy, such it took. Bill habits around checking Anthropix's own documents and release notes directly, rather than memorizing a fixed line-up from any code.

[11:18] So which one do you actually pick? Here's a practical thumb rule. You can use Sonnet for anything like junket coding, help, multi-step workflows using tools, day-to-day drafting, and research. Teach for hypo when you're automating something at scale

[11:31] and don't need deep reasoning on every single cost. Teach for opus when the task is genuinely hard, complex analysis, tricky debugging, start-up-read-writing, where the stakes are high, and you want the best available reasoning

[11:43] without paying full-fledged prices. And teach for fabul only when the task is at the very edge of what's possible. Research-grade problems, extremely long autonomous joint situations where being the right matters more than being fast or cheap.

[11:56] A practical habit worth building is start with Sonnet by default and only step up when you actually hit a wall. That is when the output is shallow or clearly needs more depth than it got. Setting down is just as valid a move.

[12:09] If you notice you're using Opus for something genuinely simple and repetitive, that's a signal to try HiCo instead and see if the quality holds up at a fraction of the cost. So how does Anthropik and the rest of the industry actually decide which one model is stronger than the other?

[12:23] Through benchmarks. Standardized tests, the whole field runs models against so the claims made by competing labs are at least somewhat comparable. A few words knowing my name. GPQI is a graduate level science reasoning test genuinely hard enough that even domain experts only get a portion right without help.

[12:41] Making it a good proxy for deep, careful reasoning rather than pattern matching. Coding as a genetic benchmark like SWEA bench verified through models at real unresolved GitHub issues and checked whether the generated fix actually passes the existing test suite.

[12:55] Newer tools like Anthropix's own frontier bench and GDP-Val AA lean further into real-world knowledge work tasks. One honest warning worth carrying into any content you build on this topic yourself is that benchmarks code move fast.

[13:08] Leaderboard reshuffles with basically every release from every major lab. So if you are citing a specific number, always lean back to the light source rather than baking a snapshot into evergreen content. It will be stale within a week. Treat benchmarks the way you would treat a stock price.

[13:21] Directionally useful, never something you screenshot once and fast forever. One more thing worth understanding is because it's a great example of where Anthropism is heading beyond general purpose chat. In June 2016, Anthropity launched to our clients and here's the important nuance.

[13:36] It's not a new, more powerful model. It's a dedicated workbench running on the exact same cloud models everyone else already has access to, purpose-built for scientists and researchers. It comes pre-wired to more than 60 scientific databases spanning genomics, single-cell biology, proteomics, and chemistry

[13:54] so a researcher can ask a question in plain language and have specialist agents tool and synthesize across sources they'd otherwise have to query one by one. Every figure generates ships within the exact code and computing environment that produced it.

[14:07] The resulting science stays reproducible. Why does this matter if you're never going to touch a protein database? Because it's a pattern to watch across anthropic roadmap. General purpose intelligence wrapped in vertical workflow-specific products

[14:20] tuned for a particular kind of progression. So now we have come to the end of this lesson. So let's just quickly have a recap of whatever we've learned. Cure comes in three everyday tires. Hypo for speed and scale, Sonnet as the reliable default,

[14:32] Opus for hard problems on a more reasonable budget, plus a frontier tire, fabel and mysers above them sonnet 5 is the same as default for most engendic and day to day work. Sebaqo Opus or fabel only wear tiles genuinely demands more depth.

[14:46] Benchmarks like GTQA, SWE bench and Centenia bench offer a rough fast moving comparison between models. Always check light numbers rather than just start seeing a static figure. Cloud lines shows where anthropo-zero map is heading.

[15:00] Not one bigger general model but focus connected to a spilt on top of the model we already has. Welcome to this session. In this lesson, we will build a strong understanding of how

[15:12] CLOAR is actually trained to behave and not just what it can do, but why it responds the way it does. We will begin by looking at constitutional AI, anthropic score training method and understand why it exists as an alternative to purely human-level training. Next, we will also explore the two-stage

[15:28] process that method actually follows. We will also discuss the responsible skill in policy and AI safety levels, a separate framework that governs when Anthropic releases more capable models at all. Finally, we will look at what all of this means practically for how you experience Claude

[15:42] day to day. So, let's get started. If you have ever used an AI assistant that refused a completely reasonable request, hedged everything into meaninglessness, or lectured you when you just wanted a straight answer, you have felt the downside of the safety training time badly.

[15:56] This lesson is about how Anthropic tries to avoid that outcome. Training clouds to be careful about real, meaningful harms without making it evasive, preachy or practically useless to the overwhelming majority of requests, which are completely ordinary.

[16:09] The technique behind it has a name, Constitutional AI. So, Constitutional AI or CEAI is an anthropic method for training a model's values and behavior using an explicit written set of principles, a constitution, rather than relying purely on humans hand-labeling thousands upon thousands of individuals good and bad responses one at a time.

[16:28] The second approach called reinforcement learning from human feedback works and is still used elsewhere in the industry but it's expensive, slow to scale and every individual human laborer brings their own blind spots and inconsistencies into the training data

[16:43] A constitution gives a model a single consistent inspectable set of rules to reason from instead of something closer to training a new employee against a written company handbook than training them purely through trial and error

[16:56] Critically, there isn't a secret internal document. Androgan publishes its constitution publicly, so anyone can go read the actual principles the model is training against and see how they've evolved over time as the technology, the use cases, and the risks have all changed.

[17:11] Constitutional AI training can be understood as a two-stage process. The first stage focuses on teaching the model how to recognize problems, elicit responses, and improve them. The second stage uses those input judgments to shape the model's behavior more systematically

[17:25] through reinforcement learning. Let's start with stage 1, which is the supervised learning stage. This is sometimes described as a critic and device logic. The model is first given a prompt and asked to produce a normal response. At this point the answer may be helpful but it may also contain issues.

[17:39] It could be too aggressive, too permissive, misleading, unsafe, biased or simply inconsistent with the principles and they both want the model to follow. To ensure that immediately asking a human reviewer to fix the response, the model is

[17:53] given another instruction. It is asked to examine its own answer using principles from the constitution. For example, it may be asked to check whether the response is unnecessarily harmful, whether it respects the user, whether it avoids encouraging dangerous behaviour

[18:06] or whether it is refusing something more broadly than necessary. The important part is that the model is not simply told the answer is wrong. It is asked explain what is wrong with the answer. That is a critic step. The model might identify a response

[18:21] with too much harmful detail that it makes an unsupported assumption about the user's intent or that it refuses an entire request even though part of the request could safely be answered. Once a critic is generated, the model is asked to produce a revised answer so the sequence becomes

[18:37] very simple. First generate an answer, then critic that answer against the constitution, finally rewrite the answer using this critic. You can think of it like an experienced writer working with an editorial setlist. A writer may produce a first draft, comment against the style guide, identify weaknesses,

[18:52] and then rewrite the draft before publishing it. The difference here is that these sources can be repeated automatically across a very large number of examples. Over time, the model begins learning patterns from those revisions.

[19:04] It becomes better at recognizing what a constitutionally aligned response is supposed to look like without meeting the entire critic and lead-head process every time it answers a user. But that is only the first stage. The second stage uses reinforcement learning.

[19:17] Now, the goal is not just to create better individual responses. The goal is to train the model so that it increasingly prefers better behavior in the first place. Here, the system may generate multiple plausible answers to the same prompt.

[19:29] For example, imagine there are two responses. Response A might be technically correct but unnecessarily harsh or overly restrictive. Is what we might still follow the safety principles while being more helpful, more precise and more respectful.

[19:42] The model is then asked to compare those responses using a constitution as a evaluation framework and determine which method follows the stated principles. That comparison creates what is called a preference signal. In simple terms, the training system now has information saying,

[19:56] for this prompt, this type of response is preferable to that type of response. Those references can then be used to train another model, often called as a preference model or reward model which learns to predict which responses

[20:08] are more aligned with the constitution. The main language model can then be optimized using ghost signals. This process is often referred to as RNAIF which stands for reinforcement

[20:21] learning from AI feedback. That name is important because it highlights one of the biggest differences from traditional reinforcement learning from human feedback or RLHF. In RLHF, humans usually review responses and indicate which one they prefer. That works, but it can become expensive

[20:38] and difficult to scale because generating large amounts of high-quality human feedback requires time, training, and careful quality control. RN-AIS introduces AI-generated feedback into that loop. So, instead of requiring a human to manually evaluate every type of response,

[20:53] an AI system can perform many of those comparisons using constitution as a reference. That does not mean humans completely disappear from the process. Humans are still involved in designing a principle, evaluating the overall system, testing its behavior, and deciding

[21:07] what kind of outcomes are acceptable. But AI feedback can dramatically increase the number of training examples that can be evaluated. And that capability is one of the major reasons constitutional AI is interested. The fourth stage teaches the model how to

[21:22] critique and revise its own behavior. The second stage turns those constitutional judgments into training fitness that implements what kinds of answers the model should first generate in the future. So the goal is not simply to build a model that knows how to refuse dangerous requests.

[21:35] The larger goal is to build a model that can make more nuanced decisions. It should recognize when a request is harmless and it should recognize when only one part of a request creates a safety problem. It should remain helpful wherever possible.

[21:47] And when it does need to refuse something, the response should ideally be clear, proportional and understandable that isn't simply shutting down the entire conversation. That combination of self-critic revision and AI generated preference feedback is what makes

[22:02] this two stage constitutional AI process different from simply writing a long list of safety rules and hoping that model follows them. Now there is also one important thing to understand about clause constitution. It's not meant to be a permanent notebook that gets written once and then can never

[22:17] change it. Antwerp's original work on constitutional AI introduced the basic idea of giving the model a written set of principles that could guide how it critiques, revises, and evaluates its own responses.

[22:30] But the situations people use AI for have expanded enormously. CLAW has no longer been used only for simple questions or basic writing tasks. People now use AI for coding, research, education, business decisions,

[22:42] sensitive personal conversations, creative work, legal and financial information, and many other situations that can involve very different risks and expectations. That means the principles guiding the model also need to evolve.

[22:54] As anthropic encounters new kind of real-world situations, the constitution can be updated to provide more specific guidance about how clause should be hit. For example, newer versions may contain more detailed thinking around areas such as child safety,

[23:07] intellectual property and copyright, privacy, emotionally difficult conversations, harmful requests and situations where the user's intentions are not immediately clear. And this is an important distinction.

[23:19] The consultation is not simply a list of things Claude is allowed to say and things Claude is forbidden from saying. It is closer to a framework for judgment. The goal is to help the model reason about competing consultations. How helpful should it be? How cautious should it be? When should it challenge the user?

[23:34] when should I simply provide the information requested and when is a refusal actually necessary. And if the refusal is necessary, how much of the request can still be answered safely? These decisions are rarely completely black and white

[23:46] and that is why it is better to think of constitution as a living document rather than a fixed tool notebook. As AI systems encounter more complicated situations, the principles can become more detailed and more refined.

[23:58] This also means that when you see a simplified explanation of a court's constitution in a quote or a video or article, you should not read that explanation as the exact wording of Claude's current policy. It's better to understand the underlying philosophy.

[24:11] If you ever need a precise and most recent wording of a particular principle, Anthropos publishes its own documentation, and that should be treated as the authoritative source. So what we're focusing on here is the bigger idea behind the system. And despite individual roles changing or becoming more detailed,

[24:25] the broader objective remains fairly consistent. Claude is intended to be honest. That sounds simple. But honestly, honesty can sometimes mean telling a user something they do not necessarily want to hear.

[24:37] So if a model is uncertain, it should also ideally communicate that uncertainty instead of pretending to know the answer. If a user makes an incorrect assumption, the model should not simply agree because agreement would make the conversation easier.

[24:49] If something cannot be done easily or reliably or safely, the model should be willing to say so. But at the same time, the goal is not to create an assistant that refuses everything remotely complicated. That would not be very useful either.

[25:01] A strong register needs to distinguish between a genuinely harmful request and a reasonable request that simply involves a sensitive subject. So it needs to push back when necessary but do so instructively.

[25:13] In other words, it should not automatically cave to a bad request just because the user insists. But it also should not automatically shut down a legitimate request just because one part of the topic found risky.

[25:25] That middle ground is extremely important. The ideal behaviour is something like this. Understand what the user is actually trying to achieve, identify whether there is a genuine problem, or limit or redirect only the problematic part, and remain as helpful as possible with everything else.

[25:42] There is another principle underneath all of this that is easy to overlook. Chore is generally designed to treat the person on the other side of the conversation as a capable adult. That means the assistant should not unnecessarily lecture the user, talk down to them or assume bad intentions without the evidence.

[25:58] It should explain its reasoning, yearly provide useful alternatives, then appropriate and allow the user to make informed decisions. So when you look at the constitution as a whole, the goal is not simply make floor safer.

[26:10] The deeper goal is to create a model that can balance several things at the same time. being helpful without being leftist being cautious without becoming useless being honest without becoming unnecessarily confrontational and being respectful without blindly agreeing with everything the user says That balance is why the constitution continues to evolve The illusion wording may change as technology changes and new use cases appear

[26:33] the center's philosophy remains the same. Build an AI assistant that can exercise better judgment, communicate its boundaries clearly and still treat the user as someone capable of having a serious and productive conversation.

[26:45] Now let's look at Responsible Scoring Policy and Safety Levels. Separated from the slowly connected to the constitution, Anthropic also runs something called the Responsible Scaling Policy or RSP. These are constitution shapes how an already released model behaves in conversation.

[26:59] The RSP governs a different question entirely. Whether Anthropic should train and release a more capable model at all and what safeguards needs to be verifiably in place before that happens. The framework is built around AI safety levels or ASLs,

[27:13] a ladder of increasing risk and correspondingly increasing required safeguards and they're usually similar to biosafety levels in a research laboratory. The exact technical thresholds that define each level are detailed in anthropic published quality

[27:26] and they get revisited as capabilities evolve. The concept that matters is this. Before anthropic scales of a model's raw capability, it comes publicly in writing to scaling of matching safeguards first and to pausing or restricting a release if evaluation shows a model has crossed into a riskier band

[27:43] without the corresponding protections in place. Let's make this less abstract. Imagine someone asks you all for detailed information on a topic that could plausibly be used for harm, but the request itself is ambiguous.

[27:55] It could easily be a student's homework or it could be something worse. A qualitative trained model tends to swing to one of two bad extremes. Either it answers in a full detail without second thought or it refuses outright with no explanation.

[28:08] A model trained against the consequences is pushed towards a third option, engaging with a legitimate everyday version of the question. while declining a specific narrow slice would provide pure uplift towards harm and leave honesty about why that line is being drawn.

[28:22] That's a consultation doing its actual job in practice. So what does all this actually mean when you're using CLAW? The practical payoff is that this training approach directly shapes the way CLAW responds when a request becomes difficult, sensitive or ambiguous.

[28:35] Instead of thinking about safety as a simple switch, answers or refuse, flood a strain to make a more nuanced judgment about what parts of the request it can help with and what parts may need to be limited. And that is why when God refuses something,

[28:48] a good response will usually try to explain the reason rather than simply stopping with a flat. I can't help with that. That explanation matters. If the model only refuses without a context, the user has no idea what tipped the boundary.

[29:01] So was the entire request a problem? Was it one specific detail? Was the intent unclear? Or was there simply a safer way to achieve the same goal? A more useful response tries to make the distinction here and very often Claude will also try to offer an additional version of what you actually need.

[29:18] For example, imagine you ask for something that contains one problematic element but your larger goal is perfectly reasonable. A richer system might reject the entire request, a more nuanced system might say, in effect, I can't help with that particular part but I can help you accomplish a broader goal in this safer way.

[29:33] That ability to redirect instead of simply refusing is one of the most practical consequences of constitutional lie. It also changes how ambiguous requests are handled. Not every annual question has harmful intent.

[29:45] Sometimes a developer is testing a system, sometimes a student is doing research, and sometimes a writer is just exploring a fictional scenario. Sometimes a security professional is trying to understand the risk so that they can prevent it.

[29:58] So if the model automatically assumes the worst possible interpretation every time, then it would become frustrating very quickly. So the goal is generally to interpret reasonable ambiguity in good faith while still paying

[30:11] attention to what the request would enable. And that creates a very different interaction. Instead of thought behaving like a rule engine that is constantly looking for reasons to block you, the goal is for it to behave more like an assistant that is trying to understand

[30:24] what you actually want while still respecting its bounties. And once you understand that, your words' behavior starts to feel much less arbitrary. You begin to recognize that many of its responses are based on the balance between usefulness, risk, intent, and context.

[30:37] That understanding can also make you much better at prompting. Suppose Claude pushes back on your request. A common reaction is to keep repeating the same prompt, make it more forceful, or try to trick the model into ignoring the concept.

[30:49] Usually, that is not the most effective approach. A better approach is to clarify your legitimate goals. explain the context, describe your audience, describe what you are trying to accomplish, and if necessary, remove the part of the request that is actually creating the problem.

[31:01] For example, instead of saying, give me everything about how this system can be exploited, you might explain, I am reviewing the security of my own application, help me identify common vulnerability categories, and create a defensive testing checklist.

[31:13] Now, the purpose is clear. The model has more context about how your legitimate objective and can often give you a much more useful response. The same sense will apply to outside safety related topics. FLOOR generally performs better when it understands why you need something, who the output is

[31:29] for, what constraints matter and what successful output should look like. So, understand the philosophy behind FLOOR is not just an interesting piece of AI theory. It has a practical effect on how you use the model. So, when you understand that FLOOR is trying to reason about the intent and consequences

[31:45] of the request, you can structure prompts and give it enough information to make better decisions. the model like something you need to fight against. Instead, you start working with the way it reasons. And that is where the dual benefit appears.

[31:57] Clearer context leads to better judgements, better judgements lead to more useful answers, and even when Claude does need to draw a boundary, you are much more likely to get an explanation, a useful alternative or another path towards the results you were actually trying to achieve.

[32:12] That is the practical value of the entire constitutional AI approach. It is not only about teaching a model when to say no. It is about teaching a model how to remain as helpful, transparent and reasonable as possible even when the answer cannot simply be yes.

[32:26] So, now that we have completed this lesson, let's just quickly have a recap of this lesson. Constitutional AI trains flawed against a public evolving set of written principles using a two-stage critic and reinforcement process, instead of relying purely on human

[32:39] labellers. The two stages are supervised self-critic and device, followed by reinforcement yearning from AI feedback which compares response pairs against the constitution. The responsible scaling policy and its AI safety levels govern a separate question,

[32:54] whether it's responsible to release a more capable model at all. This training is why Claude tends to explain refusals, offer adjustment alternatives and assume good faith by default rather than reflexively refusing anything sensitive.

[33:11] So, welcome to this demo. let's now understand each and everything from the basics. So let's basically start from the start and understand how Cloud AI works and how Cloud Code works from the basics. So first, let's

[33:25] understand about the interface of how Cloud desktop is. As soon as you install the Cloud desktop version and you open it, this is how it appears. So this is the homepage. Here you have all your reasons and you can also pin the prompt that you want to quickly refer to it

[33:43] later. Then here you have code version. So, instead of writing the code and executing everything in detail, you can just give a prompt to the cloud and then it will generate everything for you such as websites or anything

[33:56] that you want. So, there is also projects. These are the projects that you have built so far. and then dispatch is something you've done something in your desktop and you want it to

[34:08] be in your phone as well so if you're not available if your laptop or desktop is not available with you you can quickly switch on this pair your phone and it will dump everything you

[34:20] have in your desktop to the cloud in your phone so you can basically refer everything that you've done in the desktop in the phone as well so that is the feature of the dispatch and customize you

[34:32] can customize everything like the skills, connectors, which we'll talk further. You can customize them and rename them accordingly and add them as well. So this is the account

[34:44] details of your cloud and then privacy billing. Basically, I'm using a cloud pro version. If it is a free version, you will only be getting a normal cloud. You will not get any cloud code or cloud co-work. So these are the features of the cloud desktop. So let's

[34:59] So before we start, we all have one question that why would we reach for cloud, specifically if we have tragedy, complexity, any AI with us. So let's just compare tragedy and cloud now.

[35:11] Let's give the same question to each of the AI tools. So, I have written a random question and I am asking this to three of the Airtools, Claude,

[35:32] Chajikati and Jain. And let's compare the difference that we get in the answers. I have asked the question, give me three non-obvious risks of launching a subscription box business in a smaller city.

[35:49] So it has given me three to four points, three points, and then it has given the explanation of why it has given that point as well. So it doesn't scale the way you think they will in big metro or close to major career hubs.

[36:02] So it's given the reason of why it has given that statement as well. So according to the question that I've given, it has given me three points that the fulfillment cost doesn't scale the way we think and the local supplier and packaging relationship.

[36:15] And also it has given us that local market doubles as your only source of word of mouth and your talent pools. So you can exhaust both of them simultaneously. So this is the response of the cloud. Now let's give the same question to CharGPP.

[36:42] So you can see the difference as soon as the Chhajipati gives an answer. It is, the information is less and also has given us very less information compared to Claude. In the smaller city, the visible audience is limited.

[36:56] Once your most interested customers have subscribed, it becomes expensive and existing subscribers make chances after a few boxes because their novelty wears off. So Chhajipati has given us the basic version of whatever the question we asked.

[37:10] It has given the basic answer. that the customer's intake can hit faster than expected and local supplier dependence can create hidden operational risk and the shipping economics may work against the location.

[37:22] But whereas Clod, it has given us the advanced answers, it has also given us the number of spreadsheets of boxes also that can be compared. Suppose, for example, it worked in a spreadsheet of 200 boxes a month,

[37:37] can quietly get worse, not better as we scale. it has also given us the example of whatever we've asked as well. But here it is just the theory part and it's just up to the point what

[37:49] up to the points of question that we asked. So we can also check with Gemini.

[38:09] So, as you can see here in Chajipiti's version of the answer, it is clear and it has given us a clear explanation with actionable points. And in Gemini, it gives multiple angles and also gives us the contextual ideas like once you exhaust 10 to 12 local suppliers, your cost per item spikes because you are forced to source from regional or national distributors.

[38:33] So, it is also giving us the different contextual ideas as well. but some points may feel less focused or very obvious. That is a negative, that is a drawback of Gemini. And in Sagittity, it can sometimes feel predictable

[38:45] and is also generalized here in Sagittity as well. But if you see Claude, it gives the answers in a very structured and it provides more specific and less obvious risks answers as well.

[38:58] So it explains the whole reasoning behind each and every point and it gives more practical insights, which is relevant to launching a subscription box business in a smaller city, which is what I've asked. So now you clearly understood the difference between

[39:11] tragedy, drama and fraud. And let's now go further and give an image to Claude. So I've given an image and I'm asking him

[39:23] what is making this display hard to shop at a glance. As we see this image, we can quickly analyze what is the negative impact of this image. is that it's very clustered and it's not clear

[39:36] and the person cannot identify the item they want as soon as they see this. So let's see if Claude is able to respond to it.

[40:03] So, it has given us the answer. It has given us four points, saying that first is no category zoning, that is the snacks and personal care, pharmacy items, everything is kept at one place, and our eye has to jump

[40:20] between completely unglated product logics. So, instead of moving through one category at a time, we tend to look all at once, and we have no clue what we have to buy. So, that was the first point.

[40:33] and multiple competing hotspots of color intensity. The chitbags, that is yellow and red corporelase, horlicks, they everything are on the top shelf and are of the same color.

[40:46] So they are similarly saturated and similarly sized. So we cannot focus on one particular thing where all of them also have same colors and same size of texture. So with three or four zones fighting for attention at once,

[40:59] our eye doesn't know where to land first. The next is vertical hanging strips break the shelf grid. It is talking about these strips that are vertically hanging. So, work directly behind them and force your eye to sky in an extra dimension on top of the normal left to right top to bottom tree.

[41:16] So, this is also one disadvantage. And next is inconsistent phasing and orientation. This is also true. Some products phase fully forward with labels flat to viewer and others are loosely partially turned.

[41:29] The next is, it's almost no negative space or gap between the product clusters anywhere in the frame. So it also asks the solution of how to solve this problem.

[41:59] So, I have asked it how to solve this issue. It has given us a few steps. Split the wall into clear blocks. That is, we have to split the block for different categories such as

[42:16] personal snacks and pharmacy etc. And the next is put them in a particular, in the respective shelf and not on the other or vertically or horizontally or flanking.

[42:32] We should put it in the exact shelf and then turn angle or loosely stacked package so label face straight out and sit at a consistent height within their row. And also color and saturated placement.

[42:44] Instead of letting red, yellow and pink package compete evenly across the whole wall, group same color or same brand packaging together within each zone. So this creates one strong visual anchor per zone rather than five weak ones fighting for attention, which is what's currently making the whole wall look like a single undifferentiated field.

[43:03] So it's asking us to do the packaging of the same color and same brand category together. So it is visually good and the users tend to see it and buy it often.

[43:18] Next, leave a visible 1 or 2 inch gap instead of putting everything at once as to give small gaps between the zones. And reserve eye level shelves for high in margin of dead selling items.

[43:31] So it is also given as one point here which says that the middle two shelves get the most visual attention. That is so true when you go to shopping and when you look at the shelves, your eyes tend to go to the second and third shelves instead of the completely bottom and the top shelf.

[43:49] So, we have to make sure that the best margin or anything which is in discount or anything should be kept in the second or third row so that the buyer can tend to see that easily and take that item.

[44:02] And keep the top shelf light and low clutter. So, it has suggested that we have to keep the top shelf light and low clutter. So instead of putting all the good items on the top shelf, you have to put the backup boxes or which is not so important to us on the top shelf.

[44:23] Now this is about the image. Let's quickly import a document and see if it can understand that as well.

[44:40] you

[45:10] So, we understood how it responds and how it gives the answers if it's an image.

[45:33] So, let's quickly move into the document. The document that I have is a fictional document which has, I'll share the document. So, we're talking about this data. This is a fake candidate's data, fake numbers, fake IDs, everything.

[45:49] I'm just trying to put it to Claude and trying to get the answers from it if it will recognize the consistent or missing data and will it warn us accordingly.

[46:02] So, I'm giving a prompt in such a way that pull out anything that looks inconsistent,

[46:15] missing or unclear before this goes out to the printer. So, basically what I'm trying to tell here is, in Clod the text, images, documents are

[46:31] all handled inside one single conversation with Claude reasoning about each rather than just mechanically processing them one after the other. So, the strong reasoning and multimodal input which Claude has is one of the best features

[46:44] of this Claude desktop release. Also, it has given us, it has went through the 30 records. So, I wanted to put all of the inconsistent data seeing if the Claude would identify.

[46:58] So, it has given us the row and the issue. Let's quickly analyze it side by side. It has given us the email was missing here in the row 11, sorry customer ID 11.

[47:15] For customer 11 here the email ID was missing, true and for customer 12 the state is wrong If it is Chennai, it should be Tamil Nadu and it's Kerala.

[47:29] And the email is also in cab. And so the date also is different. The date format is different. Date is not different. The format is different. And the sales also is in lowercase, which everywhere it is active is in capital and this is lower.

[47:46] So it has gone through the sheet and given us the exact wrong, exactly where it is lacking. similarly for customer 13, 14, 15

[47:59] in the same way it has given us the answer so we know that it is able to find the wrong statement and given us the answer so here I have given the document directly

[48:11] but what if I tell you that I can also connect the cloud with the database, with the sheets, Google Calendar or anything that I want and then it will directly pull the data from there and give us the answer.

[48:26] Let me show you. If you go to connectors and click on add connectors and browse connector. Here you have a number of applications or websites which you can connect and get the response quickly.

[48:46] So let me connect the Google Drive. So, make sure to click on continue with access, asking us the access if it can get connected.

[49:05] So, yeah it is connected. Now since I connected the Google Drive, let's make few changes here in the sheet and give the same sheet again and see how it will respond.

[49:18] I'm deleting few rows. So let's save this and give the same question.

[49:48] So as you can see it is going through the content log.

[50:05] So, it has rewritten the data and given us which is inconsistent and which is unclear,

[50:29] which is having formatting issues, it has given us everything in clear format. The customer, Levin, it has this particular issue,

[50:41] which has no email, phone number, country code, and the dashes in the date. So, accordingly, it has set everything in a particular format and given us the answer. So, it can identify the mistakes in the text,

[50:54] in text, in images, in documents, in everything. So, this is how Claude basically works. Let's go more further and understand what are different things that are present here like skills, connectors, projects etc. So, before that

[51:12] let's quickly understand if it will give us the formulas also if we ask. If it will calculate the math and give us the formula just with one prompt so write a simple formula

[51:27] or 10% early booking discount back only if the booking is made more than two

[51:57] weeks in advance. So I've given a question. So it gives me a spreadsheet formula and calculates 10% early booking discount but only if the

[52:09] booking is made more than two weeks in advance. So this is the formula and this is how we have to place the particular rows in A2 it should be booking date, B2 weekend date and C2 price. So this checks if the gap between the booking and

[52:24] event is more than 14 years which I've mentioned in the question for everyone of the formula. You can also check it out by making a fake spreadsheet which has the different prices

[52:36] and check if this formula is working with us. So now that we've seen what after this I will put something else so now that we seen what skills and collectors are let put that into a handle and understand how this works

[52:56] Before we move into skills and connectors, let me tell you about the memory of Claude. As you know, memory is something that if I ask about some particular thing today, it will memorize that particular thing and accordingly it will respond to the questions that I ask later on.

[53:14] so let's fill the memory here first go to settings So, we're going to go and manage.

[53:43] And we will manage everything.

[53:55] I'm giving some random prompt saying that I run a small pottery studio and I sell the

[54:13] finished pieces online and also teach weekend classes in person.

[54:25] and I like product descriptions.

[54:43] It's been warm and a little playful. never cause

[54:55] gate of death so I am managing the memory in cloud saying that I am running a pottery studio etc so whenever I ask questions about

[55:09] the marks or pottery it will know that I am working and so and so and it will give us the answer accordingly So, it has managed memory and it also understood that we are working on so and so in a portrait

[55:37] studio. So, I have managed the memory. Now let's go into new chat and without mentioning anything about fortune, we'll just ask a random

[55:49] question. Write a short product listing for a set of hand-glazed ceramic mugs.

[56:13] so I've given in the memory that I like the description I seem to be playful instead of being very formal and corporate so it has gone through that and has given us the statement that

[56:31] each of these months went through the whole song and dance in a very playful and informal way so they are enough for daily use like dishwasher, microwave safe

[56:43] it's special enough that you probably still hand wash them out of sheer affection hold about this so and so and it's plenty of room for your coffee, tea I don't want to say who insists on filling it to the brim with hot chocolate

[56:58] it also like tweets the text in such a way that it matches the instructions that I have given in I can also manage the answers of how

[57:11] Claude will give us like if we go to settings and go to instructions here from now

[57:23] give me the answers and bullet points to every question I ask.

[57:40] So now, let's ask another question.

[57:56] See it has quickly changed it. Before it was in a paragraph. I have not given the clause

[58:08] instruction. When I gave the instruction it quickly changed into bullet points. So you can change into paragraph, bullet points or memory cards or images. Whatever you want

[58:20] you can give the instruction and it will do it accordingly. Now if you don't want it in bullet points, you can quickly go to settings and then delete that instruction. This will give you a normal. It is also understood what

[58:33] the memory had and then it has given the answers in everyday drinking cups, as well as what mugs are suitable for my potty studio. It has given different options like oversized espresso cups, textured or

[58:45] hand-carved mugs, etc. So, this is how Cloud works. So, let's quickly go into the new chat and ask some random is related question

[58:57] which is not related to the pottery studio at all. So let's see if it includes the pottery in that answer as well because it should not.

[59:20] So I have not mentioned any particular cuisine as well. would have given some random dish that is shrimp and salt, pepper and it is paprika sauce, garlic, butter, which does not go into pottery studio.

[59:34] It does not bring anything related to pottery in this. It only remembers that when we ask about such particular questions that are related to pottery. This is how it works.

[59:48] Now, if you see here, you have different options. Fable, opal, sonnet and haiku. Now let's understand what are these models. So first, let's select HYPU and let's enter some random prompt saying that summarize this customer complaint in one sentence.

[1:00:14] so this is one random customer query customer complaint saying that the order was supposed to arrive 3 days

[1:00:26] but it still has not received it so this is basically 5 to 6 lines but generally the customer complaints which are above like 15 to 20 lines they can't sit and read everything so in such cases

[1:00:39] if we go to HICO and just asks us to summarize this customer complaint in one sentence. It has given us the complaint is about three-day delivery delay, a lack of communication regarding the reason,

[1:00:51] and is requesting an immediate status update, confirmed delivery date and resolution. We can quickly analyze and we can understand, we can come to a conclusion also what would be the solution

[1:01:03] for this particular client. The solution would be that we should write a written mail that if there's a mistake from the delivery or any issue related, we have to just go back to the customer

[1:01:16] and send them the mail regarding the warehouse delivery or the delivery agent issue or something. And we can get back to the client. So that is the use of Haiku. If you want very simple, structured,

[1:01:29] and very easy answers and basically summarization of a particular topic, we can quickly go to Haiku and get there. Now if you go to Opus and they'll ask it to draft a full response to this customer

[1:01:52] that addresses the root cause of the delay and offers to concrete resolution options

[1:02:13] So, since we have not given the context about what the issue was and why the delivery was delayed, it has given us the brackets and so and so, like the default text and we have

[1:02:30] to replace it according to our issue. If we know the issue, the client name, the order number, city, everything, we can just give it the information about it and it will give us the whole subject.

[1:02:45] It will draft the mail as well. So, it has given us two responses. First one is explainable and offer choice and the next is add first and then confirm. So, let's see the difference.

[1:02:58] Here, we have to replace with the order number and the cause of delay in the next step. So, thank you for writing and I'm sorry your order did not arrive. So, here is what happened and it has given us the explanation. So that is the first case.

[1:03:11] The next one is you're telling sorry because your order has not reached. And then the cause and the problem, the solution, then that comes next. So basically we have done something about that issue.

[1:03:24] Because it's written like we have already done that the replacement has been dispatched this morning by priority courier, tracking the number with a confirmed delivery date. So we have already done the action and then we are getting back to the client.

[1:03:36] But here, we are just explaining and we are asking if they want the refund, cancellation or the replacement. So, you can choose the mail and the answer accordingly and you can get back to the client.

[1:03:50] So, that is where you use Opus. It's very high and useful complex task. So, before you send the prompt in Clod and if you look at the model picker, just ask yourself whether the particular task, the particular prompt is simple, everyday or hard

[1:04:07] then pick the model accordingly rather than just leaving whatever was last selected and using that by default. So right now I am using Haiku and Sonnet.

[1:04:20] Sonnet is also something which is used for medium level tasks like not so simple and not so hard so you can use it accordingly. So let's quickly know about skills and connectors.

[1:04:35] Here's a situation you have probably lived through. You ask your author something. The answer is good, but not in the format your team uses. So you explain the format. Next week, you try it, you explain it again. That repetition is what skills cause.

[1:04:48] A skill is a folder. Inside it are instructions. And sometimes scripts and the things files. The author loads that folder when a task calls for it. And follows what is inside. Three things make this work. First, you write it in plain markdown. No coding required. You can attach scripts later for something more advanced.

[1:05:05] Second, cloud loads a skill only when it is relevant. So the context window does not fill with instructions you do not need. Third, and this is a payoff and the result is repeatable. Your brand guidelines, your report format, your analysis workflow, same standard every time.

[1:05:20] So here is the thumb rule. If you explain something to cloud more than twice, that is a skill waiting to be written. So, there are four kinds of skills. Skills reach Claude through four routes. One which is which saves you from building something that already exists.

[1:05:34] First, anthropic skill. Build and maintain the anthropic. Document creation for Word, Excel, PowerPoint, and PDFs is here. You do not switch these on. Claude reaches to them when the task fits. Second, custom skills.

[1:05:46] The ones you write. A specialized workflow, your own voice, your own checklist. You upload it from settings and it stays yours. Third, organization provision skills. On team and enterprise plans, an owner pushes a skill out to everybody at once.

[1:05:59] Nobody approves anything and that is how a company standardizes the way work gets done. Fourth, partner skills. These live in a directory and come from companies like Notion, Sigma and Atracian built to work alongside that company's connector.

[1:06:13] One more thing, the skill format is published as an open standard so a skill you write is not locked inside one tool. Now the part people get tangled in. Cloud has three features that all look like giving it instructions.

[1:06:25] The difference comes down to one question. When does information load? Project knowledge is static background. It's just inside a project and it's always there and every chat will start there. A skill is a procedure. It activates only when a task calls for it and it works everywhere in Cloud.

[1:06:40] Custom instructions are broad preferences. Tone, style, how you write things, phrase. They apply across all your conversations all the time. Here is a way to hold it. Project knowledge is a client brief. Custom instructions are how you like to be spoken to.

[1:06:54] A skill is a standard operating procedure for one job. And this that a connector gives Claude access to a tool. A skill teaches Claude how to use that tool the way your team expects. Project knowledge supplies a background for the job.

[1:07:07] So now let's quickly understand what our skills and connectors are. Let's go to new chat. So here if you select, you can see something called a skills.

[1:07:19] So you can go to skills, manage skills and then you can create with cloud, write skill

[1:07:40] instructions and upload a skill as well. So basically what a skill is, it works little differently from connectors. like instead of connecting to live outside data as you can see i've already connected the google

[1:07:52] drive to the connector so instead of connecting to the outside data the skill packages up a specific and repeatable way of doing something so the cloud follows the exact same process every single time

[1:08:04] so instead of you re-explaining the format tone and structure repeatedly from scratch in every prompt skill does it all the time at once so if you've given a particular description and instructions in the skin and if you ask multiple questions hundreds or like multiple questions 50 to 100

[1:08:22] questions at once it will go back to the skin read the instruction description and then give you the answer accordingly instead of you telling it multiple times in every 50 questions suppose for example you want a particular answer in a particular format so you go to claude and in the

[1:08:41] chat you will write to give the answer in so and so point like brief description etc not only brief bullet points sorry bullet points and description not only bullet points and paragraph there are different formats of giving an answer like memory cards in a image format

[1:09:00] visually appealing in so and so color subheading fonts etc everything so if you have 20 to 30 prompts or questions or anything that you want cloud to give the answers to and you want everything

[1:09:14] in a particular format so uh for every question you can't go and give the instructions repeatedly all the time 20-30 times so for that skill is the best use you can give the instructions and

[1:09:28] description and it will generate the answer in that particular way for all the 20-30 questions that you give so that is the main use of the skill I'm adding a description saying that summarize

[1:09:42] this week's class bookings and top selling products in a friendly

[1:09:55] 2 paragraph name Always end with

[1:10:10] next week's class schedule

[1:10:25] So, we have created studio update. Now in my Google Drive, I have a fake Fortress Studio Google Sheet which has different bookings,

[1:10:43] different data basically and I am giving it the still studio update to the cloud and the The summary is also given. So you just have to click enter.

[1:10:55] If it is not reading your content, you can quickly add a prompt saying that go through my Google Drive and then pull up the data and give the answer. But I haven't even given the prompt saying that

[1:11:08] please refer to the Google Drive because I've already connected the particular application to the cloud. It has quickly fetched the information from Google Drive and it has given the answer. So it has given us the Studio Week recap from August 3rd to 9th

[1:11:23] where we have hosted 14 classes with 4 and so sessions. I will quickly show you the Google Sheet. So this is the sheet that I have given in the Google Drive,

[1:11:36] that I have posted in Google Drive. See here you have 14 sessions and 14 customers. Customer name, date, session type. So whatever sessions are done in the pottery studio, like open studio, glazing sessions, date night, private lesson, kids pottery, everything is mentioned here.

[1:11:55] And that in these amount paid, payment status, if it is not paid, pending, refunded, everything, the instructor and the particular customer name and the booking ID. So this is the data that I have put in the Google Drive.

[1:12:07] And Flod has fetched that data. Accordingly, it has given us the answer. So it has not given us the exact information. so I asked it to go through this particular spreadsheet and give me the answer, give me

[1:12:19] the mail accordingly. So, these were the 14 classes which were done, 14 sessions which were done in the portrait studio and just about what has been done is given.

[1:12:35] Next we have connectors. So, everything so far has been about what about flawed nodes. Connectors are about what slot can reach. A connector links slot to an outside tool or data source, your calendar, your files, your issue tracker.

[1:12:48] Underneath, connectors run on the model context protocol, usually shortened to MCT. Until it's built and released as an open standard, the steaming server can serve more than one AI tool. Connectors do three things.

[1:13:00] They reach, search a folder, pull up the week's meetings, find the open tickets, and they act. Create a task, update a record, post a comment in the app you already use. And some of them show you something. Certain connectors render an interactive component right into the conversation so you get a real interface instead of a wall of text.

[1:13:18] This is the whole idea. Cloud stops being a place you paste things into and starts working when your data already lives. So where connectors come from? There are three routes to a connector and they tie different levels of assurance. First, a verified directly connector, built or reviewed against Anthropian directly standards.

[1:13:34] If a verified one exists for the tool you need, start there. Save this option and about a minute to set up. 2. A Community Connector Also in a directory but built by third party developers. Each comes with that developer's own terms and privacy policy, usually shown during the sign in step.

[1:13:49] Read them. This is not the same as an anthropocrybule integration. 3. Build Your Own You point cloud to a remote mcp server you control. This is how teams connect internal systems no directory will ever cover.

[1:14:01] Available on every plan, although a free account is limited to one custom connector. Whichever route you take, access is granted per user through a standard sign-in flow. Connecting an app does not hand over standing access to your account.

[1:14:15] Connecting Response Key Now, the parts that deserve your attention. A connector can be given permission to change things, not only read them, create, update, delete. So, build a few habits early. First, actually read the tool request. It tells you what fraud is about to do.

[1:14:30] Approving individual actions is slower than approving a reading and it is how you avoid a surprise. Second, be careful with the love always option. Save it for servers you genuinely trust to run without you watching.

[1:14:42] Third, know where these run from. Remote MCP servers are released from Anthropix Cloud, not your own machine, even in a desktop app. So if your server is behind a company firewall, the connection simply fails until somebody opens the path.

[1:14:55] Turn off connectors you are not using. Fewer live tools means forever things that can go sideways. go sideways and for larger teams owners can authorize connectors centrally so staff inherit access on first login so if you just go to connectors add

[1:15:12] connectors let's say you have a presentation or you have to submit a PPT tomorrow and then you have no time you can just go to cloud and connect canva to it

[1:15:30] So as you can see the canva is connected.

[1:15:48] So let's go to new chat and then connect the canva.

[1:16:00] and then create seven slides of PPT six to seven slides precisely on the impact

[1:16:15] of AI on digital marketing this is some random topic that I'm giving you can give your topic and you can also customize and give instructions

[1:16:27] accordingly or how you want the ppt and then telling it what to cover and how the ppt should look cover ai powered content creation

[1:16:39] customer insights automation

[1:16:57] Thank you.

[1:17:27] given a prompt to create six to seven slides of PPT on so and so topic and cover these topics, AI power, colonization, connotation, customer insight, automation, real world examples

[1:17:40] and to use modern visuals and how to prepare the PPT basically. To use visuals, icons, minimal text and it should be visually engaging and presentation ready. So, let's see.

[1:17:57] So as you can see it has created the PPP impact of AI on digital marketing.

[1:18:26] also customize those particular PPT you can give changes that

[1:18:44] so it will again redesign the particular PPT and give us the particular insights you can customize it accordingly you can give it that you can give a proper instruction saying that you need the icons as well and particular images if

[1:18:59] there are any and make it UIUX friendly like use so and so particular wireframes or prototypes or so and so typography colors use this particular color when this color appears and use different images icons which are important for this particular topic

[1:19:21] You can play with it accordingly and generate the PPP that how you want. I'm just showing you how the skills and connectors work and how you can make your work easier. You can connect it to Canva, Slack or there are many applications that you can connect

[1:19:39] lot though so like google calendar as well so if you have any particular meetings or anything it will give you so suppose if i have to schedule a pottery studio meeting today at 10 o'clock and i

[1:19:57] don't know if my calendar is free so i can ask today i can know if my calendar is free but if i want to know after one week or something and my calendar is already booked like if i'm already

[1:20:09] having any meetings later after one week or something I can just give the I can just connect my Google Calendar to cloud and ask it to generate the timing whenever I'm free and schedule the meeting accordingly so it acts as an

[1:20:24] assistant basically so instead of asking a particular assistant you can ask cloud to generate particular meeting for so-and-so timing when I'm not booked the one I'm free basically. So it will go through your calendar and it will identify

[1:20:39] when and at what time you have meeting and whenever you're free, it will give you options if it is weekend or any time and it will ask you if it can schedule, it can also schedule. But you should be careful as well when you're connecting different applications like this.

[1:20:54] So yeah, you can connect to Gmail, Slack, Notion. You can also connect Gmail and ask particular questions like what is my recent email and what was the summary about or these are just

[1:21:08] the basic things actually you can go through like you can give the email and ask different questions like what is the summary of this particular mail on so and so date and then it will set the

[1:21:22] information and give you the answers as well so there's a zig field google drive calendar github canva github you can connect and ask about some repository and the code as well you can connect

[1:21:35] it to notion there are many apps that you can connect it to so you can make good use of this cloud so it has given us a ppt so it has given us this before as you can see this

[1:21:53] heading was overlapping the information and we cannot read the information. So it has given us this. I only have 6 to 7 slides.

[1:22:05] You can also tweak it a little bit and yeah, that's how the cloud works. Basically what I'm trying to tell you is instead of you doing things manually from the start,

[1:22:17] do smart work and connect everything to cloud and you can finish all your work in 3-4 days which would take 5-6 or 1 week at least. So that's how the connectors and cloud skills work.

[1:22:34] So let us move the output. Some things cloud produces are answers. You read them and move on. Other things are work. A document, a script, a prototype. Those should not be buried in a transcript.

[1:22:46] And that is what an anticast is. When cloud produces something substantial and self-contained, it opens in its own window beside the conversation. The trigger is roughly this. More than 15 lines. It stands on its own without the surrounding chat.

[1:22:59] And there's something you will refuse or come back to. What counts? Markdowns, documents, diagrams, single-page websites, vector images, code snippets, interactive components you can take. The important part is what happens next.

[1:23:12] An anticraft is built to be iterated on. Ask for a change and it updates in place. You can keep several open in one conversation and tell Claude which one you need. One setup note. Artifacts need code execution and file creation switch on settings.

[1:23:27] Everything you need sits in the artifact window. And most people never find half of it. Start with everything in place. In a markdown artifact, highlight the exact text you want changed. Click edit with Claude and describe the things right there. No explaining which paragraph you mean.

[1:23:41] This tip alone will save you more time than anything else on the site. Next, versions. Every revision is kept. Use the version selected to compare drafts or rollback. Nothing is lost when you experiment.

[1:23:53] Then, viewing and exporting. Open the code behind any article, copy it, download the file. Copying code is also how you build on an article somebody else published. And finally, publishing. It adds the article to your sidebar and lets other people open it.

[1:24:09] If they have a cloud account, they can remix it which makes their own copy without touching yours. If something breaks, look for the Try fixing with Claude button near the error. Here is where Artifacts stop being documents and start being small applications.

[1:24:21] These have abilities to do that. First, an Artifact can call Claude itself. So, you can build something that answers questions, coaches somebody through appraises or adapts to what the user types. When you share it, users count against each person's own subscription, not yours.

[1:24:35] 10 people or 10,000, it costs you nothing extra. Second, Artifacts can read your connectors. So an artifact can read from and drive to the tools you already use. A small internal dashboard becomes a real possibility.

[1:24:49] Third, storage. A published artifact can remember data between the sessions. That is what makes a tracker or a leaderboard actually work. Storage is either personal where each user sees only their own data or shared where everybody sees the same thing.

[1:25:03] Two cautions. With shared storage, other users can see what you enter and unpublishing an artifact permanently delays everything it has stored. Now, let's quickly go to Articraft and understand what it is. So, Articraft

[1:25:15] is basically, it's a feature in Clog. It is used to create usable and interactive outputs instead of just giving you a text format. So, go to Articraft.

[1:25:30] So, you can just give a prompt build near interactive personal interactive

[1:25:43] expense tracker articraft. So instead of telling you how something works, it can actually build a theme for you inside the

[1:25:57] articraft window. So it will basically include all the interactive webpages or mini websites or dashboards, data visualization, calculators, tools, games, documents, things.

[1:26:11] You have just given not even a prompt. You have just given a line, build me interactive personal expense tracker. And it has given us particular, a basic dashboard basically.

[1:26:28] of let's say if I have $600 and transport

[1:26:42] holiday you can also add Okay, the amount that I have added already has been updated and it is showing us the

[1:27:02] total spent and number of transactions won. So if I add some other listing, food, breakfast, so it has changed accordingly.

[1:27:20] So here also you can tweak it a little bit, tweak the prompt and tell it to add little icons or build like with so and so colors, typography, use so and so font and you can

[1:27:34] build the dashboard accordingly. So this is a basic expense tracker. So now that you understood what, how cloud basically is and what are skills, connectors, RTCrafts and what are the differences between CGPT, cloud, Gemini, let's quickly build a

[1:27:50] project. So, welcome back. So now let's look at projects. A project is a self-contained workspace with its own chart history and its own knowledge base.

[1:28:03] Everything about one piece of work lives in one piece. Why that matters is simple. Without a project, every conversation starts from zero. You're explaining the client. You're uploading the brief every time. Inside a project, three things change.

[1:28:16] Your charts stay together in one list not scattered across weeks of history. Your knowledge is uploaded once. Documents, notes, transcripts, brand assets, cloud draws on all of it in any chart inside that project.

[1:28:29] And your instructions persist. Set the tone, the role or your own one and they apply everywhere in that space. Think of it as onboarding. It is a document you hand a new team member on day one.

[1:28:41] So, you stop repeating yourself. Projects are every plan. Free accounts can hold up to five. There is an obvious question here. What happens when a project gets big? Every model has a context window, a limit on how much it holds at once.

[1:28:53] So surely you hate a ceiling. You do. Here is how Cloud handles it. Step 1. You upload. Docs, notes, transcripts, whatever the work depends on. Step 2. Cloud loads it as a background, available in every chart in that project.

[1:29:07] Step 3. As your knowledge approaches the limit, retrieval takes over. Instead of holding everything at once, Cloud pulls that relevant part when they are needed. This is what people mean by rack or receivable augmented generation.

[1:29:20] It expands capacity by up to 10 times while keeping answers quality. The best part is that you do nothing. There is no budget. The project scales and you keep working. One note is receivers for projects is done on a paid plan.

[1:29:34] Pro, max, team and enterprise. On team and enterprise plans, a project stops being personal. It becomes a place a team works. There are two permission levels and the distinction matters. Can you? Members give the knowledge and instructions and chat inside the project.

[1:29:49] What you cannot do is change it. This is the right setting for most people. Can edit. Members update instructions and knowledge and manage who else has the access. Give this to the people who own the work.

[1:30:01] Sharing is flexible. Add name to people one at a time, add a list of email addresses at once or open a project for your entire organization. On the receiving end, a shared with me tab connects the projects other people brought you into. One control for administrators and owners can switch organizations,

[1:30:16] wide sharing off completely, and the option then disappears for everybody. Let us bring all the four together. Project holds the context. The background this work depends on. Connectors teach the systems where the real data already lives.

[1:30:28] Skills carry the procedures. The job gets done your way every time. Artifacts are the output. The thing you keep, iterate on, and hand to somebody else. In one sentence, you work inside a project, Claude pulls what it needs through a connector,

[1:30:41] A skill tells it how your team does the job and what comes out is an article you can publish. So now let's go and build it. Let's go to cloud code and go to new. So here you can go to articles as well and you can create new articles.

[1:30:57] What I've shown you before. Let me just tell you how the cloud code works, you just have to instead of writing the whole code, you will give the prompt to

[1:31:12] the cloud and it will generate the thing for you, generate the project or website anything for you, HTML code, all the sessions basically. So I have created a folder saying cloud code demo just have to click on local and then bring that particular folder here

[1:31:35] you will get this folder so we are creating a folder because whatever html files or json files that it will create it will store in that particular folder and we can also change those files

[1:31:47] accordingly according to our wish so it will store in that folder and we can go back to that and look at it. I'm using the same this thing like the personal expense dashboard only I'm building so I'm giving a

[1:32:03] prompt I want you to build a personal expense

[1:32:17] dashboard from scratch in this folder I'm giving it a description as an instruction use only HTML, CSS and

[1:32:34] vanilla JavaScript Do not use any frameworks or external libraries.

[1:32:52] read index dot html

[1:33:07] style dot CSS script dot chasis I'm personalizing my

[1:33:20] dashboard I am adding what all features I want in my dashboard

[1:33:35] so total expenses like the basic and important features in a particular dashboard monthly budget remaining budget

[1:33:51] number of transactions a list of recent transactions

[1:34:05] Let me also add the category breakdown So I have also added the category breakdown

[1:34:17] So we have food, travel, shopping, girls entertainment and transaction form, search category, whatever features I want. So it has

[1:34:30] created the expense dashboard. As you can see I have asked it to put these features on the top total expenses, monthly budget, remaining budget and the number of

[1:34:42] transactions and the category breakdown so according like on which category how amount of transaction is there I mean is spent so this is the category big

[1:34:54] breakdown and these are the recent transactions like this is a basic the fake transaction of how much every category is taking so if it is the

[1:35:06] internet bill it's like $65 and the movie ticket is $28 travel to $10, dinner, running shows, like the whatever I've spent is given the

[1:35:18] description category, date and the amount. We can also add the transactions for example let's say shopping amount spent $300, category shopping and today's

[1:35:37] transaction and it will quickly be added here that is $300 it will be minus and it will be updated in the total expenses as well so this is the monthly budget you can also tweak and change the monthly budget according to our wish and

[1:35:52] the remaining budget but this is a very basic expense dashboard let's give you instructions few more instructions saying that add a budget warning system

[1:36:06] if spending reaches 80% of the monthly income and the monthly budget

[1:36:21] then show a warning. If spending exceeds the budget show a stronger alert

[1:36:39] and a visual progress bar for budget usage and make sure

[1:36:52] everything updates instantly when I add or delete a transaction.

[1:37:12] I can also ask it to not redesign. If it redesigns from this item,

[1:37:24] you can add just a prompt saying that do not redesign the application, application just modify the existing and verify the new feature works so let me tell you what actually happens when you assign or when you give the

[1:37:37] project etiquette so let me just quickly tell you what happens when you ask the cloud to develop so and so particular project so you give a prompt and the cloud understands the requirement

[1:37:49] first and it creates multiple files so let me show you the files that it has created this is the cloud folder see it has given us the JavaScript, the style, index file like the HTML index. This is the HTML file. You can quickly

[1:38:06] copy paste in terminal and also you will get the same dashboard basically. So it has generated multiple files in your folder, existing folder. So it will generate multiple files. It will preview

[1:38:19] the app that it has made and then it will do the requirement changes and again you will give it the prompt like the modifications and all it will again modify the existing code it tests its own application first so when you go here it will first run different

[1:38:38] whatever you have given a changes it will run by itself and then it will see if everything is going well and then it will give you the dashboard for you so it tests on its own application and then it finds the issues and fixes it and

[1:38:52] And then it is ready for us. It shows us the exact distance. Budget is 45%. So if we add more budget, let's say groceries.

[1:39:11] Let's add and choose budget. Let's say if I bought a car. $500. the budget is 65% let's add more

[1:39:36] fold let's see if it will give us a warning sign if it passes 85%

[1:39:52] let's also see let we can change this actually if it is like fifteen hundred dollars then it will give us the month see you've exceeded your monthly budget by three hundred and thirty five

[1:40:07] dollars so you can put the the remaining budget is also like in minus like minus three to five dollars so according to your monthly budget it will give you a warning sign as well so however you want

[1:40:22] your dashboard you can give the changes accordingly and it will generate it for you without you like no if you don't know coding also it's fine you just have to know the prompting data of how

[1:40:35] to give a prompt and what all features you want explain it to the cloud properly and it will generate a beautiful amazing website and dashboard for you so uh without us entering a single line

[1:40:48] of code we got a dashboard so yeah you can create different dashboards let's say uh for example

[1:41:00] if you ask it to create a dashboard for fitness, include number of steps,

[1:41:18] fold precautions,

[1:41:35] So this is the another project where I have asked the CLODS to create a dashboard for

[1:41:50] fitness including number of steps, the food that needs to be taken and to prepare it in a healthy way. So this is the number of steps.

[1:42:02] we can also add the steps here and the weekdays, the food that needs to be taken for breakfast, lunch, dinner and the healthy options for snacks and the protein, carbohydrates, fats

[1:42:17] that needs to be taken on a whole and if you have, this is the number of glasses that you need, glasses of water that you need to drink daily. So, you can keep this as a tracker

[1:42:29] whenever you have this much amount of water you can quickly come to this tracker and fill this and finish this you can customize it accordingly I

[1:42:44] have given my instructions saying that I need the number of steps the food in a healthier way and healthier options so it is also given us the few warnings and

[1:42:57] and precautions that needs to be taken for walking and the food when that needs to be taken and how much amount of food and how much amount of walk does your body need.

[1:43:09] You can also customize it according to your weight and height as well. So it has given us the, before we start, we need to have 5 to 10 minutes of stretching or warm up and then we have to start walking.

[1:43:21] and we have to drink water every 20 minutes and keep our posture up tight and cool down and stretch the muscles. And these are the warning signs. If you do everything at once, it will not work

[1:43:35] and you will have these chest pains, tightness or pressure and the breathing related issues and fainting, irregular heartbeat, etc. So you have to start from the scratch and if you want to start from the scratch,

[1:43:48] you have to follow these steps, precautions and you have to go day by day, you have to improve yourself. So this is how the wellness dashboards look like. You can also customize the title, the whole structure of how the dashboard looks as well.

[1:44:03] So this is about the project skills connectors in Flot. So now let's understand what all we have learned in this course. There are a few important things to remember from this.

[1:44:16] First, choosing the right Flot model matters because different tasks require different levels of speed, course and reasoning. True context is equally important since CLOAD performs better when it understands your goals, audience and constraints.

[1:44:28] Features like memory and knowledge help you keep ongoing work organized while vision and artifacts allow you to work with visual inputs and turn ideas into reusable outputs. Skills and connectors make repeated tasks more consistent and make

[1:44:40] CLOAD work with relevant external information. And finally, CLOAD becomes much more powerful when you move beyond simple combinations and start using it for multi-step tasks, coding workflows and automation. And that brings us to the end of this course on Cloth AI. Over these

[1:44:55] four lessons, we have built a complete framework for understanding and using Cloth. Starting with choosing the right model for a task, moving into how Cloth is actually trained to be safe and genuinely helpful, going hands-on with the features that makes it useful for

[1:45:09] ongoing daily work, and going as deep as the terminal itself with Cloth code. The real value of this course isn't in memorizing any single feature in isolation. It's in because in the internet these pieces work together. The same underlying model runs throughout the all four lessons.

[1:45:22] What changed is the service that connected tools and cartels built around each other one for a specific kind of work. Whatever really worked and the ability to pick the right tool at the right depth for the tasks in front of you is what actually surpassed casual use

[1:45:34] from generally the best of use. Thank you for watching and I will see you in the next one.