Gemini 3.5 Flash: How to better use agentic AI?

How to best use Gemini 3.5 Pro? That's exactly what this video is about, because things have changed. Until now, we had unlimited plans. We could work all the time, whenever we wanted, with Gemini, but things have changed, and one subscription after another, and yet another, is starting to put a heavy strain on the budgets of both employees and companies. We have to reinvent businesses, rewire our brains to make them work in tandem with artificial intelligence. And now that we've all gotten used to working almost without limits, suddenly, we find ourselves blocked, blocked for a week before we can continue our work. And yet, I have a Pro plan. So I've found solutions to unlock additional credits, and we're going to talk about how to optimize the use of AI through the prompt, the plans, how to use it, and that's exactly what we're going to talk about in this video: how to best use Gemini 3.5 Flash. The arrival of Gemini 3.5 Flash marks Google's entry into the agentic ecosystem. We now have Antigravity 2.0, the Antigravity CLI, the Antigravity IDE, and the SDK. The key difference is that Gemini 3.5 Flash isn't simply an LLM that answers questions; it's an engine that makes decisions and uses tools to get work done. This changes the dynamics of token consumption. The entire market is now geared towards agentic systems, towards AI agents. You are becoming managers, managers of AI agents. So, everyone will be in charge of a team. It simply won't be a team of humans; it will be a team of AI assistants. Should we be afraid of all this? No, not at all, it's nothing to be afraid of. Will we be competing with machines? No, I don't think so either. The position we'll be taking with regard to the machines is that we'll be their boss. So the attitude everyone will have is that when you work with all these agents, it's as if you have a team of young experts working with you. They're not perfect, you have to make corrections from time to time, but it's as if you were the head of a team. This is exactly what's happening in the world of work and in the business world: you 're going to be increasingly faced with managing virtual employees to whom you'll have to give orders. This is exactly what will happen with the arrival of Genspark. But today, it's only available to Ultra accounts, the most expensive accounts. Why? Because these systems are very expensive in tokens. Let's talk about subscriptions. Most people are simply familiar with the standard subscription, Plus, Google AI Pro, and Google Ultra. The main change is that the Ultra system now comes in two versions. There's an Ultra plan that offers limits 10 times higher than the Pro plan for €99 per month, and there's still the €219 per month plan that offers limits 20 times higher than AI Pro. But that's not all. To use these new AI systems, you also need what are called image and video generation credits. All these plans combined start to add up to a significant sum, around €220 per year. That's where I had to find alternative solutions. And by the way, if you want, you can do what some others have done: go to this website and get a year's access for the price of two and a half months. And that's exactly what the Ultra system allows you to do, the Ultra model at €219 with 25,000 credits, with 20 times more tokens than the AI ​​Pro system with that interface. I'm telling you about it because I got one and it really helped me make progress in my work. And that's the The first point that allows you to save money is this: when you have to work and juggle multiple subscriptions with Claude, ChatGPT, and Gemini, you quickly realize you've become completely dependent on artificial intelligence. Beyond the Google AI Pro and Ultra plans, there are also enterprise plans, which are less well-known at the moment. There's a 14-day trial offer with a 50% discount on Workspace for the first three months, which brings you down to roughly twenty euros per month for an enterprise plan with access to the Gemini ecosystem. It's the equivalent of having an enterprise version of Gemini Pro. This means that, first and foremost, you get video conferencing, email addresses, and a few additional tools geared more towards businesses than consumers. Some of you will therefore also benefit from looking at this type of plan, especially since you currently have a 14-day trial. If you enjoyed this video, please feel free to support it by leaving a helpful comment, a like, or a boost—it would be greatly appreciated. Let's continue the video together. With the Google ecosystem, we have the arrival of Antigravity 2.0, CLI, IDE, and SDK. There are so many that it's hard to know which ones to choose. But what I can tell you today is that I'm going to show you how to install the CLI version. For the CLI system, you actually need seven command lines. If you're on Windows, you take the Windows command line, copy it, and that's where I'm going to show you an absolutely brilliant trick. I was working, coding with the Gemini 3.5 Flash interface, and suddenly, I received the dreaded message: "You have no more credits, you have used up all your credits. See you next week." So, work doesn't wait. You've probably gathered that Google Gemini 3.5 Flash has one goal: to get you to buy tokens. To make a quick comparison, Gemini 3.5 Flash costs $1.50 to buy and $9 per million to sell. Overall, I can tell you that the price is very reasonable, especially compared to ChatGPT 5.5, which costs $5 to buy and $30 to sell, or Claude Opus, at $7.50 to buy and $37.50 to sell. And that's without even considering the arrival of an outsider, MiniMax 3, which is currently on sale until the end of the week. With a 50% discount on the listed token price—that is, 25 cents to buy and $1.50 to sell—it's practically equivalent to 3.5 Flash. But for those who want to work with Gemini, as you've probably guessed, Google is pushing you to use tokens. So, what's the value of this Gemini 3.5 Flash system? How do you use it? And how can you optimize it a little more if you get stuck working on the Antigravity CLI? I'll give you a tip. For those of you who want to install the Antigravity CLI, I'll show you how. Go to VS Code, to the terminal section at the bottom, click on the terminal icon, and enter the line of code that corresponds to your interface. In my case, it will be PowerShell. Press Enter and let the process install. Once the interface is installed, simply type the word "agi," press Enter on your keyboard, and you 'll launch your Antigravity CLI. To see all the commands, press a slash, and you'll see a list of them. What you'll be interested in is the "config" command, which allows you to enable the model's reasoning verbosity. You can set it to "high" or "medium," and you can also grant it "always proceed" permission. This means it will automatically launch the tools without constantly interrupting you. If you need to change models, type "model" with a forward slash, press Enter, and choose the desired model. model. You'll notice that you have access to Claude Sonnet 4.6 and Claude Opus 4.6, but I've often found myself unable to access these models due to a lack of availability due to the high volume of requests. Generally, it's quite difficult to work with Sonnet 4.6 or Opus 4.6 in Claude 4.8 CLI. So, how do you resolve this situation, which happened to me: I was working and suddenly ran out of credits. I'll show you how I did it: I simply typed the command "logout". With "logout", I logged out, and the tool will ask you to authenticate with a Google account. It's quite common to be able to use secondary accounts. So, simply use a secondary account to initiate the authentication process. A window will open; copy the code, paste it into the interface, and it will automatically prompt you to connect to all your Google accounts. That's how you'll retrieve the login code. Paste it below, press Enter, and you're done. You now have access to another account. The unique feature of Claude Antigravity is that you can store everything you've done locally, and you can retrieve all your conversations. I'll show you how. Here, you can see that you have free search queries again (you get about twenty a day). This is always useful for finishing any work in progress. To resume a previous conversation, click "resume," search for the conversation, press Enter to retrieve a discussion from another session, and then continue working. That's exactly how I was able to finish this part of the coding. For those of you taking the Level 3 course, I 've coded what's called a "status line" (I'll explain everything in the course). It's a context monitoring bar. It shows you how many queries you have left running, how many tokens you've used, and which model you're currently using. This allows you to track things more precisely. It's something you can easily do with Claude 4.8, a little less so with Codex, and it's quite complicated to implement with Gemini. I've already coded everything for you. So I'm switching systems, and automatically, you know which model you're using, the remaining context, in short, you're able to track and precisely monitor your consumption, and that's really the key point. Today, you need to optimize how you work with AI, and to do that, you need to know how to use it. That's why we're going to talk about prompt engineering and the capabilities of this model now. I ran some tests to compare the evolution of Gemini 3.5 Flash with Gemini 1.5, and I was pleasantly surprised. But I'll also reveal the shortcomings of Gemini 3.5 Flash. First test: I asked it to code a game, and simply comparing the visual rendering between the Gemini 1.5 (on the right) and the Gemini 3.5 Flash already shows the significant technical leap. I'm guiding it simply with my mouse: it's incredibly smooth and very pleasant. Plus, there's sound; it's very playable and really well rendered. There's been a significant improvement in terms of gaming performance. You have the option to purchase upgrades for your system, something you don't have on the other, much more basic model. It's clear that there's been a very significant improvement in the coding aspect. What's really cool is that both worked perfectly on the first try; it really felt like playing in a small arcade. But you also have to I'll show you the limitations of Gemini 3.5 Flash. I asked Gemini 3.5 Flash to create an infographic for me that showcased Transformers models, the neural mechanisms of AI. And that's when I noticed two fairly recurring problems. First, the image has absolutely nothing to do with the subject. It mixed concepts of electrical coils with Transformers models. If we make the exact same request to ChatGPT: "generate an infographic," providing it with a very basic prompt (my goal being to see if it can manage autonomously, understand the content and context to generate work without needing endless explanations), let's look at the result ChatGPT gives us. I can already show you what happened when I had Gemini Flash 3.5 create infographics. This is the output I got from Gemini 3.5 Flash with Imagen 3. But I must tell you that what Gemini, Imagen 3, and its understanding of the context produced is truly incomparable to what ChatGPT does. For me, there's a clear difference in quality, both in terms of representation, graphics, and clarity, in favor of ChatGPT compared to what Gemini 3.5 Flash with Imagen 3 provided. Now we'll compare how ChatGPT 5.5 generates the infographic: assuming it correctly understands that there are different blocks related to artificial intelligence, Transformer architecture, and pre-training. Here, we have exactly the topic we're discussing, and, to be honest, ChatGPT seems much more efficient. In my experience, it seems that Gemini's neural engine was trained to be very fast, but as a result, it sometimes makes errors in understanding and association, leading to hallucinations or misinterpretations. This should raise a red flag, because Google has clearly made a choice: to switch the vast majority of users to Gemini 3.5 Flash and reserve Gemini 3.5 Pro for Pro and Ultra plans (the €99 or €219 plans that you can sometimes use). Visually, they've made an effort, and I'll explain how to achieve this type of result in terms of graphics and animation. But it's a model whose limitations must be monitored: it becomes saturated very quickly. As soon as you send it complex context to process, because the neural engine has to make choices, it makes mistakes, resulting in the risk of hallucinations. This lack of clarity is also reflected in the rendering under certain conditions. Be aware, it's very good in certain situations, but I had trouble getting it to launch the Canva integration during my work. Let me know in the comments if you encounter the same problem: I'm asking it very clearly to launch Canva, to select the Canva option, and I have to write it quite explicitly ("Open Canva" at startup or at the end of the prompt) for the template to activate the Canva interface. On that note, I'm going to explain how to create those very successful layouts with Gemini 3.5, because there are many animations in the graphic rendering, and that's the topic we'll cover next: how to prompt effectively with Gemini 3.5 Flash. To optimize both the cost of tokens and the overall cost to the company of using AI, you need to change how you use it. I asked Gemini 3.5 Flash to give me a lesson on prompt optimization based on Google's official documentation. I'm showing you this presentation because, graphically, I find it very polished, dynamic, and it offers interesting animations and visual effects compared to other AIs. To achieve this type of result, I 'm going to compare it with what was possible with ChatGPT 5.5. It These are the same requests. On the left, we have Gemini 3.5 Flash, and on the right, we have ChatGPT 5.5. The difference, as you're starting to see, lies in the animation. ChatGPT 5.5 is more streamlined, cleaner, more readable, and even more graphical, so you can create something more polished. But you always have to consider the cost per million tokens. The question to ask yourself is: is what I'm going to ask it for sufficiently well-structured, clean, and clear for Gemini 3.5 Flash to be usable? It's four times cheaper than ChatGPT 5.5 (which is a cost advantage) and five times cheaper than Claude Opus 4.6 or 4.8. But if you exceed the model's capabilities, it might malfunction, and you won't be able to trust it anymore. How do you prompt effectively with Gemini 3.5 Flash? This system has changed. We've moved to an agentic system. We're no longer simply dealing with word completion. Therefore, you need to transform your prompt request into a generation system that is controllable, testable, and reproducible. To do this, whereas you used to write very generic sentences (which are clear to you, but represent an infinite number of probabilities for an AI), for example: "analyze customer requests and tell me if they are urgent," it's very difficult for an AI to understand what is urgent and what isn't. To optimize this type of prompt, you have to understand that AI is not intelligence in the human sense; it's not capable of guessing logic: it only has its learning. Humans clicked for days to tell it what is expected of it, and when it doesn't understand, it invents. It also has a certain variability (it wo n't respond in exactly the same way), hence the difficulty of framing the system. This is why a prompt isn't simply a work instruction, but a system loop that defines: the context to use, the tools to employ, and the criteria to determine whether the work has been executed correctly. Otherwise, it must restart the cycle, which constitutes an iterative control process. To begin prompting effectively, you need to define blocks of instructions using delimiters. In systems like Gemini or Claude, XML-like structures are widely used. Sections are created, and then rules are written: "as input, here is the data you have." The constraints, behavioral rules, and output format are defined. Simply telling it "don't invent," is pointless, because AI, by its very nature, only produces token probabilities. What's needed is to implement logging systems. As an AI architect, you need to read and understand the AI's logs: how it worked, how it reasoned, did it make the right decision? Did the provided prompt help it understand the appropriate decision in a given circumstance? This is where the prompt structure completely changes the interaction with AI. And this is how you need to interact with AI agent systems. Everything coming onto the market is agentic systems, like Gemini Spark, the personal agent that will work 24/7. If you're still using old prompt methods, now is the time to change how you interact with AI. The big mistake we make today is believing that because AI is free and readily available, we're smarter. Access doesn't make use. You have to be smart to use intelligence. This will allow you to better define the model, create criteria for its operation and enable it to analyze and audit itself. AI is capable of comparing the The user's request, along with what they've done, prompts them to think, "This is n't compliant; I need to start over." And you need the skill to read the logs and understand how the model works. Up until now, what you've often seen on social media is what's called "voodoo prompts." Gemini 3.5 Flash itself explains it: voodoo prompts consist of writing general, subjective instructions, telling the AI, "Be an expert, explain clearly, think step by step." I think you've heard this "think step by step" advice dozens of times. Gemini 3.5 Flash tells us: "This might work on occasion, but it doesn't create stable behavior, which is a problem in production. It results in vague sentences that give the model too much leeway, executions vary, and the level of detail or assumptions are too implicit. If the AI ​​doesn't understand, the rendering will be unclear." An example: "Summarize this report professionally." A structured way to say it: "Summarize this report in a maximum of five points by selecting the risks (defined according to these criteria) and the decisions to be made (according to financial and legal criteria)," specifying the tone, format, length, and target audience. This helps stabilize the results with objective criteria. And if we have to design a structured instruction, the goal isn't to communicate better with the model, but to transform human intent into an input contract where you give the AI ​​a set of operating rules and parameters so it understands where each block is located. That's why we'll transform: "You're a React expert, make a good, clean, and professional TypeScript hook" into something much more structured by defining delimiter zones containing the instruction: "generates a strict React TS hook, uses `console.log` and `useEffect` on the data output, handles errors and resets, and closes the block of instructions." This is a professional approach to developing prompts for AI agents, and it's how all AI agents are built today. For a long time, the saying was, "Give Gemini or ChatGPT a role." Gemini 3.5 Flash now tells us something: "You're a React expert." This doesn't install a skill into the model. It only changes the context that influences the distribution of the next token. Keep in mind that the model doesn't have a personality; it only calculates a probability: the probability of the next token based on the previous token, attention blocks, and the context. The answer depends on everything you've said before and the constraints provided. If you tell it "you're an expert," Gemini 3.5 Flash explains that this is too broad and ambiguous: "an expert" could be understood as an expert in blog tutorials, components, Stack Overflow opinions, or official documentation. The term "expert" is meaningless for the purpose of AI, and studies have amply demonstrated this. It's part of the "prompt voodoo," which helped those who didn't know how to structure a prompt system, but it should no longer be used with agentic systems because it doesn't answer the three essential questions of an agent system: What is the context of the data to be extracted? What tool should it use to do so? What is the decision-making pattern of the iterative loops? These are the three points that need to be answered. Why are we so precise about the structures? Because models like Gemini 3.5 Flash are based on what is called attention. Imagine the eyes of the AI, which can look at some sentences, but not all at the same time. Often, we send it hundreds of pages of documents because we We are told that the model has a million contexts, but a model's attention span is limited; we can't allocate an infinite number of attention blocks. Therefore, the model must make choices. In the Voodoo prompt where you tell it "you're an expert," the model chooses several probabilities. In a model where you define deterministic behavior, the model knows exactly which bases it needs to retrieve in terms of token probabilities. Remember, you are dealing with a machine that calculates probabilities based on semantic context. The goal of AI is to reduce entropy by eliminating ambiguous paths. To be good at prompt engineering, you don't just need to know how to structure your prompt into sequences; you also need to know the precise semantics to control the model's determinism. Our relationship with language is not probabilistic. For example, if you ask "create an intermediate-level course," the model has to guess what an intermediate level is (for students, experts, or managers? It's not the same thing). In contrast, in a more deterministic system, precise semantic terms are defined, including words like "weight," "neuron," "neural layer," and "Generative Pre-trained Transformer." This forces the system to target the right vectors and create equation analogies that will allow the model to better choose its words. This is the principle of garbage in, garbage out: what you put in is what you get out. This is how LLMs work, and this is where the entire logic of agentic systems unfolds. Now that you understand that Gemini 3.5 is an agentic system, I can tell you that I've tested it as a complete AI agent, capable of searching for information on Google, communicating on Slack, leveraging a RAG database, and performing internet searches. The agentic system understands that it has tools to use, makes tool calls, and can chain together loops with dozens of tools. Gemini 3.5 Flash performs as well as ChatGPT 5.4 as an AI agent, but it's significantly less expensive. It excels in agent architectures once properly configured. Advanced agentic systems rely on a kernel, just like operating systems. This core is configured and then given access to a set of tools, specifying behavior and stop rules so that Gemini 3.5 Flash understands what it's allowed to do and how many times, in order to successfully complete its mission. I tested it on classic chatbot systems based on RAGs, and I was impressed by its speed and cost-effectiveness: it's an excellent candidate for equipping businesses. If you're just starting out and understand that AI is the growth driver for all companies, optimizing prompts will transform your productivity and profitability. It's not basic prompts that will do the work for you, but your skill in configuring and managing teams of AI agents. That's what you'll learn in the training courses. Level 2, "The Best of AI," allows every entrepreneur to acquire these skills at their own pace in less than 15 days with 85 hours of coursework, including preparation for Google's prompt engineering certifications, 15 pre-configured AI agents, videos, and methodologies. For advanced professionals, Level 4, "Chatbot, AI Agent, and RAG," is designed to transform AI into a working machine for SMEs and freelancers. Everyone will be using AI by 2026, but no one knows how to make it work. That's what I'll teach you: how to be the first to deploy, secure, and sell AI agents via API or on- premises. All the details are in the description. The main difference between a classic chatbot and an agentic system is... The problem is that the chatbot model doesn't need to make decisions; it relies rigidly on a RAG (Relational Aggregate Grid) and memory. In an agentic system, if a question is asked, it must analyze the context, make a decision, choose whether to connect to the internet, which tool to use, where to connect, and which data to retrieve. This is the reasoning system, and to achieve this, system instructions are configured via a behavioral kernel. This isn't a simple prompt; it's a comprehensive structure that manages uncertainty and redirects the model. AI is extremely powerful, but it can drift in all directions. A good architect knows how to frame the model and protect it against prompt injection attacks (such as requesting system instructions or connected tools). A model configured with precise instructions can adhere to robust behavior. These aren't superficial roles; they are neural parameters. A robust prompt should contain: an objective (the desired deliverable), a context (stable data to guide probabilities), and variables. Unlike older, rigid workflows like Make, an agentic system can adapt and realign itself autonomously thanks to rules and validation criteria that enable automatic AI control. The AI ​​self-analyzes and returns a testable and auditable output format (such as JSON). These elements structure a robust prompt, reducing the risk of the model confusing context, instructions, and examples. The prompt becomes a deterministic program containing objectives, input variables, constraints, prohibitions (such as never using a specific function), conditions (if/then), and acceptance criteria. This allows for anticipating edge cases (errors, missing data). A human-in-the-loop (HITL) validation system can also be integrated, where the agent can send an email or Slack message if it needs assistance. This is how AI becomes a true partner that you control. Stop saying, "You're a React expert, make a clean and professional hook." Use a block-structured prompt: "Generates a strict React TS hook, handles loading, error analysis, and API requests. Returns only the complete TypeScript code. Technical constraints: TypeScript is mandatory, no other libraries allowed, targets React Web version 18, no console.log, and no mounting requests via useEffect." Define precise acceptance criteria to validate status, data, errors, and reset functions to finalize the workflow. AI connects words through spatial analogy ( words are vectors in space; semantic proximity is defined by cosine similarity). This is what brings out lexical nuance. Deconstructing your prompts and words requires intellectual effort, but nothing is left to chance because Gemini 3.5 Flash's probability distribution is extremely sensitive. Enrich your vocabulary, because a limited vocabulary produces imprecise structures. What you should no longer do is send vague bug descriptions like, "The application crashes when I try to load my avatar, I get a 500 error, the button remains gray." Translate this into a structured way for the AI, as actionable error reports with precise variables. This new ability of Gemini 3.5 to code native animations relies on HTML5, CSS stylesheets (keyframes, transitions, transforms, filters, backdrop-filter, gradients, shadows, pseudo-elements), Tailwind CSS (grids, flexbox, spacing, responsiveness, layouts) and libraries like Reveal.js with the use of specific patterns (data-auto-animate, data-background-color, fragments, (fragment-blur-in). I'll cover all these structures in the course so you can design perfect animations for web pages, applications, or presentations. It's these technical details that make all the difference between a simple user and an AI professional, capable of efficiently delegating your work to agents. See you soon.