How to Write PERFECT Agent Prompts in 2026 (Complete Guide)

12:38

How to Write PERFECT Agent Prompts in 2026 (Complete Guide)

Corey McClain 14.02.2026 1 041 просмотров 46 лайков обн. 18.02.2026

Machine-readable: Markdown · JSON API · Site index

Смотреть на YouTube

Поделиться Telegram VK Бот

Транскрипт Скачать .md

Анализ с AI

Описание видео

START HERE TO WORK WITH ME Create client-attracting content in one focused hour: clarify your message, generate a high-retention content, and repurpose it into ready-to-post assets → https://www.incomecreator.pro/ltao1 How to Deploy Agentic Prompts Without Coding (Projects, Routers, Memory & NotebookLM) The script explains how the creator began building agentic prompt systems in a project called Koshi using a prompt library (protocols, collaborator prompts, checksums, and evaluator prompts) to compare outputs across models and stress test solutions. The episode focuses on the creator’s “number one” way to deploy agentic prompts—using ChatGPT or Claude Projects—so solopreneurs and creators can automate workflows without coding, MCP servers, CLI tools, or other developer setups. It reviews two core pillars (library and logic) and introduces two additional pillars (artifact and memory), including the role of a router prompt that activates and deactivates specialized prompt files at the right time. Using a “YouTube Made Simple” example with multiple agents (idea, hook, series, thumbnail, title, intro, etc.), the creator shows how large prompts can exceed Custom GPT’s 8,000-character instruction limit by storing them as project files and routing between them. The script also describes a personal “Decision OS” project that uses a decision matrix library and a life blueprint to prioritize tasks based on spiritual/mental state, physical wellness, family relationships, and business finances. It covers working backward from a defined final output, breaking it into components, and building templates/examples to maintain consistent quality. The creator discusses memory access and how expanding native memory reduces the need for workarounds, highlights ChatGPT’s project memory isolation setting, and notes Gemini’s lack of projects while recommending NotebookLM as an equivalent via notebooks that can be added to Gemini 3 Pro as a dataset. The script contrasts this architecture with Claude Skills, arguing that grouped prompts connected by a router in a closed environment are more reliable for complex workflows. It closes with a call to check out a one-hour content strategy kit and a next video on the library/logic deep dive. 00:00 Intro 00:42 Multi-Agent Library 03:15 5 Pillars of Agentic Systems 04:20 Designing an Agentic Workflow 06:15 Building the ‘Constitution’ 07:49 Special Web Deployment 09:14 The Big Opportunity: 10:24 Why Claude Skills Aren’t Enough 11:24 Who This Architecture Is For HOSTINGER (Use my link + code COREY) n8n hosting / VPS setup → https://hostinger.com/corey Website hosting (WordPress) → https://hostinger.com/coreymcclain Email tools (Hostinger Reach) → https://hostinger.com/coreymcclain10 Use code COREY at checkout. TOOLS I USE + RECOMMEND (some links may be affiliate links) YouTube Growth + SEO vidIQ: keyword + competitor insights to plan higher-performing videos → https://vidiq.com/coreymcclain Email + Automation ActiveCampaign: email marketing + automations to turn viewers into leads and buyers → https://www.activecampaign.com/?_r=QIMM97L6 Funnels + Email + Courses (All-in-One) Systeme.io: build funnels, run email, and host products/courses in one platform → https://systeme.io/tr/2/161/3209095746/13990488/82595378fca639b50826408bbdff0b2f5e1d8a0a Recording + Editing Descript: edit video/podcasts like a doc (cut words, clean audio, add captions) → https://www.descript.com?lmref=2vYyUw Riverside: record remote interviews with local tracks (clean audio/video for editing) → https://riverside.fm/creators/affiliates/correy-mcclain Music + Sound Effects Epidemic Sound: royalty-free music + SFX with a creator-friendly license → https://share.epidemicsound.com/5ufv6h Repurposing + Captions Opus Clip: turn long videos into ready-to-post shorts faster → https://www.opus.pro/?via=15a042 Submagic: fast captions + short-form polish Try it here → https://submagic.co/?via=corey60 Use code COREY for 10% off. Ecommerce Shopify: launch an online store and sell products the clean way → https://shopify.pxf.io/B0rMML Presentations Tome: create story-style presentations faster → https://tome.app/invite/corey-mcclain-clc8lqpiu77ea3b8dd83n0qlf Security (Optional) Foxgard Biometric Door Lock: fingerprint + app-controlled smart lock → https://amzn.to/3ZbEiM8 AFFILIATE DISCLOSURE Some links above may be affiliate links. If you purchase through them, I may earn a commission at no extra cost to you. I only recommend tools I use or believe are worth it. CONTACT Business inquiries: corey@incomecreator.pro

Оглавление (9 сегментов)

Intro

This is where it all started for me as far as building agentic prompts in this project named Makoshi. And if you look at the files, you can see there's a library here, the protocol which is the primary prompt, the check sums, the primary prompt, collaborator prompt, there's even an evaluator prompt because what I would do is compare the responses from different models and they would stress test each other to get to the right answer or to solve a constraint in the most efficient and the most effective manner. And this was a process I had built out not for its own sake, but for building other aentic prompts and workflows. And I've since then evolved past this and found more efficient ways to do this. But in this

Multi-Agent Library

video, I want to talk to you about the number one way to deploy these agentic prompts because as someone pointed out in my last video, a good example to bake this in is what's necessary. And I really want you to grasp this idea so that you can start building your own agentic prompts and really getting the value out of your $20 month subscription from chatg plus or the claw pro plan. If we open up this YouTube made simple project, you can see that there are 10 files here. And you can see that there's a thumbnail agent. There is a never again protocol YouTube made simple idea agent, a title agent, intro agent, series agent, hook agent, and then if I go into the settings, you can see that there is a router agent. everything that I talked about in the last video because this prompt in the instructions is managing all of these different files down here. And if I open up any of these files at random and we look at it in detail, you can see just how much content is inside each one of these individual prompts. This one by itself is over 530 lines, over 14,980 characters. Now, just for context, you can only place 8,000 characters inside of the custom instructions window for any custom GPT. This is almost twice that size and it runs perfectly using it this way. This is my decision OS and this one has a router prompt as well, but there are only two files inside of this library. There is a decision matrix library that helps it understand what's most important to me and it prioritizes every decision of my life based on what it knows about me from this life blueprint. And so anytime I have any confusion about the order of operations or I have several things going on, I can brain dump and just tell it what's going on and it will immediately organize everything that I should do and in what order because it knows the dependencies between the relationship of my spiritual and mental state, my wellness physically, my relationships with my family and my finances with my business. So it can help me be as productive as possible and manage my time wisely. If you'll give me a few minutes, I'm going to explain to you the proper way to deploy these agentic prompts after you write them and build them so that you can start creating these systems for yourself as well. And the best part about all of this is that you are not going to have to code or use any developer tools like you see in a lot of the other videos on YouTube. In the last

5 Pillars of Agentic Systems

video, I shared with you two pillars, which is the library and the logic. And those are absolutely critical. But there are three other pillars in this system that I did not discuss in the last video. And I'm going to share two of them with you right now. And that is the artifact and the memory. And so something I want you to understand as you're watching this video is that this is going to be different from the content that you typically see online where most of the content you see around agents is MCP servers. It's clawed code. It's the CLI, the command line interface, and it's all of this technical stuff. But you're a professional creator, you're a soloreneur, and you don't have time to become a super techy, geeked out person. You just want to use your chat GPT cla Gemini or your Grock interface and you want to automate some of your work. You want to have some workflows that make it feel aentic because it knows what to do and it knows how to deliver the same quality standard grade work that you expect every single time. then this video is going to be perfect for you because you are the people that I build these workflows for. But let's quickly

Designing an Agentic Workflow

cover some of the principles that we talked about already. One of the first things you want to do when you're getting ready to set up an agentic workflow is you want to think about what is the final output. So for my YouTube made simple project, the final output is a published video on YouTube. Now how well the video performs, that's a completely different topic. That's YouTube skills, experience, and things you need to learn. but a published video on YouTube that people actually enjoy. From that point, I begin to work backwards because that is the child output. And so now I understand what the final output is, but now I need to break it down into the smallest components possible. And so I realize that, you know, I need to get an idea. I need to have a hook. I need to think about whether this is going to be a series or not. I need to think about a thumbnail. I need to write my introduction. I need to think about the body of the content. I need to think about retention. I need to play it back and listen to it. I need to think about calls to action. And so once I've identified each of these individual items, I can then begin to build my library. All of these are things that I sit down and I decide no matter what the system is that I'm building. Sometimes it's easier and sometimes it's more difficult depending on the goal that you're trying to achieve, which makes it so important. You clearly identify what the final output is. So, for instance, let's say that I was building a copywriting agent. I would need to think about all of the different types of copy that could possibly be written. And then I would need to find a way to realistically and structurally give that agent the expert knowledge that it needed to write copy for all of these different types of copy that it could possibly write. And sometimes what it's going to need is a tool set with templates, diagrams, examples, and other items that streamline the process and make it move smoothly and quickly while also maintaining the same standard of quality on every output. That's what makes it

Building the ‘Constitution’

worthwhile. And then after the library is established, that's where you begin to set up the parent. This is where you write your constitution. lay out your guard rails. This is where you define everything that this agent will or will not do. This is where your conditional logic lives if it can. You always want to put your conditional logic at the highest place possible. Now, with OpenAI, this is not possible for the most part because you are limited to 8,000 characters. If it's a simple agent with a simple output, then 8,000 characters is going to be more than enough. But if it's something complex like YouTube made simple, then I am absolutely going to need to put those prompts inside the library so that the router prompt can actually manage them and pull them when necessary. And like I said before, this is not a problem when you're using Claude or Google Gemini. The contest window for the custom instructions of both of those platforms are probably at least twice the size of chat GPTs. If I took you back to my YouTube made simple agent, you would see that there are eight different mega prompts. These prompts are all in the ballpark of 10 to 15,000 characters long. And the router prompts responsibility is to make sure that each individual agent is activated and deactivated at the proper moment. And with this type of architecture, there's no limit to what you can automate. And there's one particular platform that you need to start using right now that we're going to talk about in a minute that is absolutely going to crush it with this in the next year or two. They're doing well with it right now, but they are going to dominate everyone else with this in a year or two. Just watch what I

Special Web Deployment

tell you. Memory access protocol is one of those pillars that I did not mention previously. But memory access protocol helps the agent remember you remember what you learn, remember the ideas you share and it grows alongside you and becomes better and improves. And so when we had this problem with the platforms with their very low memory, what ended up having to happen was that people were setting up these workarounds with MCP servers and virtual private servers. But now that the platforms are widening their memory, giving us more access to memory, all of this can be contained natively. And so if you build an agent prompt today, in a couple of weeks or a couple of months is going to change because you're going to learn and it's going to learn with you and make suggestions. And the best way to deploy these is inside of chat GPT or claw projects because it allows you to place all of your files there. your custom instructions there, but even more importantly, it allows you to group your conversations together. And there's one chat GPT memory feature inside their projects that I wish claw projects had. But it allows you to choose whether the memory for your project is contained within that project or whether chat GPT can access other conversations outside or from outside within that project. And whenever there's an important project that I create, that's the number one thing I look to do is to turn that setting on so that I don't have to worry about any noise inside of that conversation. And that allows chat GPT to be more precise with its memory and focus on what matters to me at the moment. Now

The Big Opportunity:

currently you cannot do this with Gemini because Gemini does not have projects. But the equivalent of projects for Gemini is Notebook LM. Now with Notebook LM this is so huge and this is the opportunity that I was telling you about that a lot of people are going to miss out on because they really don't understand. But the fact that you are able to upload 50 different sources inside of a notebook and then you're able to have I think a hundred notebooks on the free plan and then you can take that notebook and add it to a conversation inside Gemini 3 Pro as a data set and Gemini 3 Pro will behave as though it were uploaded inside of a project. So it's essentially the same thing. a few more steps, but it's all inside of the Google ecosystem and it operates exactly as chat GPT and exactly as claw. But as time progresses and as all of these models begin to raise the floor, you're going to see that this particular workflow is going to become so much more powerful when you begin to understand that you can actually influence you can almost control the behavior of Notebook LM, which is really just Gemini 3 flash inside of the Notebook LM environment. I know that

Why Claude Skills Aren’t Enough

some of you are going to think, well, what about claw skills? And the thing about claw skills is that they are just simply prompts or howto manuals. They don't necessarily capture the templates, the instructions, the several different prompts and then bring them all together under one roof. Theoretically, I could upload all eight of my YouTube made simple prompts and claw could determine when it needed which one. But maybe it misses one because there's no router prompt to bring it all together. I'm counting on claw to do that. And so it's just a much better architecture when they're isolated, they're grouped together, and they're connected by a router prompt. They're mentioned in the router prompt by name, and they're in a closed off environment where they can do the work that you want them to do. Clause skills are great. Don't get me wrong. They are absolutely wonderful. I just think that for this particular use case, you want to go with the architecture that I'm showing you. It's a little more work, but it's definitely worth it. The logic

Who This Architecture Is For

behind these agentic prompts that I'm talking about, these workflows and how I use projects custom GPTs, the same logic can apply to coding agents. It's the same thing, but I've created this architecture for the individual who wants the benefit of agents inside of chat GPT Gemini and Claw. Now, of course, they are not fully autonomous like agents, but they do understand step-by-step procedures. They do standardize your workflows. They do standardize quality, which gives you quality assurance, which increases your output and which can help you achieve your goals for using AI. So, if you're a small business owner or a soloreneur with an online business who uses content to get clients for their business, make sure you check out my 1hour content strategy kit. The link will be in the description. And if you want to learn more about how to build your own agentic prompts, then make sure you check out this next video right here where I do a deep dive into the logic behind the logic and the library that will tell you everything you need to know. And if you don't got time to watch the full video, just copy the transcript and drop it into the chat GPT or Claude or Gemini and they'll work it out for you. But either way, if you got value out of this video, make sure you hit the like button, subscribe to the channel, and as always, take care, have a good day, and I'll see you in the next

Другие видео автора — Corey McClain

Ctrl+V

Экстракт Знаний в Telegram

Экстракты и дистилляты из лучших YouTube-каналов — сразу после публикации.

Подписаться

Лучшие методички за неделю — каждый понедельник