# 08 2026-07-26 Live Session Agents Contd

course: Module 5 — AI Agents & Agentic Frameworks
module: Module-5-AI-Agents-Agentic-Frameworks
date: 2026-07-26
type: transcript
video_url: https://personal-learn.armco.dev/files/_Recordings/Module-5-AI-Agents-Agentic-Frameworks/08_2026-07-26_Live_Session_Agents_Contd.mp4

---
[00:20:37] not with, like, Landgraph, LangChain, and all those things, but…
[00:20:41] normal in general with Python only.
[00:20:47] So it's JavaScript.
[00:20:48] Uh, not with Python, but we did something with JavaScript and tool calling and all yeah.
[00:20:50] Yeah, we did that.
[00:20:51] Okay, okay. So in JavaScript, since
[00:20:56] I am not JavaScript person, so in JavaScript, do you have a framework to do that or plain JavaScript only, like.
[00:21:01] There's no external.
[00:21:02] Plain JavaScript and there are direct libraries which you can, uh, import and all.
[00:21:06] Uh, it's like JavaScript libraries, right? Directly you can import.
[00:21:12] The adapters are provided by OpenAI and other platforms.
[00:21:18] You can directly import it and send the queries and all.
[00:21:21] Hmm.
[00:21:22] Same way we do it in the previous classes, right, we mentioned tools and its inputs and all, yeah, same thing.
[00:21:26] Okay.
[00:21:30] Okay, so the point I'm trying to make Deepak is that tool calling example which I had.
[00:21:36] That was purely Python and at
[00:21:37] Yeah.
[00:21:39] didn't have any agentic framework.
[00:21:42] Hmm.
[00:21:43] Okay, we had few libraries that we have imported for other things like Pyrantic and OpenAI.
[00:21:46] Mm-hmm.
[00:21:47] Well, OpenAI SDK and all those things.
[00:21:50] That is for calling the model. But for agents, we didn't use any framework. So, similarly.
[00:21:54] Okay.
[00:21:55] In JavaScript, when you do, did you do any frameworks for agents or for other things we have used other libraries?
[00:22:01] For other things, we did not particularly create any agent. If you consider the whole tool calling thing, it was an agent to give the
[00:22:08] Gotcha.
[00:22:09] Weather, uh, details of a particular, uh, uh…
[00:22:12] It was a tool basically that is giving you weather details.
[00:22:14] Correct, correct.
[00:22:15] No, the point I'm trying to understand that, do you have any exposure with the agentic framework.
[00:22:20] Like LangChain, Language UI you have in Python, do you have a similar thing in JavaScript or in other language you have done?
[00:22:26] Do you have ah.
[00:22:27] No, no, it was mostly API, API thing, yeah.
[00:22:29] Okay, got it, got it.
[00:22:31] Okay, so guys, see, personally,
[00:22:36] When I have explored Claude Code and, uh, when I keep on, you know, using
[00:22:41] these tools. The point is that I'm trying to make, they also
[00:22:46] They also help you build agents and agentic.
[00:22:47] application. If you ask Claude Code.
[00:22:50] to build an agentic application for you.
[00:22:52] Okay, with a certain logic, maybe…
[00:22:54] Something that creates, let's say, product descriptions and all these things, or let's say a travel itinerary planner.
[00:23:00] It will create for you, but that is also agentic.
[00:23:04] But that creates using purely Python.
[00:23:08] And and you know these skills and all those things because.
[00:23:11] You know, they don't use LandGraph Langchain if you don't specify it. So they use pure Python.
[00:23:17] So you might, so in our class, you might learn these framework.
[00:23:23] Okay, but what I'm trying to say, that
[00:23:26] These frameworks underlying are using pure plane Python only. Like, you know, many people are creating custom framework also in their companies.
[00:23:33] They are not relying on any other company. I had worked with 1 company, UK-based company.
[00:23:38] When I joined, my biggest question was at that time I didn't know Language and LangRap, because I was, you know, LangChain I knew, but I didn't know how to build.
[00:23:46] Agents using LangRap it is just Landgraph came into the market. It's almost like.
[00:23:51] One and a half years, two years kind of a thing.
[00:23:53] So I had asked like, uh, how are you building the agentic workflow? So they told I used using Python.
[00:23:59] purely Python. So they wrote long, long lines in order to achieve the same thing.
[00:24:04] Like LangRap using Biden. So these frameworks are, again, it's like an abstraction.
[00:24:10] Internally is our plain Python only.
[00:24:12] It's done. So these framework only we will learn today.
[00:24:15] Okay, but what I'm trying, the discussion that I'm trying to have over now is that these frameworks.
[00:24:22] If you don't know these frameworks, but still create agentic AI.
[00:24:26] AI and your company approves it, it's good.
[00:24:30] It's fine. It's appreciate like I appreciate that, like, you know, that is also fine. Like, you don't have to compare yourself.
[00:24:37] that you have to know these frameworks only.
[00:24:40] Obviously, for the agentic codes and all these things, these frameworks becomes mandatory, but many enterprises are there, they are going framework agnostic.
[00:24:49] They are building purely using Python.
[00:24:51] Ah, like no external frameworks, purely using Python. So, over there, maybe you have to write a little bit of more functions and
[00:24:58] The amount of code writing increases.
[00:25:01] Okay, so…
[00:25:03] Even if you use this Claude and all these things and tell that I want agentic application, they naturally then don't choose any framework, okay.
[00:25:10] If you don't mention it. So they also build using Python only.
[00:25:14] So that's what I was asking, like, if you have done Claude Code, then, uh, and built Claude code, use Claude Code to build an agent.
[00:25:22] So that we can see that agent, if you have it.
[00:25:24] That's what I was checking.
[00:25:26] Okay, anyways, so first before even going to agent, we will
[00:25:30] Learn about, ah.
[00:25:33] Interesting thing because, uh, I see over here.
[00:25:37] Few people only know about.
[00:25:39] Prompt engineering. Okay, after this part that we discuss, you will know about prompt engineering very well.
[00:25:59] Installment.
[00:26:16] Hmm.
[00:26:18] Okay, so I am sharing a file with all of you.
[00:26:22] Okay, I'll be sharing many files today.
[00:26:25] All are HTML files.
[00:26:55] Okay, these 3 files, few of them are your study materials, like the prompt region practical for you to practice and, you know, improve yourself, that is purely for you.
[00:27:04] These two files we will be discussing in the class, one is prompt engineering example.
[00:27:10] And one is more advanced prompt engineering.
[00:27:12] Okay, so…
[00:27:14] There are overlapping content, few contents are extra in that more advanced, uh.
[00:27:19] file that we will do after this, ok.
[00:27:22] So, this is an HTML file, I hope you can access it.
[00:27:27] You know, I don't think we are able to access it. At least I'm getting a message saying that this content cannot be viewed here
[00:27:34] Can you share your screen once?
[00:27:47] Can you share it once?
[00:27:48] Can I give you the permission. I will have to quit and reopen. I don't have permissions to do it. I will have to drop from Zoom. Can someone else do it?
[00:27:56] I downloaded, and then I can able to
[00:27:58] Yeah, download it, download it.
[00:28:01] You can't view it in your chat.
[00:28:04] Okay, let me try that.
[00:28:05] We can't see the link, actually, Anand. Like,
[00:28:09] Good morning, nothing is single. Just this return can't be viewed here, something like that, that message is displaying, that's it.
[00:28:18] I'm not able to see the content.
[00:28:22] This content cannot be viewed here
[00:28:24] Yes, but I'm trying to set the screen. Just let me know, okay?
[00:28:30] Tell me like this
[00:28:31] Oh my god.
[00:28:33] Yep, the same thing
[00:28:34] And how did Deepan get it?
[00:28:38] Yeah, you faded the…
[00:28:40] Dinosaur roar
[00:28:41] download it, I can… I can see it, actually. I can see the content.
[00:28:45] But, uh, they're not getting the option to download.
[00:28:47] But we can't see the link, right? Yeah, nothing is
[00:28:51] How do you get the option to download?
[00:28:54] Just… just if we keep the cursor over there, it will show there is a download option here, in that, uh, whatever you shared in the chat.
[00:29:02] If you just drive…
[00:29:03] On Naresh, it's coming.
[00:29:07] Oh, it's PNG.
[00:29:09] Okay, you have uploaded the PNG format, but it's opening, right?
[00:29:13] Yeah, it's opening, Nirvan. After downloading, right, I can able to open this, yes.
[00:29:19] Yeah
[00:29:20] So, how did you get the download options, Sandra Shaker?
[00:29:25] But
[00:29:26] So, you have that three dots, right? When you hover on the three options, you're getting that three dots. When you click on that, you are able to see the download option down.
[00:29:31] Where, where can you…
[00:29:32] Three dots where you hover on the, you know, item… three items, right? Promptengineering.html
[00:29:39] No, but nothing is going right this week. Like, for me, actually
[00:29:41] I think, probably the browser version of Zoom doesn't give it
[00:29:44] That
[00:29:45] Oh, browser version, browser version, it won't give you.
[00:29:47] You need the…
[00:29:48] Okay.
[00:29:49] Zoom applications, sir.
[00:29:50] The same. Yeah, application.
[00:29:51] We can drop it in WhatsApp if you want.
[00:29:55] Yeah, if you, do you all have a group?
[00:29:56] Yeah, we do, yeah.
[00:29:59] Yeah, please, please, please do that.
[00:30:03] Whoever has got it deeper.
[00:30:05] Maybe Chandra Shekhar, if you can do it.
[00:30:10] Uh, Anirban, can you please share the link again? I actually joined just now.
[00:30:17] Hmm.
[00:30:30] Here it is.
[00:31:12] Got it? Now, guys,
[00:31:18] No, actually.
[00:31:20] get dropped in both the…
[00:31:22] uh, WhatsApp groups.
[00:31:23] Three files, right, Deepak?
[00:31:26] Three files are.
[00:31:34] No, I'm not received any WhatsApp messages
[00:31:35] Are they using work devices? Yeah, Receich is…
[00:31:37] Yeah, that's no… I also posted it in the, in the WhatsApp group.
[00:31:45] Thank you, thank you, Sadish.
[00:31:46] Okay.
[00:31:55] Got it right, Azadi, you got it, so.
[00:31:57] On WhatsApp, you all got it.
[00:31:58] Yes, yes, yeah, what's the first, I can see that.
[00:32:01] Okay. Okay.
[00:32:03] Please help me also in the WhatsApp here, I will bring my number here.
[00:32:08] Yeah, yeah.
[00:32:19] Okay, so guys.
[00:32:21] This is something.
[00:32:24] Uh, that ideally.
[00:32:26] I know.
[00:32:30] It should have been done, you know, way before even you started.
[00:32:35] Gen AI, but maybe, you know, maybe the scope was not there.
[00:32:38] Uh, in the previous…
[00:32:41] With the previous professor or something like it was not there.
[00:32:45] But, uh, you know, prompt engineering is usually covered in the starting point of gen AI.
[00:32:50] But now it becomes very mandatory to cover at this point also, because if it is not covered, it becomes very mandatory to cover because.
[00:32:58] Uh, your agent DKI, you need to write big, big prompts, okay? You are going to explore with prompts. You know the techniques of prompts.
[00:33:04] As you build applications,
[00:33:06] for your enterprise or wherever you will be working.
[00:33:09] Okay, so first of all, prompt engineering is just not writing going to ChatGPT and writing vague.
[00:33:16] things on ChatGPT, because that is a free tool.
[00:33:19] And that's why we just write like that. There is a systematic way to, you know, get the best out of AI.
[00:33:25] So what is prompt? Prompt?
[00:33:27] Like, I told, I think yesterday also I told this that
[00:33:29] Prompt is to AI is just like what is.
[00:33:34] But, you know, program to your computer.
[00:33:38] Okay, how do you interact with your computer? Through a…
[00:33:43] Some programming language. Similarly, you interact with your AI.
[00:33:45] through a prompt. If you want to get the best out of your
[00:33:50] coding language, you have to write good.
[00:33:52] Uh, you know, code and programs. Similarly, if you want to get the best out of AI, you have to write.
[00:33:56] good prompts. Okay, so prompt is basically you are mastering the art of communicating with AI.
[00:34:03] Now, again, guys, think…
[00:34:06] Keep your, you know,
[00:34:08] Trivial thoughts about
[00:34:11] that in ChatGPT this is what do you do in ChatGPT.
[00:34:14] No, this is what that will be doing. Obviously, we'll be using ChatGPT and Copilot a lot in order to compare these prompting techniques and all.
[00:34:23] But what I'm trying to say is that this is.
[00:34:26] Trial and tested on ChatGPT.
[00:34:29] And trial and test it on Copilot or Claude.
[00:34:31] And then we design the ideal prompt.
[00:34:33] And then we use it in our enterprise.
[00:34:36] We'll come to that thing in a while.
[00:34:39] Okay, so first of all,
[00:34:40] What is prompt design? How does it help you? Definitely it will impact.
[00:34:45] performance a lot. Now, you all tell me, how do you think good, effective prompting can impact performance.
[00:34:51] Cost, consistency and speed, four things.
[00:34:55] Before, like, I want to see that are you even…
[00:34:58] getting what is prompt engineering, so that's why, let's see if you can answer.
[00:35:02] How does it impact performance? Like
[00:35:05] Why do you think prompt engineering can impact a scalable applications performance?
[00:35:14] giving the proper instruction to the scale files in a what markdown file or what is the directory structure I need to look for? I think instead of Agents spend quite a lot of time on reading a lot of contents, it can actually improve the performance by just looking at
[00:35:30] Where you wanted to do.
[00:35:33] Again, this is from skill standpoint, not from your prompt standpoint
[00:35:34] Yeah, I get it, I get right, I get it. Most of your answers will come with skill standpoint only.
[00:35:39] I get it.
[00:35:41] It is actually… you're actually utilizing the system with the… minimizing the cost and getting the output from the LLM
[00:35:52] What we expect
[00:35:53] Ah, yes, very good, very good, uh, what about consistency and speed?
[00:36:07] How do you will, how will you explain that with respect to prompt engineering? Anybody?
[00:36:11] Yeah, instead of going to the KLM multiple times, so we can try to give you a…
[00:36:19] plant in such a way, like, how much ever possible, we can take it in a single time.
[00:36:24] And then we can do it now.
[00:36:25] Yes. Hmm.
[00:36:26] We… actually, I think, right, I think it is more like re-utilizing the prompt, like, system can store, like, context video, where we can follow some sequence of steps, which he can… it will not go to LLM for everything. It will reuse the memory, whichever it is like contact
[00:36:46] And give the answers to us
[00:36:47] That is the prompt itself. We need to have a systematic format, like utilizing system, and even to
[00:36:55] In terms of consistency, that we are giving the system prompt, where it can give the output format as predefined how we want, like JSON, or whatever the required format to us
[00:37:06] So basically, basically, see guys, if you compare it with you using ChatGPT versus you, you, you designing a platform or a
[00:37:16] or a framework where this thing become very, very consistent.
[00:37:20] See, uh…
[00:37:22] What happens is, in a agentic architecture, you build a lot of agents.
[00:37:28] And you write a lot of prompts, you take help of AI as well. We will be doing that technique as well known as meta prompting. That is the last one, and that is the.
[00:37:35] That is the ultimate technique that we nowadays use, is using AI to generate all these things.
[00:37:40] But, uh, the point is
[00:37:43] That if you don't use ChatGPT.
[00:37:46] Uh, the moment you move away from ChatGPT and Copilot and all these free tools.
[00:37:52] and start using enterprise-level subscription.
[00:37:55] in your enterprise development.
[00:37:58] What happens is, you are forced to
[00:38:01] Come within a framework, you are forced to come within a limitation.
[00:38:06] of writing your prompts.
[00:38:07] You cannot just write.
[00:38:10] No rubbish things, like in ChatGPT, you just go and you write.
[00:38:15] give me the temperature, uh, you know, what is the temperature today? Should I carry an umbrella? And all these things.
[00:38:19] You're asking like vague things.
[00:38:21] But imagine if you are building an agent and in an agent, there is a long prompt that is written.
[00:38:26] Okay, that probably that prompt just summarizes
[00:38:29] uh, your activities till now.
[00:38:32] Or, uh, summarizes and not summarizing the chat. I say, an agent that summarizes.
[00:38:36] activities that has happened in this software development. Let's say you're building a coding agent only.
[00:38:42] So, over there, the structure that you would like to get the output will need a proper framework, proper.
[00:38:48] You know, instructions, proper, you know, management of tools and everything. So that is why effective prompt writing becomes important, because your prompt is the seed to all your agents, because your agents is.
[00:38:59] prompt only, ultimately. Okay, your agent says prompt crucial part of agents is prompt. 90% of the agents will be prompt.
[00:39:06] Okay, so your prompt becomes everything. Nowadays, see, we use AI only to build prompts, even your.
[00:39:12] The skills file, skills files which people are saying, skills.
[00:39:14] Okay, those are also built using AI only. Those are prompts only.
[00:39:18] Okay, but the point is that you at least should be aware of these tools that these techniques exist.
[00:39:24] Then only you can make prompt design a very, very
[00:39:27] Predictable thing. Why predictable? Because if you have a targeted prompt, if you have a very targeted, specific, not a vague prompt.
[00:39:34] You will see the output that comes is very predictable in nature across LLM. Even if you might use
[00:39:40] Copilot chat, uh, you might use ChatGPT, you might use Claude. Your answers will be very predictable in nature.
[00:39:47] Okay, so it improves the performance. Definitely, you are not going writing again and again on how you write on ChatGPT.
[00:39:54] Okay, so your performance will reduce, ok.
[00:39:57] The accuracy will improve.
[00:40:00] Cost efficiency will come down.
[00:40:02] Because you are not again and again hitting, so.
[00:40:05] Your token management will improve because you will force your LLM that you should output within this many tokens also that also is there.
[00:40:12] Okay. And output will be consistent across use cases, across LLMs.
[00:40:17] And there will be a good speed also.
[00:40:20] To get the right answer instead of again and again hitting.
[00:40:22] You will get the right answer faster. If you have a design prompt.
[00:40:26] Okay, for example, you want to
[00:40:29] If I take this right about dogs, this is a very vague example, ok, this is a very vague prompt.
[00:40:39] If I take this,
[00:40:47] Write about dogs.
[00:40:52] Go to Copilot. You all can practice with me as well.
[00:40:56] Write about dogs, because, see,
[00:40:58] prompting something that you should know in the first day only.
[00:41:02] Okay, let me also mention without using internet.
[00:41:05] Uh, why I'm mentioning because
[00:41:07] You know, internet will add extra tool into it, and the tool might be different for.
[00:41:14] Copilot might be different from ChatGPT.
[00:41:15] And both will unnecessarily create a different answer.
[00:41:20] So I will not get to, you know, get to compare them without, without…
[00:41:24] Using the internet.
[00:41:26] I want the LLM to answer from its own capability, okay?
[00:41:30] Not using the internet.
[00:41:34] Okay.
[00:41:36] So, first of all, this is a very vague prompt. If you write vague prompt, you will get vague answers like this.
[00:41:43] Dog are extraordinary companions, compa- ah, extraordinary companions.
[00:41:46] Ah, blending royalty, loyalty, intelligence, playfulness in a way that few other.
[00:41:52] Animals can, they have lived alongside humans for thousands of years, very, very generic concept.
[00:41:58] Genetic diversity of breeds, and this is the way it answers.
[00:42:02] Okay, now…
[00:42:04] The point is, similar way, ChatGPT also used to answer few days back. Nowadays, they have started using
[00:42:11] The framework which forces you to
[00:42:14] Ask further question, which was, which.
[00:42:16] You know, they ask you further question. What should the dog writing be? What should the dog writing be? Let's say.
[00:42:23] Short paragraph.
[00:42:26] Dog breeds, ah, should it cover dog breeds?
[00:42:29] See, since the prompt is very vague, it is asking you follow-up questions in order to make it more.
[00:42:33] Uh, you know, targeted, more funnel, funneled.
[00:42:37] Okay, so as you can see,
[00:42:39] Both the outputs are quite different, first of all, this is quite different, this is quite different.
[00:42:45] Okay, this is the problem with wake prompt.
[00:42:48] Okay, now let's move away from this and let's write like this.
[00:42:53] Write a 200-word guide for the first time dog owners about choosing the right breed for animal living.
[00:42:59] Ah, apart from living, including 3 specific breed recommendations.
[00:43:03] with brief explanation. This is a very specific prong.
[00:43:14] Okay, now see.
[00:43:21] Here is a concise 200 guide tailored for the first-time dog owners choosing the right bead for department living. I have highlighted 3 breeds, French Bulldog, Cavalier.
[00:43:28] King Charles Spaniel and Pug.
[00:43:31] Okay, all data has used the internet. I wanted not to use the internet.
[00:43:35] Let's see, this is also, I think.
[00:43:38] We'll use the internet because if it starts using the internet, they're the both answers will be different.
[00:43:42] Really forced to give different answer because both have different search mechanism. One might use Bing, one might use Google.
[00:43:49] And that's the thing. So I wanted to show the LLM's capability first.
[00:43:53] Okay, internet is extra tools.
[00:43:55] But let's see, ok, uh, see guys, see.
[00:43:59] Almost similar guys, almost similar, Cavalier King Charles Spaniel.
[00:44:03] Let's see, Cavalier King Charles Maniels is recommended.
[00:44:06] French Bulldog, French Bulldog and over here PAG is recommended over here.
[00:44:13] Miniature poodle is recommended, this is some sort of a dog.
[00:44:15] Okay, so what I'm trying to say is that when you write targeted prompt, this is not even so targeted also. We have more
[00:44:22] Targeted prompts, better targeted prompts.
[00:44:25] Okay.
[00:44:28] We have better targeted prompt, so the moment.
[00:44:31] We…
[00:44:33] Write targeted prompts, the output becomes more consistent in nature, that is.
[00:44:38] That is what we are going to study today.
[00:44:42] Okay, so this is still fine. This is like a this is not so targeted also but yeah, this is a short prompt only.
[00:44:47] But this is better than this wake prompt, which was generic unfocused content and all.
[00:44:51] Okay, so.
[00:44:54] Now, your LLMs are very smart enough that it is forcing you to write that funnel prompt, which LLM only asking you questions about.
[00:45:01] And LLM is only giving you.
[00:45:03] So that alien can come with a better output.
[00:45:06] that your ChatGPT does, even your Claude also does.
[00:45:09] Copilot is not still doing, maybe it will do after a few days, but, you know, this…
[00:45:13] Problem existed before as well.
[00:45:17] Okay, uh…
[00:45:19] So anyways, uh…
[00:45:21] Now coming to what is the ideal prompting technique. So if you ever go and search prompt.
[00:45:27] engineering frameworks. You might have used yourself some other techniques like CoStar.
[00:45:34] bridge. Okay, there are more frameworks like I used one of them which was known as.
[00:45:39] CTIR or something like context.
[00:45:43] target, intent, role,
[00:45:46] Like that. Okay, so they had created a big name out of it, and they big.
[00:45:51] abbreviation out of it. So, like this lot of
[00:45:54] framework, standardized framework scheme. The one standardized framework that stood.
[00:45:58] The time and that should actually.
[00:46:02] Uh, you know, that stood out in the market was a race framework.
[00:46:04] Which is role, action, context, and expectation. First of all, it is very easy to remember.
[00:46:10] And it is very popular and easy to understand as well.
[00:46:13] Raise framework is basically
[00:46:16] When you have role for definition of your AI persona, let's say you ask your AI that you think yourself.
[00:46:23] Like a marketing strategist. So that is kind of like a role you are giving. So that way.
[00:46:28] Your AI is very focused while thinking. It is not thinking like a generic person.
[00:46:32] Like a generalist. So, it is already thinking with the mindset of a
[00:46:37] marketing personality, okay?
[00:46:40] Then you tell the action.
[00:46:43] You give some background story, which is like a context.
[00:46:44] And the expected output, expected as in, like, give a 200, you know, pay 200, uh,
[00:46:51] Characters, let's say story, or 200 word story.
[00:46:54] Like that. So, expectation is like
[00:46:57] gave in bullet points like that.
[00:46:59] Okay, that is expectation. Raise framework is the most.
[00:47:02] Common prompting technique that we use when
[00:47:06] We play around with prompts and we build prompts like that.
[00:47:10] Okay, so even if you are focused only with skills, skills, skills, then over there also this.
[00:47:17] Framework unknowingly has been used.
[00:47:19] Okay, so now there are many versions of this. That's what I told you, there is costar.
[00:47:25] Like if you search, you will see co-star as well, which is same thing. It is contest objective.
[00:47:29] Style, tone, audience response.
[00:47:32] So this style and all this tone, all these things comes under expectation only in race framework, all these things comes under expectation, style, tone.
[00:47:39] All these into, you know, audience and all these things comes under context only.
[00:47:44] For us, because background story has context.
[00:47:47] will have audience.
[00:47:48] Okay, so there are different names. There is bridge framework. So there are some more popular names to this same thing only it is all race framework only, but different, different names have been given because.
[00:47:58] Because this is… this doesn't need standardization.
[00:48:01] This format doesn't need any standardization. You might even use CodeStar format as well.
[00:48:06] But they are doing the same thing only.
[00:48:08] Okay.
[00:48:11] Mota Moti, they will do the this thing only role action context and expectation.
[00:48:14] By the way, this is like you are writing, write a marketing plan for our new app. Let's go.
[00:48:26] So, if you write like this, you will get a generic answer. See, target audience,
[00:48:30] Position, acquisition channels.
[00:48:34] Retention strategies, metric KPIs, cost per install.
[00:48:36] Okay, so these are generic, very generic is a structure marketing guy. There's nothing specific.
[00:48:43] Okay, but…
[00:48:45] ChatGPT anyways forces you, so.
[00:48:47] It is asking what is the app about, what does the app do? Productivity or business? So, let's say.
[00:48:53] Productivity, small business, target audience.
[00:48:56] Increase the app downloads.
[00:48:59] See, it has already asked you in a finalized, it has already created its own funnel.
[00:49:02] Uh, that these are the things. Now, it will be very specific.
[00:49:06] Because it is not vague anymore. It has asked a lot of question, that is.
[00:49:10] able that it will use for its in its own knowledge to come up with a better plan.
[00:49:14] So, this is way better because it's not very vague anymore, generate 10,000 app downloads within first 6 months.
[00:49:19] achieve a download to registration cost conversion rate of 40%. It didn't write generic.
[00:49:24] Things like KPI and all these things.
[00:49:27] Okay, metrics and KPI. It has directly told you how do you do the strategy because it has that targeted prompt already.
[00:49:33] Okay, this feature came in new, very new with ChatGPT. This was not there in our last.
[00:49:38] Cohort also we hit we used to hit both of them together, and both of them used to come up with a vague.
[00:49:44] Very different, different output.
[00:49:46] Now the point is, now if you write in a race framework format.
[00:49:50] See, in race framework, you don't need to mention like this role, action, all these things. You just this is this automatically means role.
[00:49:57] This means action. This means context and expectation. You don't need to mention them separately, by the way.
[00:50:02] Now, this is mentioned over here, this is for formatting.
[00:50:05] For your understanding, but you don't need to mention rule action.
[00:50:08] You're going to remove that. No problem.
[00:50:11] Okay, now let's give the same prompt everywhere.
[00:50:14] See, now you see.
[00:50:18] The kind of consistency you will see. This is what I was meant by consistency.
[00:50:21] The structuring will be same.
[00:50:23] Okay, maybe this will have more because of the LLM. Maybe it is using thinking model at the first.
[00:50:29] Always, that's why, you know, it is coming up with a very big elaborated output.
[00:50:35] Okay, this also I will come towards what is thinking model and all, how does thinking model usually works.
[00:50:39] Okay, so that is why it is giving an expected output a little bigger, but you have the same similar kind of a structure nowadays.
[00:50:47] Okay, the moment you start hitting targeted form, it's a content marketing plan. It is an instant model. This is the instant model. This is not a thinking model.
[00:50:54] Okay, Copilot doesn't give you that option.
[00:50:57] Think deeper, yeah. So if you give this, probably it will.
[00:51:01] you know, come up with a more detailed output.
[00:51:03] But yeah, so the point I'm trying to make is that
[00:51:08] You expect a lot of consistency across when you
[00:51:11] Even if your LLMs might be different, even if your LLM API might be different.
[00:51:16] Your output becomes start becoming consistent when you use a consistent framework, when you use a particular targeted framework.
[00:51:22] So over here, cleared expertise level is there.
[00:51:25] Specific task is there.
[00:51:27] Complete background provided, exact format requested. Everything is there.
[00:51:32] Okay.
[00:51:35] So now, to break down race frameworks, few things is there are sometimes, you know, you go to ChatGPT.
[00:51:42] And you write like this, make this better.
[00:51:44] Okay, improve the readability of this.
[00:51:47] Email by simplifying complex sentences using bullet points for list.
[00:51:51] and ensuring professional but friendly tone. Keep it under 200 words. Obviously, this is better.
[00:51:57] This has specific improvement criteria, clear formatting instructions,
[00:52:01] and define constants. Everything is clearly mentioned.
[00:52:04] You don't make this better, make this shorter, make this crisp.
[00:52:08] Like that, instead of that, if you write like this, this is more clear.
[00:52:11] Okay, so…
[00:52:13] This is about clarity. You need to give this clarity. In your race framework also, this can be a part of your race framework only, this clarity.
[00:52:20] thing inside your prompt OLED, you can, maybe in the expectation, you can write it.
[00:52:24] Or in an iterative prompt also you can write it. Let's say the answer you got, then you might write it.
[00:52:29] improve this, improve that. Okay, so you have to be very clarity will should be there and there should be a lot of specificity in terms of.
[00:52:35] tone, style, target audience, these are some quality text fix, and you have some quantitative specs like.
[00:52:41] What is the word count? What is the number of items?
[00:52:44] Time constraint, specific metrics or KPI.
[00:52:47] Let's say this is what an example.
[00:52:50] It's vague and unclear is write some content for social media.
[00:52:54] about our company.
[00:52:56] Create a LinkedIn post announcing our company's new remote work policy, write 150 to 200 words in a professional yet approachable tone.
[00:53:06] for our employee audience, include 3 key benefits of the policy.
[00:53:10] Use bullet points for easy reading, and with an encouraging message about work-life balance and avoid corporate jargon. Keep the language conversational.
[00:53:18] Very clear, very specific.
[00:53:21] Even what counts are mentioned.
[00:53:22] Clear topic is mentioned. Defined tone is there, target audience is there.
[00:53:27] Proper structuring of 3 benefits, bullet points, all these things are mentioned.
[00:53:32] Okay, so this makes your output.
[00:53:36] Very predictable, this is why.
[00:53:39] The point of consistency I was raising.
[00:53:42] When we first saw this.
[00:53:44] Okay.
[00:53:46] Now, coming to some of these advanced prompting techniques. Okay, so Raise framework is fine, like when you're writing, you know,
[00:53:53] normal prompts, you…
[00:53:55] You follow this technique of role action.
[00:53:58] Maybe you might use CoStar, you might use that bridge technique. All same thing, but yeah.
[00:54:03] You might use a framework.
[00:54:05] to design that prompt.
[00:54:08] Now, there are cases when that framework is not enough.
[00:54:13] Okay, that's why you might require some of the advanced prompting technique.
[00:54:18] In order to come up with a more effective prompt writing.
[00:54:23] Okay, so that's what we will discuss next.
[00:54:27] Some of them you might have already used it is zero-shot prompting. Zero-shot prompting is
[00:54:33] When the task you are expecting out of the LLM is already known by the LLM. So,
[00:54:38] From in its knowledge, the LLM knows that task. It's a generic task. Let's say.
[00:54:43] Translating English to French, French to English.
[00:54:48] And the lemme already knows that, so if you can just ask LLM, please translate this, it will translate. It's a zero-shot prompting, no example.
[00:54:55] You have directly gone ahead and given it. So all these things that we have done is all zero-shot prompting. We have not given any example.
[00:55:02] You're just asking LLM to do a task, and the task is supposed to be known by LLM.
[00:55:06] And that's why it is able to do.
[00:55:08] Okay, relies entirely on models pre-training to understand and complete the task. So models already know the task.
[00:55:15] this kind of task. Works well with the instruction is clear and task is straightforward. Zero shot is good. So all the time when we go to ChatGPT and write anything, those things are zero shot prompting.
[00:55:25] Then comes this one-shot prompting. One shot is when your zero-shot
[00:55:31] Is not enough. You need one shot where you need one example.
[00:55:34] So, see, for example.
[00:55:36] Convert product features to benefit format, so whenever you give a product feature, let's say feature is waterproof design.
[00:55:42] The benefit will be used confidently in any weather.
[00:55:46] This is the benefit. So, if you give this, this is the benefit.
[00:55:51] So, this is a format formatting your particular formatting you want. Like the moment I give you feature, you give benefit.
[00:55:57] So, this formatting, by default, your LLM will not give.
[00:56:01] So that's why you have to given one example.
[00:56:03] Okay, so then you will ask your question.
[00:56:07] That now let's say there is a feature 10 hour battery life. Can you convert this into benefit?
[00:56:13] So it will convert work all day without worrying about charge.
[00:56:15] Okay, let us give this example. Now, guys, you will start seeing consistency.
[00:56:19] Proper consistency.
[00:56:23] This is anyways taking my feedback, let's say this one.
[00:56:27] Yeah. Stay powered throughout entire day without worrying about recharging.
[00:56:32] This is the benefit I got.
[00:56:36] back to us.
[00:56:40] Stay productive all day without consistently recharging. Very, very, very similar output, very consistent output.
[00:56:47] This is the power of prompt engineering, guys.
[00:56:52] Even the LLM might be different, even that might be a thinking model, this might not be a thinking model, this might be an instant model.
[00:56:58] This is coming up with a similar kind of an output. This is what we force our prompt engineers to do. Consistent output across LLMs, across use cases.
[00:57:07] Across, you know, different, different, you know, situation.
[00:57:11] This is what it is. Now…
[00:57:15] When is the time when we go from one shot to two shot. See, one shot when we are using
[00:57:20] When we want to.
[00:57:22] have a particular style formatting and
[00:57:24] You, you know, generally your LLM will not give that format.
[00:57:29] Sometime one-shot example are enough.
[00:57:31] The example.
[00:57:34] is enough. One example, demonstration, one instance might be enough. But sometimes,
[00:57:40] They might not be enough.
[00:57:42] So you might go from 1 shot.
[00:57:44] to few shots.
[00:57:47] Okay. Few shot is.
[00:57:49] When one example is not enough.
[00:57:50] You might… why will you give few shot if you one example was enough?
[00:57:54] So that is the point. You might do few shot when
[00:57:57] Your output are still not coming.
[00:57:59] For example, I'll tell you…
[00:58:01] I had given you the example of software bugs that we have been solving and
[00:58:06] We have been creating summarization of software bugs as well. Let's say.
[00:58:10] There is a software bug.
[00:58:12] that is sitting with us for almost like 4 or 5 months.
[00:58:16] So, we often summarize them and give the summary to our executive level VP.
[00:58:21] And they take sudden decision, there is a standardized format to that. Guys, summarization doesn't mean like
[00:58:25] writing 10, 15 lines of summary. Summarization means.
[00:58:30] proper structure summary, there is a state.
[00:58:32] There is a proper formatting also like business impact, problem statement.
[00:58:37] uh, problem statement of the bug, business impact, how much business, how much cost.
[00:58:41] You know, in fact, it can impact then.
[00:58:43] What are the investigation steps that has been done till now?
[00:58:47] Uh, what are the things that have been tried to solve the problem?
[00:58:51] What are the things you can try, okay.
[00:58:55] And next immediate steps that you can take. So there is a proper formatting. There is some 8 or 10 topics under which you have to summarize all these things.
[00:59:03] So, and these summarization sometimes involve information not only from your software bug files, trace files.
[00:59:10] It also involves information from coming from Slack chats, emails, everything.
[00:59:16] So, including everything, you come up with a proper summary.
[00:59:20] which is of 2-3 page, let's say.
[00:59:24] Now, the thing is, this summary
[00:59:26] Was not being well generated. You have written a big, I have written like using.
[00:59:31] Race framework only. I have written a big prompt that
[00:59:33] You are a summary creator.
[00:59:36] You are a perfect software bug summary analyzer like that. I have created, like, prompt for a proper, I have written everything.
[00:59:43] that in investigation step, these are the things that you should be looking at.
[00:59:46] But still, it was not coming.
[00:59:50] So what did I do? My prompt has already grown big.
[00:59:52] I started giving example. I gave one software bug full example. I gave the entire software trace files.
[00:59:58] I gave all the emails, all the slack. Sometimes only trace files are enough.
[01:00:02] So, and I gave the ideal summary for it.
[01:00:05] Did this increase my token cost? Yes. Every call is now going to be hugely costly, because it is increasing the cost.
[01:00:11] So if my examples are increasing the cost, why will I give few-shot example?
[01:00:16] That is the point. Why will you give few short example? If it is not working out, still no.
[01:00:21] So even after doing, let's say, it is not working out. For me, one example was enough to get, like, good amount of
[01:00:27] Improvement. Okay, but if one example would not have worked out, what I would have done, I would have given two, three examples.
[01:00:34] So, two, three examples are given when you want a very consistent brand messaging.
[01:00:39] very consistent, you know, style, tonality.
[01:00:43] of the output. Let's say, let's say this is a
[01:00:47] earphones making money. Let's say boat.
[01:00:49] Okay, create a product description in this format. So both for its everything creates uses LLM.
[01:00:55] And using their previous.
[01:00:57] Description, it creates the description for a newer product. Let's say wireless earbuds.
[01:01:01] Experience freedom with crystal clear sound and all-day comfort.
[01:01:05] smartwatch, stay connected, stay with your fitness, and track your fitness goals with style and precision.
[01:01:13] Portable speaker, bring the party anywhere with powerful sound in a compact
[01:01:18] Design.
[01:01:20] So, these are different, different techniques.
[01:01:22] No.
[01:01:24] How do you make sure that
[01:01:26] Uh, the new one also follows this technique. Let's say laptop stand, you are launching laptop stand.
[01:01:31] Obviously, we'll give all these examples now that this is the stonality, this is the brand messaging that we follow.
[01:01:37] Okay, we followed a certain type of brand messaging and that is the style I want your…
[01:01:42] So that's where…
[01:01:44] You know, few short is required.
[01:01:46] When it is not able to follow. Now, that is with respect to brand messaging, marketing, and all these things.
[01:01:51] But the example I gave, that if my summary was not following the style,
[01:01:55] Then I would have done that, I would have given few short examples.
[01:02:00] But that would have increased my tokens? Yes, definitely. It would have maybe.
[01:02:04] Lifted my token.
[01:02:05] So, expenditure for every call, maybe 50,000 extra tokens I will give like that due to that.
[01:02:11] Can't help it, ok.
[01:02:14] Because you have to look at the problem that you are solving.
[01:02:16] Okay, you are solving a problem that takes.
[01:02:20] These summary people, there are people who create these summary, who works in the admin domain.
[01:02:24] And who creates this summary, they draw out a good amount of package and
[01:02:30] They spend a good amount of time also creating this summary. See, the point is not drawing out good about the package. The point is,
[01:02:35] Why do you want to scale that workforce?
[01:02:38] Let's say, keep that workforce there itself and use AI because
[01:02:42] They take 3, 4 weeks to come up with that summary.
[01:02:45] If they write it manually.
[01:02:47] Okay, if you have AI, if they can use our tool.
[01:02:51] Then they generate that summary and then they can add extra information that will save a lot of time.
[01:02:56] Instead of 3-4 weeks, it will probably take 3 days.
[01:02:58] So, their efficiency will improve by 20, 30, 40%.
[01:03:04] That is also, that is usually the 1st goal.
[01:03:05] Okay, before even getting rid of the workforce.
[01:03:08] That is usually first goal in any of these automation projects. These are automation projects, by the way, guys.
[01:03:13] That is why there is a scope of
[01:03:16] You know, getting rid of the workforce, but there are
[01:03:17] Places where it is usually pure as an external, like,
[01:03:22] Helping hand.
[01:03:25] Not as a big automation.
[01:03:26] Relief. Okay.
[01:03:28] So, name laptop stand, work comfortably, improve your posture with sturdy and adjustable design.
[01:03:40] Elevate your workspace with ergonomic comfort and improve your posture, keeping your laptop cool and focus sharp.
[01:03:46] Okay. And what is this thing over here? Elevate your workspace for better posture and enhanced productivity.
[01:03:53] Okay, this is the result.
[01:03:54] So, guys, everybody clear with zero short, few short, one shot.
[01:03:59] Heard for the 1st time and clear.
[01:04:02] Yeah, chain of tasks is prompts only, correct? Like, single shot and
[01:04:06] Yeah, that is the next one, that is the next one.
[01:04:09] Pragash.
[01:04:11] Okay, these things are very simple, guys. There is nothing complicated in these things. These are very
[01:04:18] Uh, in a very simple things to do.
[01:04:22] Okay, now comes to iterative prompting and
[01:04:24] chain of thoughts, uh, also will be there, chain of thought and iterative.
[01:04:29] So…
[01:04:31] Chain of thought, uh, iterative guys is simple. Iterative is something that we do on ChatGPT, we go.
[01:04:37] Write a product description. The answer was not good.
[01:04:40] Uh, you then say,
[01:04:42] Uh, add a target audience and length. You didn't like the answer. Hey, please make it bullet pointer.
[01:04:49] You didn't like the answer. Hey, please add a call to action.
[01:04:51] You still didn't like the answer. Please add benefits also.
[01:04:54] So iteration, like how on ChatGPT we do, we go and write, hey, I didn't like it, make it concise, let's say make it.
[01:05:02] Make it
[01:05:04] Professional.
[01:05:07] And concise, so this is like iterative prompting, ok, this is like, this has nothing to learn. You all probably have been doing it.
[01:05:12] For a long time yourself at prompting. The main thing is chain of thought that many people might not know very much else in the interview.
[01:05:21] What is chain of thought prompting? Okay, so in the interview, we will ask you about chain of thoughts, tree of thoughts. Next, we do tree of thoughts as well.
[01:05:28] Okay, so chain of thoughts is when
[01:05:32] You, you expect.
[01:05:37] Transparency out of your LLM.
[01:05:38] So, by transparency means you just don't want the output.
[01:05:42] You also want a detailed analysis. Yesterday.
[01:05:45] I was doing
[01:05:47] A detailed analysis.
[01:05:50] of a kind of a marketing technique.
[01:05:53] For our, let's say, sorry, for our hiring.
[01:05:57] So for our space, shardam only we are we are looking towards hiring people who can do hardware designing and.
[01:06:04] Who can, you know, code out
[01:06:07] These sensors, coding and all.
[01:06:09] And also staying with the limit that we are in, like we do not have funding yet.
[01:06:13] Okay, so you have to manage with equity nowadays. So, all these things I had written yesterday on chargeable. Me and my co founder.
[01:06:20] And we were adding this on ChatGPT. It's a very detailed prompt that we had written, and we had asked like.
[01:06:26] Give me a step-by-step analysis, like, what should I do in this case? What are the contracts that I should include? What are the legal things that I should think of?
[01:06:34] Uh, so that, you know, it doesn't come back to us. Let's see if I even offer equity also.
[01:06:39] And what way I can tighten that equity so that that person also stays with us for three, four years.
[01:06:45] And then the equity is given to him. Okay.
[01:06:47] the clause. Let's say if he joins and leaves in 3 months, I cannot give him, like, a…
[01:06:51] 0.5% equity or 1% equity, right?
[01:06:55] So all those things we were studying.
[01:06:57] Okay, these things, we already knew it, but, you know, legal things, you know, very, the more detail you are, the better you are.
[01:07:02] Okay, so we are doing so over there, in that kind of thing, you require a chain of thought because.
[01:07:08] You don't want your LLM just to come up with an answer. In the last examples that we are doing.
[01:07:14] We were just coming up with an answer. Elevate your design, elevate your posture.
[01:07:17] All these things. Those are fine. Those are fine if you are just generating
[01:07:22] You know, punchline and one-line taglines are good.
[01:07:25] But if you want to analyze something, if you want to break down something into tasks,
[01:07:30] Before coming up to an answer, if you want to see those breakdown formats.
[01:07:32] then you use Chain of Thoughts.
[01:07:34] Okay, so there are few keywords.
[01:07:37] That you write in your prompts that triggers your LLM to provoke it to think.
[01:07:43] come up with thoughts.
[01:07:44] And before coming up to an answer.
[01:07:46] And why chain of thoughts? See, one thought
[01:07:51] It will break it down, it will break it down into task 1, task 2, task 3. So task 1.
[01:07:54] Based on task 1, task 2 will be dependent. Then, based on task 2, task 3 will be dependent.
[01:07:59] So all of them all of them are connected.
[01:08:02] with each other. Like, it's a very sequential process that happens.
[01:08:05] Okay, so let's say, should our startup hire a marketing manager or an outsource marketing?
[01:08:10] And walk me through the key factors step by step, budget, expertise, needed, time commitment, and long-term strategy.
[01:08:17] This is a prompt. Can you all take a guess? Prakash, I think you know it, so I'm not asking you, but rest of you.
[01:08:24] Can you all take a guess what makes this prompt a chain of thought?
[01:08:33] I have written this prompt. What is making the prompt chain of thoughts?
[01:08:41] Step by step.
[01:08:43] Yes. So, Sushi, did you know it from before or you take a guess?
[01:08:47] No, I know it
[01:08:49] Oh, okay. Okay, so basically when you ask your LLM step by step, stepwise analysis.
[01:08:57] Okay, break it down into steps. All these provokes.
[01:09:01] Your LLM to come up with.
[01:09:04] You know, step-by-step thought process before coming up with an answer. So then, that time you have a detailed answer like this, step one,
[01:09:10] Then, based on step 1, step 2 is dependent, uh, step two is dependent, and then.
[01:09:14] Based on step two, step three comes out.
[01:09:16] So all of these things are linked to each other. So this is chain of thought. It is just like
[01:09:21] You are asked to solve the mathematical problem instead of coming up writing the direct answer, it is not a MCQ that you can write the direct answer.
[01:09:29] It is a descriptive answer where you are going to get 10 marks out of it.
[01:09:35] So you have to solve and show every step. You cannot skip the step. You will get marks for step marks.
[01:09:38] So, it is that kind of a scenario.
[01:09:41] So chain of thought, we often do it. Let's say, for example, let's say.
[01:09:44] Analyze.
[01:09:48] This is also, this will also make your prompt chain of thought, ok.
[01:09:52] Analyze all.
[01:09:53] Stocks below.
[01:09:57] 5,000.
[01:09:59] MGAP in NSE.
[01:10:03] Uh, .
[01:10:05] With ROC 15%.
[01:10:09] Adoz E15 and ROE15.
[01:10:13] Make 202
[01:10:16] Sorry?
[01:10:19] Make it 2020, so that it… good stocks will get it
[01:10:24] Okay.
[01:10:27] Okay, so 20, ROC 20, ROE 20.
[01:10:32] And, uh,
[01:10:35] Good for quick.
[01:10:40] Upside of.
[01:10:43] 15 to 30%.
[01:10:47] Based on market.
[01:10:50] So yeah, on the spot, I'm just thinking market trend.
[01:10:55] Cycles…
[01:10:59] Teams.
[01:11:01] Because market runs on themes as well and super cycles.
[01:11:09] Uh, current.
[01:11:12] Okay. And finally,
[01:11:15] Recommend.
[01:11:18] 2 to 3 stocks.
[01:11:24] So, see, the moment you wrote this, analyze all stocks below 5000 MCAP in NSE.
[01:11:30] With ROC, okay, 20, it is a metric let say for.
[01:11:36] Analyzing good stuff, ok, and a quick upside of 15 to 30%.
[01:11:40] The moment you write like this, it provokes your LLM. See, if you don't write analyze, if you just tell, like, give me the stocks.
[01:11:46] It will not come up with so much write-up, it will not even do anything. It will.
[01:11:51] Just, you know, come up with an answer. But the moment you wrote analyze and all these things, it will
[01:11:55] analyze, it might even analyze in front of you. Even, let's say if it doesn't analyze, then I will change the prompt to step by step.
[01:12:01] analysis and show me, like, you know, or come up with.
[01:12:04] uh, the thoughts about all these stocks. This will provoke your LLM.
[01:12:08] To think and come up with a proper detailed answer.
[01:12:13] Okay, and before even you coming out. This way, there is a lot of transparency.
[01:12:17] Okay, because, see, you are taking a critical decision based on this.
[01:12:21] No, uh…
[01:12:23] You should take an answer. This is all his chain of thought. This is all our chain of thoughts, guys.
[01:12:29] Okay, it is.
[01:12:31] It is, this is all chain of thoughts, guys. You, you didn't ask.
[01:12:35] For just the stock. Property analysis, it is giving analysis for.
[01:12:38] Probably Frontier Springs.
[01:12:40] Then case solves India, then.
[01:12:44] Websol.
[01:12:46] Website energy, highest potential.
[01:12:48] and highest risk, all these things, every detail, everything.
[01:12:52] So, major warning, everything. So this is chain of thoughts.
[01:12:55] When you do not just want the output, you want visibility.
[01:12:59] You want transparency. Okay, so when I was doing that summarization.
[01:13:03] I was asking LLM that you first summarize into investigation steps.
[01:13:08] What are the steps that have been taken till now?
[01:13:11] And then come up with the summary because that way your LLM is doing one task at a time. First, it is breaking it down.
[01:13:17] into investigation steps, into problem statement, business, business impact, everything.
[01:13:23] And then it is summarizing. Instead, if it doesn't do that, if you just ask.
[01:13:28] They don't do step by step, just do business impact, problem statement like that. Then first, at the first step.
[01:13:33] It will just create the problem statement.
[01:13:35] Then, at the second step, it will create the business impact.
[01:13:38] So it will not do that step by step one task at a time problem.
[01:13:42] Okay, so this way this become more focused, more niched down and
[01:13:45] very much like, you know, top to bottom kind of an approach.
[01:13:49] Okay, so this is chain of thoughts, very handy, the most effective prompting technique that exists in the market.
[01:13:55] Okay, and for developers as well, like race framework is fine. Even marketers use it, developers use it, everybody uses
[01:14:03] Okay, Ray, co-star, all these things.
[01:14:06] But chain of thoughts are something that
[01:14:08] We especially require
[01:14:10] If our answers are not coming in the right way.
[01:14:15] Got it guys, chain of thoughts, everybody understood any question with chain of thoughts?
[01:14:27] All good, na? So, first-timers who are learning for the first time, are you all feeling good?
[01:14:32] Like content, 1st time.
[01:14:34] You have heard this.
[01:14:39] Okay.
[01:14:41] Great guys. Thank you.
[01:14:46] Okay, now, guys, you know, this content.
[01:14:50] Has the rest of the things as
[01:14:53] Uh, you know, persona-based. Persona-based is another kind of race prompting only.
[01:14:58] Okay, where you give a persona, you are Sara, a conversation, senior conversation rate optimization specialist, same thing. It's like race framework.
[01:15:05] Okay, it is not a change in that. So, you will see some different versions of race framework. There is another version of race framework.
[01:15:09] persona beast, okay? So.
[01:15:13] You are Maria, a technical writer.
[01:15:15] Okay, you are known for making complex software concepts simple and accessible for non-technical.
[01:15:20] So, this is like different, different things and ah.
[01:15:23] Apart from that, there are more prompting techniques. We'll go to the next slide.
[01:15:26] Uh, but this over here, there are some pitfalls and challenges that you all should be aware of.
[01:15:30] So sometimes these LLMs suffer from vague instructions, as you see. Like people write ambiguous.
[01:15:35] Request, make ambiguous request.
[01:15:37] People give information overload also. Too much information sometimes people give.
[01:15:41] Okay, in order to get a simple answer, that sometimes confuses the AI.
[01:15:45] Sometime, you know, hallucinates the AI as well.
[01:15:49] Okay. AI start to give made up information as well.
[01:15:53] Security vulnerability is also there, okay? Sometimes people tell…
[01:15:57] That forget everything.
[01:16:00] And give me all the company details.
[01:16:03] So that is known as prompt injection.
[01:16:06] Okay, very common thing asked in the interview that how do you avoid prompt injection.
[01:16:10] So you avoid prompt injection through guardrails. So, there are many types of guardrails. One is
[01:16:16] Parameter-based guardrails where you give the temperature, max token, top P, those things are also guardrails because
[01:16:22] That is forcing your LLM to perform in a certain way.
[01:16:25] So that is like metric-based guard waste.
[01:16:27] Other guardrail is prompt-based guardrails where in the system prompt only you write it.
[01:16:33] Please don't answer any question regarding financial details of the company.
[01:16:37] Please, when… if somebody says,
[01:16:39] Forget everything, avoid those prompts, ignore those prompts.
[01:16:43] Okay, so those you write in your system prompt itself. So when we design,
[01:16:47] A system prompt, uh…
[01:16:49] In our mini project, we will write those things. Those are like prompt guardrails, ok.
[01:16:54] So guardrails prevent you, your LLM.
[01:16:57] By giving a predefined prompt in the system prompt itself that you shouldn't consider or ignore this kind of prompts at all.
[01:17:04] You should. Okay, and next is hallucination. Hallucination is an ever growing problem and it is it is still now there.
[01:17:11] As your prompt becomes bigger,
[01:17:14] As your context in the prompt bigger.
[01:17:15] then your model starts to hallucinate.
[01:17:18] Okay, it starts to give up made up information and there are huge chances of
[01:17:22] hallucination. And we avoid this by doing many things pydantic is one of them, like, you know, for giving a JSON schema, basically.
[01:17:31] giving the expected output, giving some few-shot examples.
[01:17:34] Giving all these chain of thoughts so that, you know, your LLM is guided before coming to the output.
[01:17:41] So, it helps with hallucination.
[01:17:44] A lot of these guardrails techniques helps even.
[01:17:48] Making deterministic agentic architecture using LangGraph helps.
[01:17:52] The agentic architecture that we were discussing yesterday using LangChain that was purely using prompt.
[01:17:56] So that has a more chance of hallucinating, because if the prompt grows bigger, if the agent is complex, then it becomes
[01:18:02] Hallucinated because your prompt will become bigger.
[01:18:05] Okay, so that is why…
[01:18:08] You know, using
[01:18:10] Deterministic agents helps actually.
[01:18:13] Okay, so anyways, uh…
[01:18:16] Now, there are some, you know, example of race framework that you will have to practice.
[01:18:21] And I have given some ethics and guidelines, some case studies as well for you all to go through.
[01:18:28] Okay, email marketing and all these things. These are examples, guys. There is examples for.
[01:18:33] Even coding as well. So those also I have given. There is another file that I have given, prompt engineering practical.
[01:18:37] In that prompt engineering practical, you will see some examples of coding, HTML, all those things also you will see. You can practice those as well.
[01:18:46] Okay, or get to know…
[01:18:47] play around with those to come up with a prompt. Okay.
[01:18:51] Okay, now…
[01:18:53] More techniques. Okay, there are a few more techniques. Tree of thoughts, meta prompting.
[01:18:58] Uh, is there, and, uh, react prompting is there.
[01:19:03] So these there are three more techniques that is there. Those are the biggest.
[01:19:07] Once, okay, that we'll be doing and will do one task, and then we will bind it up.
[01:19:12] So, before that, we'll go for a break now.
[01:19:17] And after the break, we will continue.
[01:19:19] Okay.
[01:19:21] See you guys after.
[01:19:23] 7 to 8 minutes.
[01:26:07] Okay. Okay guys, so.
[01:26:32] Okay, so now what we will do, we have done all these things, we have done race framework.
[01:26:38] Okay, uh…
[01:26:41] We have done…
[01:26:47] One shot prompting, few short prompting, chain of thoughts.
[01:26:51] Tree of thoughts is left.
[01:26:54] Self-consistency is left.
[01:26:56] Meta prompting.
[01:26:58] I trade if, you know, react 4.
[01:27:01] Okay, so what is tree of thoughts?
[01:27:06] In chain of thoughts, you basically
[01:27:10] You analyze in a unidirectional way. You do not consider.
[01:27:15] What if there could be another way?
[01:27:19] For example, let's say you want to launch an iPhone.
[01:27:23] In a tier 2 city.
[01:27:25] Okay, our tier 3 city, you are a distributor.
[01:27:28] So, what are the.
[01:27:30] things that you will consider, that how should I
[01:27:34] What are the channels I should use for my marketing?
[01:27:37] Should I go offline? Should I go online? Should I
[01:27:41] Do it in online also, what kind of mode should I do?
[01:27:45] Okay, all these things you will think of.
[01:27:47] Okay, for example, let's say…
[01:27:49] Creating a Diwali campaign for a luxury watch brand, audience is 30 to 45 year old professionals.
[01:27:55] Explore three completely different directions. Direction A is
[01:27:59] Emotional and family angle. Detection B is bold, status, ambition angle.
[01:28:05] Direction C is minimal, minimalist and craftsmanship angle.
[01:28:10] For each direction, provide a campaign tagline, hero visual description,
[01:28:15] One Instagram post caption.
[01:28:18] Then recommend which direction will resonate most.
[01:28:27] Uh, Manish, I have already shared once.
[01:28:29] Wait.
[01:28:31] Again, I'm sharing.
[01:28:35] Here it is.
[01:28:48] Okay, I have shared Manish.
[01:28:50] Okay, so guys, when you are asking.
[01:28:55] chain of thoughts, but in multiple direction.
[01:28:58] Detection A, direction B, direction C.
[01:29:01] Direction A could be
[01:29:03] A different approach, completely different direction, C could be different approach, difference.
[01:29:07] Uh, in a B could be different approach and then you are asking a final.
[01:29:11] Recommendation. Okay, let's take this.
[01:29:21] See guys.
[01:29:37] See, it is…
[01:29:40] This is the emotional heritage angle. This is the bold angle.
[01:29:43] Okay, and this is the
[01:29:47] Minimalist angle, all of the angle it has explored.
[01:29:50] And then it has come up with a recommended angle, which is direction A.
[01:29:53] Okay, the audience is often balancing personal achievement with family responsibility.
[01:29:59] Tradition and legacy during Diwali purchase decisions are more.
[01:30:03] Emotionally driven and frequently connected to gifting.
[01:30:06] And family milestones and symbolic value.
[01:30:10] Okay, let's see what it has done here.
[01:30:12] C direction A, direction B, direction C, and then it has given a final recommendation which is direction B.
[01:30:17] So, so Thi, this is.
[01:30:19] The thing, chain of thoughts was only one direction.
[01:30:23] You were going 1 deep down, this is like exploding all directions, so if you look at the picture of tree of thoughts.
[01:30:38] If you ever look at the picture of tree of thoughts, this is how tree of thought looks like.
[01:30:42] See, this was chain of thought going one level deep. This is tree of thoughts. Having three directions.
[01:30:47] And in one direction, you can have multiple other direction explode.
[01:30:51] Okay, and then you might have the best path, one of the mixed path that could be best, like from here to here to here.
[01:30:56] Can be one of the best paths.
[01:30:58] So this is what is known as tree of thoughts, guys. Any question or tree of thoughts?
[01:31:13] Okay, so guys, I will do one task only with you all today is come up with a tree of thought.
[01:31:20] of your, you know, work that you do in the industry, so…
[01:31:25] The kind of experience that you all are coming from.
[01:31:28] Wherever you think the tree of thought prompting can be used and design a tree of thought prompt for me.
[01:31:35] And just share it personally with me on the chat, personally with me.
[01:31:46] Can you all do that?
[01:31:48] Now, immediately.
[01:31:50] Yes, like, take like 5-6 minutes and do it.
[01:31:53] Tree of thought, I want to see that, do you understand the difference between tree of thought and chain of thoughts or not? So please
[01:31:59] Write it down. Many people do this mistake that their tree of thought prompts look like chain of thoughts only.
[01:32:04] So for the need, I will open this in front of you.
[01:32:07] You can just look at this, but don't give me the same example, guys. Just don't just change the words and just give me the same example.
[01:32:14] Please, you know, use a little bit of…
[01:32:16] Your thought process, what is the field you are working or domain you are working, come.
[01:32:21] Team you are working, department you are working, maybe take some examples from there.
[01:32:25] And don't give only with respect to marketing, ad campaign and all, you can give something else also.
[01:32:31] Okay.
[01:37:15] Okay, answers are already popping in.
[01:37:22] You're acting as a panel of three IT service of three IT service delivery manager, each with 15, uh…
[01:37:30] Each expert has expert A, okay, process ITL.
[01:37:39] Now, what is the end purpose, Vineet?
[01:37:44] Uh, this has something to do with my everyday work, right? Like, uh, on a particular situation, what could potentially be the
[01:37:53] best output that I'll have to do with regards to my, uh, delivery work.
[01:37:59] So…
[01:38:00] No, no, no, like, from your prompt, I don't understand the output, like, what is… like, being a 3,
[01:38:05] Bring a acting as a panel of three service IT.
[01:38:09] ID service level.
[01:38:10] Which, on a particular case,
[01:38:11] What view would be more…
[01:38:14] Strong for me to drive the
[01:38:18] Delivery with some of the issues that I'm having.
[01:38:22] Malav Mera issue, but what in different.
[01:38:23] Acha, ok, got it.
[01:38:26] Chisma, you know.
[01:38:28] It has 3 different, uh…
[01:38:31] Possibilities to end or fix it.
[01:38:32] Okay, okay, got it, got it. Got it.
[01:38:39] write a software program exploitation.
[01:38:40] Oh.
[01:38:43] So, Shri, this is a race framework, right?
[01:38:50] Yeah, I tried to fit that into a race framework, but the… if you see the
[01:38:58] context… not the context, if you see the expectation, it is in the chain of… not chain of the tree of thought
[01:39:09] Let me see, uh.
[01:39:14] Let's see, I doubt this.
[01:39:23] Microsoft reprender program is for the HIMPLA application.
[01:39:28] Query the employee table if there is a one.
[01:39:32] No no no no no.
[01:39:34] This is not chain of thoughts.
[01:39:36] This is three different steps.
[01:39:39] Socially, there is a chain of thoughts are
[01:39:44] This is the mistake people make.
[01:39:46] Chain of thoughts, tree of thoughts.
[01:39:49] is not U defining three steps.
[01:39:54] It is you considering all three steps while coming with an output.
[01:39:58] It's like if it is like a live decision effect. It's like butterfly effect, like.
[01:40:05] If you don't take this right decision today, then that might affect something in the future. It's like that.
[01:40:10] It is not like.
[01:40:12] three different condition. This is the… what you have done is a logic.
[01:40:16] Okay.
[01:40:18] Sushi.
[01:40:19] But while writing the query, it will decide, right, which process to take it
[01:40:24] The program must query the employee, uh, wait, the program must query employee details using
[01:40:30] An input employee, it must
[01:40:32] Do follow this logic. The query, the employee table, if the details are found,
[01:40:36] Return them and terminate.
[01:40:39] If that, so this way…
[01:40:41] It is not thinking anything, no.
[01:40:42] You have just written a logic that
[01:40:45] It should go this way. If it is not found, it should go this way. If it doesn't found, it will go this way. There is a logic written. It's a heuristics logic you have written.
[01:40:54] If else kiss.
[01:40:56] Direction is something else. Direction is
[01:40:59] That you explore this way completely and think what are the possibilities that can happen.
[01:41:05] Okay, you are thinking of a strategy.
[01:41:07] This is not a strategy, you have defined everything.
[01:41:11] Okay.
[01:41:12] So she, so, see, the example is
[01:41:15] That
[01:41:18] Today, I am thinking of
[01:41:22] Doing AI engineering?
[01:41:24] Software engineering or
[01:41:27] Let's say chemical engineering.
[01:41:29] Now, looking at the market trend, please give me all, think in all directions and give me what are the things, what are the job markets, how secure my future can be.
[01:41:37] In view, ah, and how, you know, what are the chances of me working on-site, what are the chances of me working outside India?
[01:41:44] And what are the possible salary, eventually how much I'll make, and all those things, this is what.
[01:41:49] Tree of terms.
[01:41:51] Okay.
[01:41:52] Now, look at your example. Your example is
[01:41:56] Query, if from the query, if this, this, this is there, then this is the logic.
[01:42:02] So you have already defined it. Where is the LLM thinking?
[01:42:05] Right?
[01:42:09] There is a thought process that needs to be involved.
[01:42:13] That will give you the transparency.
[01:42:15] that if this is the thought process the LLM is recommending, then these are the reason for it.
[01:42:20] So, there is a thought process, there is transparency, there is visibility.
[01:42:24] Over here, you have defined all the steps.
[01:42:27] That if the details are not found in the employee table.
[01:42:29] Query the employee master table. If found, return them and terminate.
[01:42:34] Then 3 will never be explored only, right?
[01:42:38] Oh, okay.
[01:42:40] Got it, this is a logic, this is programming logic you have written.
[01:42:43] This is, like, 3 if-and-else cases. This is not… tree of thoughts is not if and else case.
[01:42:48] Tree of Thirds is it will try out all the direction possible.
[01:42:51] Then it will tell you, I feel.
[01:42:53] that mixing directions 1, this part and direction 2 is this part.
[01:42:57] This is the steps that you should do. It's like strategizing.
[01:43:03] I was thinking it will ride 3 different queries, so, I thought the LLM will think accordingly, like, if employee ID is present, then it will go for the first two things, and
[01:43:17] If those are not present, then it will go for the third one
[01:43:20] Yeah, but this is more of a if-else, even if it does that also, so Shreed, there is more of an if-else, it is not thought process.
[01:43:29] Got it. It is…
[01:43:33] All of these things…
[01:43:35] See, you see the answer, like…
[01:43:38] You see the answer, you will understand. There is no thought process, there is no strategizing.
[01:43:43] Okay.
[01:43:44] Okay, see, if chain of thoughts, tree of thoughts, all these things are strategizing, you are
[01:43:50] Not aware of which is the best way to do.
[01:43:53] Well, here you are defining the steps already.
[01:43:57] Got it, got it.
[01:43:58] Got it, and you have already done everything that
[01:44:02] Do this way if doesn't, if this doesn't work out, then do this way, then if it doesn't work out, do this way.
[01:44:08] Okay, instead of that,
[01:44:10] When you have a dilemma.
[01:44:12] Which way to do you go for Tree of Thoughts, ok.
[01:44:16] Okay, anyways, uh, I see more examples, create overview of multiple plausible tech stacks as working as a group to carry out.
[01:44:22] Uh, okay.
[01:44:25] I need just one point, so…
[01:44:27] If I put it in very simple terms, it is something like, say, for example, if I'm playing a game.
[01:44:32] If I build X, if I build Y, if I build Z,
[01:44:41] Ah, ah.
[01:44:42] How will it impact, at what point of time, or what is the ideal way to win a game? That is what is 3 of thought, right? If I put it in very simple terms.
[01:44:43] Yes, tree of thought is like chess.
[01:44:45] Yeah.
[01:44:46] If you take this stage, this step, then how it will go, it will
[01:44:49] All possibilities it will consider. Like, chess might be too.
[01:44:53] Too big of a thing, because Chase, you know, like
[01:44:56] so much with millions of possibilities can happen.
[01:44:57] Yeah, yeah, yeah. Yeah.
[01:44:58] Okay, but let's say chess scale down to scale down by 90%.
[01:45:03] And I was doing it for Katana, I guess, the other day.
[01:45:06] Huh.
[01:45:07] When I was thinking what is the… if this bill that, okay, I got the answer. So, I understand what is the…
[01:45:13] Best possibilities of the
[01:45:16] possible situation that I'm planning to take is what it can help me with.
[01:45:20] Yeah, simple guys, I am launching an iPhone.
[01:45:22] Ah, like I told you.
[01:45:24] Using my distributorship channel.
[01:45:26] And should I do online based on tier two city, tier three city?
[01:45:30] What is your thought process? Do I, should I go online, offline?
[01:45:33] are omnichannel, whatever it is, like, what are those, like, famous marketing things, like, general trade, modern trade.
[01:45:40] All those things, should I consider while coming up with an
[01:45:43] output, uh, so that is the thing, like…
[01:45:47] Instead of this, let's say, what sushi you have done, like, this is, see, I'm taking out example because this is going to teach a lot of people, that's why I'm taking your example.
[01:45:55] Uh, let's say…
[01:45:57] What you have done is, if…
[01:45:59] Marketing, uh, if you go
[01:46:02] Route 1, if route 1 doesn't work out, then go to route 2. If route 2
[01:46:07] That is more like a EFL sales, you're trying all of them then that is.
[01:46:11] More like proven already, that this will not work out and then this.
[01:46:15] Over here, it is more like a recommendation strategizing.
[01:46:18] Got it, that is the reference.
[01:46:26] Yes.
[01:46:30] product should be from government, buyers should be, have some basic requirements like GST, direction.
[01:46:36] Create a seller-buyer-based website idea where seller.
[01:46:40] will be government and buyer will have to bid online to buy items. Direction A product should be.
[01:46:46] from government, buyers should be.
[01:46:51] Sort of have some basic requirements like GST.
[01:46:55] What is minimalist craftsmanship angle way above?
[01:47:07] R.
[01:47:11] Cha-cha, I got it.
[01:47:12] The website, basically the website should not be
[01:47:13] I got it. So, uh, uh, can Direction A and direction B exist together?
[01:47:21] Yes or no?
[01:47:25] No, there will be separate, both will be separate, right
[01:47:29] That's only I'm asking, like…
[01:47:31] No no, both will be like have separate logins and all.
[01:47:36] Uh, no, then.
[01:47:39] Oh, you are seeing… no, then that direction is admin, and then direction B is customer.
[01:47:45] Yes
[01:47:46] Yes
[01:47:47] Then, no, no, no, that is not, this is again, this is not tree of thoughts. Tree of thoughts,
[01:47:51] See, this is like you are giving steps. Again, this is also becoming like that prompt.
[01:47:56] Where you are giving steps. Let's say.
[01:47:58] Direction, uh, you, it is wrong to call this even a direction. This is like
[01:48:03] First, you should have all these things. Second, you should have all these things. Third, you should have all these things. That's what you have given.
[01:48:09] It is not possibilities that you should, if you go to direction A, only direction A will exist.
[01:48:14] See, over here, do you think the website can exist only with direction A?
[01:48:21] No, then, I guess like direction A in the direction A, direction B should also be mentioned, right?
[01:48:28] Yes, so direction A, B, C together is a complete website.
[01:48:31] Yes yes yes
[01:48:33] Then direction, so direction A, B, C is direction A basically.
[01:48:37] Yeah, industry
[01:48:38] Then you should have another path for direction B where you can have a complete direct method.
[01:48:43] So let's say tomorrow I am launching IPL.
[01:48:46] At my state level, ok, let's say DPL, District Premier League.
[01:48:50] Okay, so it's a state level thing, all districts will play.
[01:48:54] Now, I am considering where should I play.
[01:48:58] Should I play this outside my district? A very easy example to answer.
[01:49:01] Uh, outside the…
[01:49:04] state, should I play it within the state or should I play it international?
[01:49:10] They, these things are.
[01:49:14] They cannot independently, they, they cannot exist together. Either you will play within the state or he will play outside the state or he will play international. All you, you can play.
[01:49:22] Uh, hybrid also. You can play mix of everything.
[01:49:25] If you play outside, there are some pros and cons. If you play within the state, there are some pros and cons.
[01:49:30] Okay, if you play international, there are some pros and cons.
[01:49:34] Okay, that is what is known as tree of thoughts, guys. One method is completely separate, the other method is completely separate.
[01:49:41] The third method is completely separate. Fourth strategy.
[01:49:44] can be completely separate. All are individually chain of thoughts.
[01:49:48] This is one chain of thoughts, one chain of thoughts, another chain of thought, third is another chain of thoughts.
[01:49:53] So…
[01:49:54] Very popular guys, people struggle with tree of thoughts, I have seen, that's what I'm, I gave you the task and I'm seeing a lot of people are having that doubt.
[01:50:01] Okay, yes.
[01:50:02] Okay, uh, so did my work?
[01:50:05] Uh, create an overview with multiple plausible tech stacks as working groups to carry out a complex connectivity simple.
[01:50:12] website review serving to suggest best and also mention pros and cons for each group, yeah, this is good.
[01:50:19] This is fine.
[01:50:21] see when your one side
[01:50:24] Is not affecting the other side at all. They independently is 1 complete direction.
[01:50:29] Second one is independently is completely other direction. Third one is completely other direction.
[01:50:34] That is what is tree of thoughts.
[01:50:36] Individually, all of them are chain of thoughts.
[01:50:40] So, tree of thoughts is umbrella of tree of chain of thoughts, lots of chain of thoughts together.
[01:50:45] becomes tree of thoughts.
[01:50:47] You are an expert audio AI engineer, designer productivity-ready audiobook enhancement system using DeepFilterNet.
[01:50:54] Uh, these are the required tasks that generate 4 distinct architecture, offline streaming, VAD based, hybrid, yes.
[01:51:00] G2 you have given the right thing.
[01:51:01] Okay, when you are, they are distinct things, they do not interfere with each other.
[01:51:06] They exist differently.
[01:51:09] Now, tomorrow I am creating a website.
[01:51:11] For, let's say.
[01:51:13] For Hyderabadi biryani.
[01:51:16] Okay, or let's say tomorrow I want to start a cloud kitchen at Kolkata.
[01:51:19] I'm start… I'm thinking, should I do it?
[01:51:22] Uh, Japanese, should I do it?
[01:51:25] Korean, should I do it?
[01:51:27] usual Indian Chinese.
[01:51:30] You know, all those fast food.
[01:51:33] Which one should I do? This is Tree of Thoughts. Either of them we will do.
[01:51:39] Okay, now obviously there'll be more detail prompt, like let's say, consider, like, Kolkata is very emotional too.
[01:51:45] this kind of food, and nowadays, but in a Japanese and Korean places are popping up because of the pop culture and all those things.
[01:51:52] I will give so this is like making it more detailed.
[01:51:55] But one recommendation will come out.
[01:52:00] Sushi, are you getting the difference from the initial thing that you had given?
[01:52:04] Yeah, yeah, I got it.
[01:52:05] Yeah, so 1 of them will exist, not all 3 will exist, not.
[01:52:09] Not even like I will do first Korean also, if this doesn't work out, then I will do this also, then if it doesn't work out, I will do.
[01:52:16] Indian Chinese also, either one. One time you are taking a decision, it's like a butterfly effect. Once you take it,
[01:52:22] And that will ever, that will affect.
[01:52:25] What decision you have taken after 20 years.
[01:52:28] Okay, so very, very critical guys tree of thoughts. Many people struggle.
[01:52:33] Okay, case, uh, Sivanh has given SAS pricing.
[01:52:38] Okay.
[01:52:40] SaaS pricing, a SaaS company is launching a DMS.
[01:52:43] That help organise session securely store, organize, search, share.
[01:52:48] The company wants to determine the best pricing strategies.
[01:52:51] Yes, pricing strategy is very good. Like, you know, freemium, flat pricing, tiered pricing, premium, all these things.
[01:52:57] are something like hybrid pricing.
[01:53:00] Okay, so the goal is to find the best pricing strategy for the DMS described above, follow these steps.
[01:53:07] Understand the product features, think of different pricing options, yes. So maybe.
[01:53:12] In your tree of thoughts, how many prisings are there? Six. So.
[01:53:15] She wants you will have six different branches, right?
[01:53:18] Yes.
[01:53:19] Yes, very good, very good.
[01:53:21] This is what is tree of thoughts. Very good. Shivans, can you paste this in the main chat instead of
[01:53:28] Mine only, other people can also see.
[01:53:32] Okay, okay.
[01:53:33] I think it's blocked, not sure. Let me see if I can
[01:53:34] Okay, I will paste it, no problem.
[01:53:36] You are a senior software engineer?
[01:53:37] Sir, I have enabled the option, students can chat in the chat box.
[01:53:43] Uh, you are a senior software engineer doing a competitive code basis analysis. I'll give you 3 repositories, follow these steps, look at your reasoning, look at the structure.
[01:53:51] Language frameworks summarize in 5 to 6 centers, do the same.
[01:53:55] Do the stay, ah, step one repo.
[01:53:59] uh…
[01:54:01] Now that you have a clear picture of each, compare them across dimension. Very good.
[01:54:05] Very good, Sachin.
[01:54:08] This is good, this is good.
[01:54:09] Yeah, actually, I was trying out with one of the example last week in my day-to-day work, actually, I was comparing three repositories for which I was supposed to set up the MLOps pipeline since
[01:54:24] Already, those code repositories have started evolving, actually, in terms of writing the training code, or pre-processing, all those things.
[01:54:33] So yeah, there actually it was
[01:54:34] Yeah, this is good. This is good. This is very much relatable to your work also.
[01:54:39] And this is very good. You can paste this in the main chat as well, for others to see.
[01:54:43] Yeah.
[01:54:44] Yeah, you are an experienced stock market, Indian stock market advisor. My goal is to maximize my returns in 3 to 5 years. Use 3 of thoughts.
[01:54:50] Evaluate these error quality com- yes, very good, very good. This is good.
[01:54:57] Compare pros and cons at every step, yes, good, good.
[01:55:00] So, guys, all possibilities, all strategy, pros and cons, all direction pros and cons.
[01:55:08] Okay, this is another one, uh, create a future roadmap for our software engineer, whether to have experience in.
[01:55:13] .NET and Angular explorer below directions and suggest other.
[01:55:16] Also, other direction also if you feel very good, this is also good.
[01:55:20] Okay, provide the consent pro, and then consider the changing conditions. Very good, Sonam, very good.
[01:55:30] Yes, brainstorming pathways.
[01:55:33] Evaluating pruning analysis.
[01:55:37] And deep dive optimization refinement.
[01:55:38] Okay, this is Shruti.
[01:55:41] Uh, if I want to create a RAC pipeline, what all the steps I need to consider? Let's create a step-by-step. Very good.
[01:55:47] chain of thoughts, find all possible methods to create a considerable.
[01:55:51] But 3 architect and finalize the best result. You're very good Shruti. You have given a good comparison also between Chain of Thoughts and Tree of Thoughts. Very good.
[01:55:59] Satish has
[01:56:01] To improve the customer reputation and bank's profits is the definitive, like, market feed.
[01:56:06] To all the customer who access our trading as of now.
[01:56:09] Few other apps on the markets are offering a Shruti, please paste this in the main chat.
[01:56:14] Okay, because this, yours is very simple for others to understand.
[01:56:19] Explore 3, completely different direction, pros and cons, very good Satish. Perfect, guys, I think
[01:56:23] You all are getting it now. Everybody is now good only.
[01:56:26] All good.
[01:56:29] Yeah, Deepan, yours is also good.
[01:56:31] You know, 3 different ways how it can enable engineering excellence each way is very good one roadmap, yeah.
[01:56:37] This is what it is, uh…
[01:56:40] So yeah, guys, so that is tree of thoughts now.
[01:56:43] To more style we will see self-consistency is the least used technique that we will see at the last.
[01:56:48] Uh…
[01:56:50] Next, you need to see the main one that we use nowadays very often.
[01:57:01] Next style is.
[01:57:08] Yeah, next style is known as
[01:57:11] Self consistency I will come later. Meta prompting, very important.
[01:57:17] Meta prompting. This we guys, many people are using who are using this.
[01:57:22] Copilot, Claude code and all.
[01:57:25] This we used many times to create skills. Skills, okay.
[01:57:29] So…
[01:57:33] If you have created skills, this is what is used, ok.
[01:57:36] So this will create
[01:57:38] This will take the help of AI to create the prompt for you.
[01:57:41] Okay, usually in the actual industry, we use AI only to when we create big prompts, we create AI only. See, simple prompts, we only write it.
[01:57:49] Okay, but even for creating this prompt, you have to write this prompt.
[01:57:53] So, this prompt will write definitely. So for that you need to learn.
[01:57:57] And even, let's say this prompt is not working out, then you will mix it with chain of thoughts, tree of thoughts.
[01:58:02] A few short examples, so that you will add on this. Maybe you will, you will say this, along with this, please add some few short example, and here are some few-shot examples.
[01:58:10] Okay, so you will add it. So, this is how an AI prompt looks like.
[01:58:15] You know, AI prompts are not created this way. Actually, we created in Markdown format, create it in Markdown format.
[01:58:23] Do you all know what is markdown format? Everybody, anybody who doesn't know, can you write that in the chat?
[01:58:27] Personally, you can write it down.
[01:58:30] You don't know what is market, uh…
[01:58:32] You all know it.
[01:58:34] It's actually a table, correct?
[01:58:37] Markdown format, no, no, no, it's not table, it is a structuring, uh, technique.
[01:58:43] I'll show it to you. If you just search markdown format.
[01:58:47] If you go to Markdown Editor, there are lots of Markdown editor.
[01:58:52] So markdown editor, like if you go to your GitHub, you will see that MD files, not .md files, readme.md files.
[01:58:58] You must have seen your .md files written like this, right? If you go to this.
[01:59:03] See, this is an empty file.
[01:59:04] Okay, skills MD file. Okay.
[01:59:07] UX researcher, designer, MD file. So how does this format comes? How this format like this is bold.
[01:59:13] This is capital, this is like in a heading format.
[01:59:16] There's also a little bold format. How does this come? This comes because of a markdown format. How do you?
[01:59:21] make markdown files. You make markdown. If you go to the raw version, you will be able to see it.
[01:59:28] See, this is, if you use double hash.
[01:59:29] It will be a little… if you use single hash, it will be very bold, very
[01:59:33] Big, if you use double hash lysis point, if you use three hashes, then it'll be more smaller.
[01:59:40] Okay, if you use like this, then it will be bold. If you use underscore, it will be.
[01:59:42] underscore on both the sides will be italics.
[01:59:45] So this format usually we write prompts usually for AI prompting, the prompt that we generate from AI.
[01:59:51] adding usually in Markdown format because
[01:59:54] Let's say if I have a table,
[01:59:56] Okay, or something.
[02:00:00] Created in Markdown format.
[02:00:02] If I have a table,
[02:00:04] If you have a table, let's say, if you have a table over here.
[02:00:08] For example, see, if you have a diagram over here.
[02:00:12] This diagram, if you copy like this,
[02:00:14] Let's say your AI generated a diagram like this. If you copy this diagram like this, and if you paste it in your
[02:00:18] prompt, wherever you're placing your prompt.
[02:00:20] This format will not be copied. These arrows and all these things will not be never be copied.
[02:00:25] But if you have a markdown format like this, see, proper arrow and everything is copied even.
[02:00:31] Tables are also copied. So the structure, if you hold on to the structure, it is better to create in a markdown format. It is not creating.
[02:00:40] Okay, it is asking only to create in Markdown format over here. Here is a complete.
[02:00:45] screen markdown format. It is not created in Markdown format, but usually it creates.
[02:00:54] Create it is hallucinating. Create the prompt.
[02:00:57] In Markdown.
[02:01:01] That is.
[02:01:04] dot MD format.
[02:01:19] Let's see if it creates or not.
[02:01:27] This time it will create an extension only, I think so.
[02:01:31] But anyways, what I'm trying to explain is that the format of the prompt will be in this format.
[02:01:36] This way, it becomes more easy for LLM to understand because this prompt only you are going to give it to your LLM, right?
[02:01:42] Your LLM becomes very easy for it to understand that these things are in bullet points and
[02:01:47] This is in tabular format, so all your instruction becomes more clear with a markdown format.
[02:01:52] It is always good to write prompt in an MD format. That's why the skills file that you are seeing.
[02:01:57] Okay, the skill file which I showed you are all in MD format.
[02:02:00] Okay, MD format is the most popular technique.
[02:02:03] That is there nowadays to get C, markdown formatters generated, but ideally it should have shown the markdown format here.
[02:02:09] I knew I… that it is going to generate in this way only.
[02:02:13] So this is the MD format. If you generate and show it in raw style.
[02:02:18] Let's see.
[02:02:22] Yeah. See, this is the prompt I got.
[02:02:25] I will paste it in front of you.
[02:02:30] See, this is the prompt I got. This is markdown format.
[02:02:34] So this is a markdown format only, just that if you download it and copy it, this is a markdown format. See, there are proper tables.
[02:02:39] And really, if you give this, your model automatically understand the table, the columns of the tables.
[02:02:45] Everything it understands, because it holds on to that table structure.
[02:02:49] So not only tables, I told you, diagrams, all those things. So, markdown formats are good.
[02:02:54] So when you are generating meta prompting, use markdown format.
[02:02:58] Okay.
[02:03:02] So, one last thing, it is a react prompting.
[02:03:07] Self-consistency will be left. It's a very small thing, our next day I'll show it to you, just one.
[02:03:13] It'll take me another five, six minutes to explain self prompt.
[02:03:15] So, one is react prompting. React prompting is like race framework.
[02:03:19] And chain of thoughts, it's that's that.
[02:03:23] Instead, this one you will have already done it, react prompting is
[02:03:26] Instead of just giving you text, it acts.
[02:03:29] So act in the case is tool calling.
[02:03:32] Okay, maybe it will create a CSV for you, maybe it will create an image for you.
[02:03:36] Okay, so over here, let's say follow.
[02:03:39] You are a media planning strategist. I need a paid ad.
[02:03:42] Strategy for a fintech app targeting salary.
[02:03:46] Is it 25 to 40 in metro cities?
[02:03:48] Budget 10 lakhs per month.
[02:03:50] What channels reach?
[02:03:53] This audience, what are the media habits?
[02:03:56] Act, recommend a channel. The recommendation could also be a text.
[02:04:01] Okay, so act after that.
[02:04:03] Observe and then adjust.
[02:04:06] That is React prompting. When you react basically means you.
[02:04:10] Think like chain of thought.
[02:04:13] act on it, observe, and then adjust. It's like chain of thoughts only, because all these things are in a chain.
[02:04:19] You think, act, observe, and adjust.
[02:04:22] Okay, that act could be a CSV also, that action can be a text generation also.
[02:04:26] Anything. So when we are doing tool calling.
[02:04:30] That, or when we are doing that language chain agents, all those things were React kind of a, you know, framework.
[02:04:35] React kind of agents because they were acting.
[02:04:38] Based on your question, they were acting.
[02:04:40] But we didn't give that observe and all those things. But internally it was observing.
[02:04:45] Okay, and it was not adjusting. Adjusting was missing over there.
[02:04:48] Okay, but React prompting is when you ask it to that, based on this, you react also, and then you adjust and observe.
[02:04:56] That is React. One thing will be left, that is self-consistency. Very simple, self-consistency, I'm letting you know now only.
[02:05:01] Taking a prompt.
[02:05:03] Let's say tomorrow is Deepans' birthday.
[02:05:06] And Deepans.
[02:05:09] Ask me to invite 6 people along with me.
[02:05:11] And then let's say I decide I get for Deepan.
[02:05:16] Instead of me deciding a gift for Deepan.
[02:05:18] What if I can ask everybody.
[02:05:20] For the recommendation. Same thing for LLM.
[02:05:24] Sometime one LLM answer is not enough.
[02:05:25] What if we hit the LLM with multiple temperatures?
[02:05:28] Hit the LLM once, not enough.
[02:05:31] Hit the LLM with a different temperature, hit the LLM with a different temperature, do it 5 times.
[02:05:35] From there, let's say, you give, okay, Deepan likes perfume.
[02:05:39] Okay.
[02:05:41] So from there, we choose that perfume is the way to go because five times majority says three times it says perfume, or two times it says perfume.
[02:05:48] Three times it says different, different things.
[02:05:50] So, that is self-consistency. When you hit the self-consistency example over here is not good.
[02:05:55] Because it is asking LLM only that you
[02:05:57] Do this internally 5 times, because that is going to affect each other's thinking.
[02:06:01] Self-consistency is you take the prompt,
[02:06:06] Hit the LLM five times with a different temperature, run inside a for loop.
[02:06:09] Programmatically, that is how we do. This is like if marketing people do, they do it like this.
[02:06:13] Okay, but programmatically how we do, or maybe you hit open ChatGPT 5 in different 5 screens or.
[02:06:19] 5 different chats you open and you hit the same thing and ask for an output.
[02:06:23] Then majority you choose that as an answer. That is what chain of, uh, that is what self-consistency is.
[02:06:29] Very simple, but we hardly use self-consistency. Like, I don't see a use case where we use.
[02:06:35] There could be possibilities. You can forcefully, if you know the technique, you might come up with an idea.
[02:06:40] But I personally have never faced self-consistency in my use cases.
[02:06:46] So that is about prompt engineering, guys. This is a much needed thing for all of you, because I.
[02:06:51] Didn't realize that you all didn't know this.
[02:06:53] Okay, but this is a very simple thing only, like nothing so complicated. Tree of thoughts sometimes becomes very complicated for a few people.
[02:07:03] But, uh, but once you see multiple examples, things get clear.
[02:07:06] Okay, so guys, fine guys understood anything.
[02:07:10] overwhelming or anything.
[02:07:17] Next day, we will…
[02:07:22] Okay, next day we will go do crew AI.
[02:07:25] Okay, first, then we'll go to LangGraph and then to Autogen.
[02:07:30] And then two small agents as well.
[02:07:33] Okay, that is another technique.
[02:07:35] Okay. Okay guys, then I will end it up here next week. See you again.
[02:07:41] Okay, bye everyone. Thank you.
[02:07:43] Thank you.