--- prompt: |- Anand delivered a talk / workshop on AI - Context Engineering. Saurabh introduce him. Debi helped read out the questions. The audience asked questions in the last part. --- # 2026-05-23 IIM SG Alumni AI Unboxed Context Engineering These are the chats used to prepare for the workshop: - [IIM Singapore Alumni AI Workshop Plan](https://claude.ai/share/635c2e28-8a7d-4c9e-9dfe-0b3aae3fcd62) - I've been invited to run 4 AI workshop sessions for IIM Alumni in Singapore. Some are senior. Some are junior. Search online to get a sense of the profiles. It's mostly finance and banking professionals but take a look. Then, suggest the duration of each workshop and topics for the 4 workshops, knowing my style, past work, and what'll be useful for them. Give me options to pick from, with reasons. - I'm thinking: 90 min each, with these topics: Context engineering, Agentic analysis, Vibe coding, AI strategy. What do you think? - OK. I'll use these 4 then: 1. Context engineering. 2. AI tools & workflows. 3. Agentic analysis. 4. AI strategy. - [IIM Alumni Singapore workshops content - ChatGPT](https://chatgpt.com/share/6a11a0be-d43c-83ec-9f84-70a4939b11ff) - I'm running four 90-minute workshops for IIM alumni in Singapore on the following topics: Context engineering. AI tools & workflows. Agentic analysis. AI strategy. QUESTION: What should I cover in each session? (Audience research, Local MCP, ...) - [IIM Alumni Singapore workshops content - Claude](https://claude.ai/share/2f54eab9-4d97-4b23-b5fa-5b79ed368d20) - I'm running four 90-minute workshops for IIM alumni in Singapore on the following topics: Context engineering. AI tools & workflows. Agentic analysis. AI strategy. QUESTION: What should I cover in each session? (Audience research, Local MCP, ...) - Here's another answer from ChatGPT. Take what's better, drop what's worse, explore any new thoughts this leads you to, and revise your response based on that. - [IIM SG Alumni Workshop Context Engineering Prep](https://chatgpt.com/share/6a11a1df-77e8-83ec-bd16-259cd51b35a8) - I am preparing for the IIM Singapore Alumni session, the workshop that I will be delivering in a short while, about context engineering. I will ask you a series of questions. I am dictating this, so there may be transcription errors. Use past conversations of context, factor in phonetical errors, give me crisp responses that I can read while bicycling, but think hard and research extensively before answering me. Let me know when you are ready for me to ask my first question. - I want to show how they can use copy-paste as a mechanism to provide context in a non-trivial way... - Based on my past conversation, are there any other cool tricks that you have seen me use where I have uploaded content that would fit in either of these categories? - Based on all my blogs and past conversations and any other information you have or can find about me, what are the various ways in which I am providing context and therefore doing context engineering that is beyond what most people do? - Explore a 4 layer approach: Prompt storage, Export digital exhaust, Agent consumable content, Agentic tooling - Create a 5x5 matrix of components and actions - Look, it doesn't have to be a five by five. In fact, the less there is to fill or the less there is to cover, the more value there is. Do you think we would actually be better off with fewer than five elements on the rows and or the columns? - Is there a different way of structuring the columns? What are possible alternatives and which do you think would be the best and why? - I prefer prompt, files, memory, and tools. Any other options that I absolutely should add? - Now, fill this in for me, specifically, how am I doing each cell? - Give me the collection of the most powerful techniques in each cell. - Suggest audience exercises. Links - [IIMPACT AI Unboxed Teams Background Design](https://chatgpt.com/share/6a114e9b-0c6c-83ec-b9df-0f2f15149325) - [IIMPACT AI Unboxed Teams Background Design - Gemini](https://gemini.google.com/share/f15ef277be83) - [IIMPact Logo](https://files.s-anand.net/images/2026-05-17-iimpact-logo.avif) - [Nayana Jain blog post](https://www.s-anand.net/blog/people-skills-with-ai/) - [Argue against my blog posts and find errors](https://claude.ai/share/b7b0ab70-9235-4f2a-997e-1118c449b90b) - [Public Profile of Anand](https://chatgpt.com/share/6a118465-e9cc-83ec-9684-c04f199a7157) - [Memory test on Gemini](https://gemini.google.com/share/f92a1b18ac8c) - [Prompts](https://github.com/sanand0/blog/blob/live/pages/prompts/) - [Skills](https://github.com/sanand0/scripts/tree/main/agents) - [LLM Pricing](https://sanand0.github.io/llmpricing/) - [LM Arena Leaderboard](https://lmarena.ai/leaderboard) Forms - [AI Unboxed - Context Engineering](https://docs.google.com/spreadsheets/d/1jCIpMGJ8o1yGPjVL7W4XUSiwtaeNzhKZe4APrOjX1GA/edit) - [AI Unboxed - Context Engineering - Exercises](https://docs.google.com/document/d/1vyOTJFvTtJyRsJNMyMSv302dizTIe4MuDNVHCA5reGw/edit) ## Transcript **Saurabh**: [00:00] Good afternoon and welcome, everyone, to the AI Unbox series. We started this in December last year where we did a session talking about Chinese internet. We had one of the Chinese VCs present, we had a couple of Chinese founders, and to round it up, we had Anand running us through a workshop. **The response was tremendous, and there was a real need for everyone—people said we should do more of that.** **Saurabh**: [00:33] In this whole attempt, here we are. Anand has been kind enough to spare us 8 hours of his time, flexible, over four Saturdays across four months to run four different workshops. So, we're looking forward to a really exciting time. Just some house rules: **This session is being recorded.** Everybody who’s registered will receive a recording after the session. The way we're structuring it is one and a half hours with Anand. Any questions, please post in the chat box here, and then we have 30 minutes which we're dedicating to him addressing specific issues and things that people might have. **Saurabh**: [01:21] And just thank you to the team organizing this. Debi Guha, of course, my partner in crime, she's been the most active proponent of this workshop. Kishore for managing the backend—we did try Zoho's solution, but came back to Teams, but this platform works well. So, Kishore, thank you for managing the backend. And Nana Jain, of course, for all the creativity. All the lovely posts that she puts across—**my latest prompt in Claude is "write like Nana Jain" and it comes up with these posts.** So, that seems good. This is not—over the year we plan to have many more sessions with Anand and then we have other ideas because, you know, we want to keep the whole AI theme going for most of the year. So, thank you for your participation and let's make this as much fun and as much interactive as you want. As Anand said to us, it will go in the direction where people want it to go. So, Debi, over to you. She's the host for the day, and with that, Debi, the floor is yours. **Debi**: [02:29] Okay, I'll just take one minute of everyone's time to introduce Anand. I don't know how many of you were present during the December session, but we're extremely privileged to have Anand. He is one of us, he's an IIM Bangalore alumnus. I don't know how many people know, but for some reason, **he was called "Bhalla" in IIT Madras**, which everybody needs to find out why he was called that—I'm still yet to figure that out. And then in IIM Bangalore, **he was called "Professor" as well as, instead of S. Anand, "Stud Anand".** If you look at the impressions of the yearbook that IIM Bangalore has, I think the heading of his section says **"God of all things."** **Debi**: [03:22] I think more than anything else, I've got to know Anand over the last few years after he moved to Singapore. What surprises me is his intense curiosity at several things. I think one of the things he had mentioned to me, and I later read in one of his blogs, was that **his love is music, books, and maths.** So, I'm trying to figure out that I think maybe it's the maths part of it, together with the music part of it, which has now moved him towards AI. This is an area where he's obviously spending a lot of time. He's a founder of Gramener, a company which he set up after his stints in Infosys Consulting and BCG, and it was perhaps one of the first companies from India which started creating data stories. So, over to you, Anand. Anand has been very generous with his time during December as well as volunteering to spend four Saturdays with us. So, I would encourage all of you to open up with any questions that you have, post it on the chat, and Anand is more than happy to answer them. Thank you, and over to you, Anand. **Anand**: [04:34] My daughter asked me, "Why do people need an AI session? Can't they figure it out themselves?" Very valid point. What she didn't add was, "And too, why do they need an AI session from you of all people?" But here's the thing: I was talking to Bradley Chamber, who runs the design and AI sessions at SUTD the day before, and some of the teams at NUS yesterday, and the question was more or less around the same thing. On the one hand, there is a bit of an AI backlash; on the other hand, there is a "how do we teach AI?" and the educators are all thinking, **"If people can just read stuff by themselves, what are we doing as faculty?"** **Anand**: [05:22] Something interesting happened, actually. I was attending one of Bradley Chamber's sessions at SUTD—I taught for 15 minutes in his session, and for the remaining 45 minutes, I was attending his session. It was interesting because that was an amazing session in how you use AI in design. I went through the same set of slides that he presented later on; that magic was not there. They were good slides, but what he presented, when he presented it, the experience was very different. I'm thinking that **the value that these workshops add is probably not as much that I am going to give you some content that will be useful. A big part of the value is you have decided to dedicate a certain amount of your time to do something related to AI.** That you're sitting along with a whole bunch of other people who are also thinking along those lines. That environment is probably going to provide a whole lot more value. **Anand**: [06:23] Therefore, to everyone who's going to be watching the video, please keep in mind that the environment adds a bit of value beyond this. See if you can put yourself in that kind of an environment—I don't know, group watching maybe, whatever, to get that kind of benefit. I will be using the recording, chat transcripts, etc., for derived output, some of which I'll be distributing to all of you. So, please keep that in mind when you're asking questions on the chat or whatever. We will be keeping the questions to the chat window initially, that is the first hour and a half till 4:00 PM Singapore time, and then we can get onto voice discussions. But until then, please keep yourself muted. Sandeep, I'll request you to please mute others in case there is any accidental voice conversation that happens. **Anand**: [07:18] We will be using a Google Form to run some of the interactions, especially for you to—this is a workshop, you'll be doing the work—for you to share the output of a few questions. The sheet that we will be using, I have just put on the Teams chat. We will be using the Teams chat for some of the ongoing communications. Your questions can certainly go into—not Teams chat, there is a separate Q&A tab that you will find. It's probably best if you put your questions on the Q&A tab. But I have one question for you which I will share my screen and explain. It's a simple question: **What paid AI do you have?** Everyone who's registered will find your email ID here. Please try and find your email ID, it's sorted alphabetically. Don't change any row other than your own row. We have track history enabled, so we should be able to revert it, but it just helps if you don't mess it up too much. The answer can be just a comma-separated list of, like, Claude and Perplexity, or Claude, Perplexity, or whatever. **Anand**: [08:41] But the key question is what *paid* AI do you have, and the paid part is important. **Why is paid important? The difference in the quality of models is enormous—just take it from me. But that is only 70% of the answer. 30% of the answer is if you're paying for it, you probably will use it more. If you're not using it enough, pay more.** It's the reason why I guess we hire a BCG or a McKinsey consultant, or arguably any luxury item, but it has an effect. You're taking it seriously, just like a gym fee, right? So, this session requires you to have—"requires" is a strong word, you won't be able to do some of the things that I will be talking you through unless you have a paid version of either Claude or ChatGPT. An institutional paid version, personal paid version, doesn't matter; any paid version is fine. In India, there is a "Go" version, ChatGPT Go, that might not suffice for some of the things, but any paid is better than the free version. Take it for a one-month subscription at the very least; long-term subscription is entirely your call. And if you have an office version, that's perfectly fine. **Anand**: [09:59] As we go through this, since we have about 76 people, I'll typically wait till about half the people have responded. And we have about 50-odd people who responded, that's more than half. I'll continue, the others just go ahead. Sushmita, you're not able to access the sheet and got that you don't have a paid version. Now, what we're going to do is look at some of the features in these. Now, for the exercise that I'm about to suggest, you will need either ChatGPT—paid or unpaid might not matter—or Gemini. Both of them will support what I'm about to share. Like I said, free versions of these are fine. **Anand**: [10:42] What we're going to do is something that I just did a few minutes before this session. You see this background on Teams behind me, right? **That background was created by AI.** And it's a fairly simple prompt that we used. All I did was upload the IIMPACT logo, in this case to ChatGPT, in another case to Gemini, and asked it to create a Teams background. The full prompt is this, but really you can use any prompt you want. I said, "Create a background, my style," and I'll come back to the "my style" bit in a while. But I also added, "Think about what will work as a great Teams background, keep the IIMPACT logo on the top right, blah blah blah." But really all you need to tell it is "Teams background using this logo" and it will do something. **Anand**: [11:45] I would love to get a screenshot on Teams with all of us coming on with this background, or at least a page full of us expressing our creativity, turning on a background that has IIMPACT. I'm sure it'll do great buzz on social media with AI-generated backgrounds and all of that. But you need the logo. The logo is at a URL I will put in the chat window. I will also place it on the form here. The exercise is "Create a Teams background with" and I'm pasting the logo here. You should just be able to click—okay, we are not allowed to click here, and then click here to open the logo. How do you use it? Well, one possibility is just right-click, copy the image, and put it into ChatGPT. Just paste it here. Or you could download the file. **Unsure**: [12:51] **Oh, can you just keep it for a while on the screen? I'll take a screenshot.** **Anand**: [12:55] I will. You may just want to put that on the chat because it's going to be less disruptive. And yeah, anyone who wants to take a screenshot, please feel free, that works too. However, **the exercise is intended to get you practiced in uploading files.** So, for those of you who are taking the screenshots, please download the file and upload the file. Then the other exercises that we'll do are of a similar kind. **Anand**: [13:25] The crux of it, and I'm going to switch screens now, is this plus button. All the agents have a plus button where you can add files and photos. **Today's session is about context engineering, and context engineering is largely about making sure that you've got stuff that you can upload—that you *actually* upload stuff.** That's not all of context engineering, that's about a quarter of context engineering, but it's an important and probably the biggest part of context engineering. Once you know what to give an agent, it will do a better job because it has information that it needs. You'll notice this plus button across Gemini as well as Claude, so here we are. This is the request. **Anand**: [14:14] Now, Gemini is pretty fast. ChatGPT is much slower. Claude, the last I checked, cannot create images, so stick to one of these. Once you have the image, you can share the conversation. How do you share the conversation? In ChatGPT, on the top right in almost every case, you can click on the share button. That will allow you to copy the link. In the case of Gemini, this is what it looks like: on the top right there are three dots, and it says "Share conversation," and you can copy the link once it has created the link, and you have it in your clipboard. You can paste that copied link into the form. I'm doing that in my case. The link that I have used to create the Teams background that you see here is the first response. **Anand**: [15:13] Your task is to go ahead and create this particular thing and let it run in the background. We are going to continue talking, doing a whole bunch of other things, and that's part of what I'm realizing with AI sessions: **You don't wait for the AI to get work done. You tell it to do stuff, come back, give it another task, tell it to do something else, and then go back and we will check on these.** Towards the middle of this session, once we have approximately, well, at least 25 of this column filled, I'll request all of you to download your favorite image and use that as your Teams background. Please put your chat here in this column, please don't put it on the chat window. **Anand**: [16:05] Okay, that—okay, we have nine in this, that's okay. About half a dozen people who already managed to—Danny, if it's you that pasted this, can you please delete that and just put in the link? It will be easier to just have the link. If it was someone else, yeah, just please just put in the links here, not the actual image. Those are really nice images, of course; we would love to see it as your background. **Anand**: [16:37] Now, what we're seeing here is—and I'll stop sharing my screen for a little bit and just talk. What we did here was very simple. We took a task. You probably said something along the lines of "Create a Teams background with this image" and possibly gave it the image. And behind the scenes, it used a capability. It probably would not have directly created that image; it would have thought a bit, it would have called an external tool and gotten the job done. In my case, I also told it one extra thing, which is "use my style." And that's because it knows all of the things that I have chatted to it with. **Anand**: [17:31] **These constitute what I am now framing as four elements of context engineering.** The reason I say I'm framing is because nobody—this is not official theory, I'm just coming up with it. Actually, I'm not even coming up with it; ChatGPT came up with it during my cycle ride back from the police station. Now, in all of these cases, the stuff that I'm about to tell you has two flaws. One, it's just my theory, one person's opinion, so take it as—take it with a pinch of salt. Second, it is ephemeral. What is true today, given the pace of AI development, will not be true tomorrow. So, don't worry too much about remembering it. **What you *do* today is what will stick in your fingers, in your memory; that's what counts.** But for what it's worth, let me go through some theory because it helps a little bit in structuring things. **Anand**: [18:25] So, there are broadly, as I see it, four kinds of elements to context: **Prompts (what we tell it), Files (which is what we upload), Memory (which is basically all of my past conversations), and Tools.** ChatGPT has been trying very hard all day to get me to use better workshop labels. I've told it very clearly, "No, I understand this, I don't understand all of your fancy stuff. We will stick to just what I understand." Also because there is a clean mapping with the tools. Prompts are what you type. Files are what you upload. Memory and tools are slightly hidden. Not very much, but tools are simply stuff that you give it access to. One of the ways you give access to tools is by clicking on the plus. For instance, in ChatGPT, I've clicked on the plus, there is a "Create image" option. When I click on "Create image," I'm giving it access to the image tool. There is a "Web search" option. When I click on "Web search," I'm giving it access to the web search tool. **Anand**: [19:35] Now, what is the big deal about context engineering and these tools? **Tools are basically telling the agent, "Look, apart from what I'm telling you, you do some work by yourself."** Creating an image is a reasonably simple piece of work. Web search is a good example where, if I want it to do something instead of me doing the web search, it can do the web search on my behalf, and that is where agency comes in. We are delegating that part of the work. And web search is actually a fantastic tool. **Anand**: [20:12] So, here's another exercise that I would strongly suggest you try out. This is to research yourself. And you can use Claude for this, ChatGPT for this, Gemini for this—any paid version. Some of these have a web search option, specifically ChatGPT has a web search option, the others don't, but in none of them do you need to specify web search; they will all do a web search by themselves. What I'm going to do is research myself. The prompt is very simple—I'm going to dictate into ChatGPT, and the reason I'm going to dictate—you could type this, I'll tell you why I dictate in a short while. **Anand**: [21:01] "I am Anand, head of innovation at Straive and former co-founder of Gramener and an IIM alumnus. Search online and find out everything that is known about me in the public." **Effectively, this is researching myself. This is an important thing for each of us to do because this is what everyone is doing about you.** When a vendor comes to meet you, when a friend is meeting you after a very long time, when somebody's investigating you for whatever—potentially for a new job, HR—this is the new way in which it's going to happen. And you may as well know what currently AI is going to say about you. The prompt is "I am so-and-so" and give it some context so that it knows what to search for. **Anand**: [22:00] The better way to do this will be to upload your CV. But—actually no "but," so I'm going to upload my CV. Where is my CV? Somewhere in Dropbox. I have this old CV, but this folder will probably have—okay, let me find my CV. I have it somewhere in Dropbox. And we are going to get into this in a short while, which is: **How do you make sure you organize your files so that you can find them?** This is not where my CV is. Okay, I moved it to—okay, find my latest CV. It's a PDF. Yeah, here we go. I have my latest CV from 2024. So, I'm going to take that and drag and drop it. Drag and drop is a pretty easy way to upload stuff. And I'm also going to tell it, "My CV is attached." By now, you know what the next exercise is. You're very welcome to get started. **Anand**: [23:25] But just research yourself. Once the chat is complete, then please share the research about yourself—no, all the stuff about yourself. "Here is my research about me," let's just say "Research me." The prompt is broadly "Research me." Give this a shot, please share what is publicly known about you. It's publicly known anyway, and we may as well know each other, and it's an interesting way of introducing each other. Incidentally, what you can do, and I'm certainly going to do, is take the however many responses that we get from this, consolidate all of those, and put them into one big catalog and say, "Here's what we know about the people that attended this session." And you should try that with your teams, you should try that with a bunch of people. **The output of prompts are also context, and that feedback cycle is a pretty important one.** **Anand**: [24:30] We'll come back to that in a while. Let's see, we have 36 people who have created backgrounds. That's great, I needed only 25 for the entire screen to be filled, but it's so nice to see three or four screens getting filled. So, do see if we can get up to 50 at least. Anyone who has not managed to create a Teams background, please just share what problems you're facing in the chat. **Anand**: [25:03] Now, what have we seen so far? We've seen that there are prompts, which are the instructions that we give AI. There are files that we upload—I just uploaded my CV, we all just uploaded an IIMPACT logo. Memory we'll come to in a bit. We also saw a little bit of tools. Web search is a tool. And what it's doing here is doing very detailed searches online. It's searching for Singapore LinkedIn, it's searching my website, it's searching Gramener's blog—Gramener is the company that I co-founded—it's searching for me in Analytics India Magazine, blah blah blah. **The more you're present online, the more it finds. The more somebody's talked about you online, the more it finds.** And you'll find that there are quite surprising things about you that you didn't know that were out there. **Anand**: [26:01] Which, of course, allows you to start looking at how you tailor your presence online. Because a follow-up prompt after that could be, "What negative news is there about me?" And then, of course, we can always go to what do we do about it. But this, incidentally, is a prompt that I've not tried. I should try it. So, okay, "Grammer as Gramener," fine. It's smart enough to have figured out the typo that I made during transcription. Blah blah blah. I'm not going to read this yet. I'll talk about this later. What I *am* going to do is share this and put it on the chat, so you will have at least one entry populated and, for what it's worth, you will also have a sense of the prompt that I'm using. 90% of this prompt is really useless. All it's saying is "research me, I am so-and-so," possibly upload your CV, don't upload your CV, doesn't matter; it should be able to find you. But whatever somebody else is going to search for about you is what we're looking for. I'm going to copy this link and paste it in the form that we're using so that you can check this out. I'll also just modify the wrapping so that it runs and we don't have overflow. I will continue with this question, which is "What negative news is there about me?" But so far, we're not doing anything advanced, anything sophisticated, no issues or no complications or no great learning so far. **Anand**: [27:54] Other than that, maybe, look, the search capabilities are a little more sophisticated than you think. Because what it did during the search is not a naive search for just my name. Let's take a look—yeah, let's take a look at the kinds of searches that it did. Once it figured out my company name, it started searching about my company's acquisition. That is slightly clever, it has a little more—it's trying to get more context. It has figured out that I have a GitHub link and it searched for me on GitHub and my other social media platforms. So, what are the videos that I'm publishing? What are the posts that I'm putting on Reddit? I didn't even remember that I have a Reddit account, and yes, there are a few posts from a long time ago out there, and who knows what people have commented about it. Searching for me by role instead of by name, and that's an interesting twist that it has. Public details, looking for more sources. Okay, it's searching for me to see if I'm present on some influencer lists—I don't think I am. **Anand**: [29:05] Yeah, but the point is this: **this is roughly like giving a research assistant or a detective a task: "Research me." And they go sit and research, research, research, use their reasonably good brains, and come up with something that is consolidated. Increasingly, we should stop thinking of ChatGPT, Gemini, Claude, as tools and start thinking of them as agents. I literally think of them as professionals.** I think of these as my auditor, my researcher, my financial manager, and so on, because it's a reasonably good mental model to have—that is whom I'm substituting. Keep this in mind, keep this at the back of your head. We are going to go back to the fourth item: memory. --- **Anand**: [00:00] We spoke of tools; I've told you about only one tool, which is the web search. We will be diving deeper into this tool in the next session, where we are going to be talking about how research and arguably analysis can be almost entirely taken away by these agents. But the—yeah, tools part we won't be, and we'll be extensively covering tools in a future session. But memory, we will touch upon a little bit right now. **Anand**: [00:30] **Memory is, among other things, where we allow it to access our past chats.** These past chats can be compartmentalized with projects, but I'll just tell you how you enable it. Depends on the tool. If you're using Claude, for instance, you go to settings and in settings, somewhere in preferences, or customization—customize—no, not customize, I think it is in preferences, somewhere where these are. It keeps moving, and I enabled privacy—yeah, shared chat, memory preferences. So, in settings, in privacy, I find it today under memory preferences. In ChatGPT, it will be there somewhere else. **Anand**: [01:21] And you can click on manage. **What I do is turn on almost everything.** Occasionally they'll toss in new stuff that I have not discovered, like today there's a connector discovery. Yeah, just go ahead, discover connectors. Connectors are very powerful; we will be touching upon it maybe a little bit today. But yeah, I just make sure that everything is enabled. Keep in mind that you do not want to treat me as an example to emulate. A lot of my information is public. You may not be comfortable, and Sonal as you mentioned, yeah, it's perfectly fine if you skip any of these options or sharing what you're not comfortable with. But I am trying to see how far I can push things. **I've given up on my privacy anyway post-social media, actually pre-social media. So, I just enable all the options including training, because my hope is that if it can be trained to my preference, then for free I've gotten a multi—well, now a trillion-dollar company potentially—to do a little bit of my bidding on my behalf, which is always fun.** **Anand**: [02:40] Once you have this enabled, you can do even more interesting things, like, for instance, you can ask it a personality profile. So, now I'm combining two elements of context engineering: **Prompt plus Memory with a simple prompt.** I'm going to change models—I'm going to tell you why I'm changing models in a few minutes—but for now, I'm going to switch to Sonnet 4.6 Adaptive and ask it: "Based on all our past conversations, if you had to define the top three traits that are unusual about me that most people would consider negative traits, what would they be? Don't spare me. Just be totally ruthless about it because that is what will really allow me to improve myself." **Anand**: [03:32] Now, let's run this prompt and see how it's working. Okay, based on—oh, okay, that was quick. Well, it has accessed the memories, but I—okay, we'll come to this. One, it says I collect insights the way others collect trophies but rarely bleed for them. Okay. Oh, I just talk about stuff, I don't commit. I'm addicted to—why am I washing dirty laundry in front of you? The point is that you can use it to research yourself. But here's one of the things you would have heard of, which is hallucination. **And the thing about hallucination is that AI sometimes gets stuff wrong, just like people sometimes get stuff wrong. And I cross-check it exactly the way I cross-check people, sometimes.** **Anand**: [04:24] For instance, in this particular case, I'm going to say, "I want you to point to the specific conversations that you have in your memory as evidence and support each of these claims with as many pieces of evidence as you can, ideally citing verbatim from those so that I will be able to go to those chats, copy-paste what you've typed, and find that exact piece of text." Now I've told it how I'm going to verify it, what verification I want, and this will give me some confidence. Is it actually doing this? **Anand**: [05:01] So, now it is looking for the relevant chats. It's found relevant chats. And now I have a little more confidence. Last time it just thought, and it probably had a sense of who I was, blah blah blah. Now it seems to be really searching for relevant chats, and the good part is I can click on these chats. I'm not going to click because it just might jump to a different window, but after it finishes, I am going to click on some of these chats that I deem safe to share publicly. And we will, yeah, see. Okay, so here is stuff that is linkable. "Skill.md conversation", "Accountability conversation". **Yes, it seems to be literally citing some text that I believe if I were to open—so let's right-click on this, open this link in a new tab—it opens a new Claude conversation, and yes, this conversation exists.** So that gives me a little more confidence. And it asked me to find fundamental errors in your own post, but notice the format: I commissioned the critique from an AI, not from a peer. So it's basically that way. Let's not again go into what my flaws are. The point is the verification method. **Anand**: [06:15] What we did here is took the memory capability, the fact that it has the ability to search through past conversations, and it really went through several conversations and is giving me evidence about the personality. **So now I have a compounding asset. The more I chat with it, the more it knows about me.** There is a lot of bad about it potentially; there is also a lot of good about it, and this is one of the ways in which I can leverage it. Because now I can start asking it questions like "How do I improve?" or, of course, it could be the other way, which is "What are my strengths? Therefore, how do I leverage it?" or "How do I ask you better questions? Where are situations where I asked you a question and you didn't give me good enough of an answer?" Now that you probably have smarter models, go back, review those, and see how we can improve. Lots of things, but the core of what I wanted you to take away from this section was **use memory—it is there. Enable it. You can always disable it.** And how do you disable it? In any chat, you can click on the top right to use an incognito mode. ChatGPT and Gemini have their equivalents where that particular chat is in incognito mode and the memory from that chat is not used. The chat is still saved sometimes—maybe not saved, I don't know—but if it is saved, you can always select the chat and delete it. So it will be permanently deleted; that's one option. It will not reflect in any of the memories or use any of the memories; that's another option. You can turn it on and off. I suggest you turn it on by default. **Anand**: [08:02] What I'm going to do now is take a few questions from the chat window once I do a quick check to see—okay, we have 13 people who have researched themselves, nice! Please do research yourself and share what you find. And 44 people who have backgrounds! So after I answer a few questions, I'm going to request a team photo with as many backgrounds as we can. But let's go to the questions. I will move the questions up here and... Anyone found the memory option in ChatGPT? And actually, let me find the memory option in ChatGPT. **Anand**: [08:40] I will go to settings, personalization maybe? No, that's—yeah. **Under settings, personalization, memory, you can turn on "Remember across chats" and "Improve the model."** There you can also, in ChatGPT, select "Manage," and here are the recent memories. "Had an Italian Tuscan meal, bread with olive oil tasted very good." Yes, that is true; I don't know why it thought this was so important that it should remember it, and yeah, whatever. **Anand**: [09:21] Next question was in Claude—yes, I promised I'd answer that, Vijay—which models do I select? Sonnet, Opus, etc. **Claude consumes tokens a fair bit. So I, if I'm in a workshop and I'm worried that I might run out of tokens, I use Sonnet. If I'm not worried and I'm not using it that much, I use Opus.** I would suggest you start with Opus. When it says, "Oh, you've used all your tokens for these five hours, you can only try after two hours or three hours or some such thing," at that point you would have realized that maybe you need to switch to Sonnet for that kind of a use, or just wait a little bit. Short answer: **Always use the best model possible.** And if you find that—you will almost never find that you don't want to use the best model. **Anand**: [10:19] With the sole exception of—Sandeep, could you figure out who is unmuted and please mute them? Because they seem to be on another call that's cross-talking. With the sole exception of Claude.ai where you might want to switch to Sonnet if you find that with Opus you're running out of your limits very often. How do you change this in ChatGPT? Very similar option. There is a drop-down here where you can configure which model you're using. Choose the best model. When new models come out, it usually auto-upgrades because for OpenAI—for all the providers—it's usually cheaper and for you it's better. But best to keep an eye out here. In ChatGPT there is also an instant versus thinking option, and even in thinking there is a standard versus extended. **I live on o1 thinking extended.** ChatGPT takes a reasonably long time to respond, but long time is what? Three minutes? I can wait that long for a really thoughtful answer. If I really need something quickly because I'm on a dinner table and I need to figure out whether Vijay became Chief Minister or not, then I'm switching to instant and that's okay. **Anand**: [11:41] Same with Anthropic. You can select an adaptive thinking toggle. Adaptive thinking means that if required, it will think for longer, and it's generally a good idea to turn it on. I am yet to find a situation where you would not turn it on unless you really want a quick answer. Sorry. **Anand**: [12:00] Next question: can I copy the prompt into the chat? I will do that, Rajan, absolutely. Thank you for flagging that. Sonal says can the file link be shared again? I'll request Sandeep, Debi, one of you could please share the Google Forms link once again. That would be great. **Anand**: [12:30] And Vikas says on enterprise versions they don't keep too much memory beyond the current thread—is that correct or is that a settings issue? I don't know, Vikas; it's entirely possible that they are controlled by organization settings. Two things happen: a) the providers themselves have different defaults and settings; second, they allow the organization administrators to control those settings. For instance, on my Straive work Gemini, which is a Google Pro account, I cannot share—not even internally, let alone externally—but I think it lets me access memory. So I'm actually going to run an experiment. These questions are how I learn, actually, so thank you, keep the questions coming. And I will ask it: "What can you tell me about myself based on our past chats? Go through all the chats exhaustively and cite verbatim from those so that I have evidence that you have looked at those chats and if possible give me the links." Let's run that. I—okay, I don't have access to your past chats. **Anand**: [13:37] But if I took exactly the same thing and pasted it into my current free Gemini—I don't know, might or might not work. I moved away from the paid version of Gemini three, four days ago, keep turning it on and off. Ah, but this seems to be a little more promising. Personalization in progress. Okay, yeah, yes. Some of this is definitely—it could not have figured this out even from the internet. So, in short, to your question: **It depends partly on whether it's configured by the provider, partly based on how it's configured by the AI admin.** **Anand**: [14:30] Thanks, Debesh, for sharing. And Abhinav asks, "Can you show how to aggregate memory across multiple providers?" This is a very interesting question, Abhinav, and also a hard one. If you could please remind me to show you this in the fourth session where we will be building tools, I will be showing you how you can build a tool that does it. But for now, what I will do is show you a weaker version of how I'm doing this. **Anand**: [15:06] That is: each AI has a certain style, a certain personality. Sometimes I like Claude's writing style, but I like ChatGPT's rigor of research. So what I do is—let me—where'd my windows go? Not this one, not this one. Yeah, this one. What I might want to do, for instance, is take this conversation in ChatGPT. Now it's a long conversation, solid content in ChatGPT, but not Claude's style of writing, which I like. And Gemini has a nice style of writing. I would like to be able to copy this entire thing there. Version one: what you can do and anyone can do is go to the bottom and click on the copy response. Paste it in the other chat. **Copy-paste is your friend.** **Anand**: [16:10] However, now if I wanted to take the entire conversation and put it into Claude, that is a bit painful. So what I have is a bunch of mini-tools that I've built for myself. The way I build these tools is I identify a need, which is basically me saying, "Dammit, I can't even do this, what rubbish!" And then going to Claude or one of these things and saying, "Build me a tool to solve this problem." One of the tools that I asked it for is I want to be able to copy the entire conversation, not just one particular chat response, from Gemini, from Claude, from ChatGPT, and give it to me as a bookmarklet. A bookmarklet is one of those things that I can drag and drop into the bookmarks bar. Once I do that, then it will allow me to run it like a tool in that particular page, which means that I can scrape this entire page. I'm going to click on ChatGPT Scraper. What it will do is open these side windows all by itself one by one—what it's thinking, what it's doing, blah blah blah—and eventually ChatGPT conversation was copied to the clipboard. I'm going to click OK. Now let me go to Claude.ai and I'm going to tell—yeah, what the heck, let's do it with Opus, but I'm going to paste the entire thing. **Anand**: [17:55] Now this is huge, by the way. The amount of stuff that—you can probably see from the really tiny scroll bar how huge it is, but not huge in the sense of big data huge, it's huge in the sense of far more than you or I would be able to dictate let alone type. Now what I'm going to do is a trick that I very strongly recommend that you use, which is: **talk to AI verbally. Just talk.** And I was doing that with many of these, but if it's a long conversation, I prefer ChatGPT. Even if I'm going to type into Claude, I go to ChatGPT and dictate into ChatGPT because as of two weeks ago, ChatGPT was still far ahead. I haven't tested it in the last two weeks with the other models; they may have improved, but let's go ahead and say: "This chat transcript has how I was preparing for the AI unboxed session which is on context engineering and it's a first of four workshops that I'm running with the IIM alumni in Singapore. What I'd like you to do is create a blog post in my style. Here are a few sample blog posts that you can refer to and also check out the description of my style and generate it accordingly." **Anand**: [19:10] This is a total digression, but I'm going to show you a bunch of things that we can hope to get to by the end of the fourth session. I'm going to use all kinds of tools. But what we just did was dictation and using a tool to build a bookmarklet and use that bookmarklet to export it. By the way, I will copy-paste the link to where you can check out my tools—it's `tools.sanand.net`—and at least the AI conversation scrapers, which I will also paste, is something that will be useful. And Debi, just saw your question: yes, I will be teaching you how to make tools of the kind that I just showed you. **Anand**: [19:48] So, we have copied from one model, we have pasted in another model. **Do this, by the way, on a regular basis because fact-checking keeps them honest and increases your confidence. They also are able to brainstorm with each other. They have very different styles of thinking.** I used to use Gemini also in the mix and then later on asked them to evaluate each other. And here is my evaluation sheet if I can find it—"LLM Comparison," here we go. Yeah. This was my rating where I would constantly—so on April 18th, I asked, "How do you compress a video to WebM?" And GPT-4.5 gave me an answer with fairly structured methodology; I rated it as the best. Sonnet 4.6 was not so good development methodology. Gemini 1.5 Pro had creative framing and insight but not as good as the others. **So by doing this on a regular basis, week after week, I got a sense of ChatGPT being good for analytics, rigor, and algorithms; Claude being good for strategy, insight, soulful writing, creativity, front-end code; Gemini being good for learning, readable writing, foreign language, people search.** And I'm going to paste this in the chat for your reference because this accumulated wisdom that you might find helpful. But the main message is: copy-paste across these. **Anand**: [21:24] Now, I'm still not done with this blog post because I'm going to show you a couple of other techniques. Now these techniques, some of you might understand, but what I'm going to do is open an MC—I'm setting up a Cloudflare tunnel which will expose my machine to Claude. And now I have a really powerful tool. Remember I mentioned that one of the capabilities that we have as part of context engineering is tooling, and I mentioned web search as one of those tools. **Running code on my machine, searching in my machine, doing whatever it wants on my machine, is another capability.** That I happen to have given it. It requires a certain amount of programming. If you use Claude Co-work or ChatGPT Desktop, you will get this capability. You don't need to do much; just run it on your machine. But because I'm on Linux and half these tools don't work on Linux, I'm forced to do it in a slightly roundabout way. **Anand**: [22:58] I am going to use a local MCP connector. MCP stands for Model Context Protocol; it's a way for people to create new tools that can do stuff. And I'm going to use something ultra-useful for everyone, which is copy-paste from a prompt library. I have a little tool that lets me automatically copy-paste from prompt libraries, but I will show you where this prompt library is in a few minutes, and I'm going to ask all of you to create a small prompt library yourself in a few minutes because this may be one of the lowest effort, highest value kind of thing. This says that I have a local MCP running somewhere, and I'm going to ask it to refer to my previous blog posts—doesn't need to refer to talks and all of that, most of this is not relevant—see my past conversations to figure out where my blog style is, and look at some recent 2024 and 2023 pre-AI posts to understand my style. Now with this longish prompt, I'm going to let it run. What it's going to do is first—actually I have no idea what it's going to do in what order, but it's going to do a whole bunch of things. What it's doing is looking at my blog style; it can't find it in its own machine. Hopefully it's smart enough—yeah, it's figured out that it has to run the code not on its machine but my machine, and it's running the code on my machine. In fact, let's take a quick look at—yeah, this is my machine that's receiving the information from Claude. If you had Claude Co-work or ChatGPT Desktop, it'll just search in your machine for the relevant files and figure out your style of writing or your client's style of writing or whatever you need and take all of what ChatGPT wrote, write it like I wrote it—probably better than I could have written it. **Anand**: [24:59] I'm conscious that what we're doing here is a very roundabout answer to one question, which is how to aggregate memory across multiple providers. But what I'm hoping you would have taken away from this is: **a) copying—for which there is a button—is actually very helpful. You should copy from one AI, paste into another for verification, for ideation, for any kind of red teaming or brainstorming exercise. And using tools you can take it to another level, which we will discuss in future sessions.** Those are the main takeaways for this part. **Anand**: [26:10] Before I go to the next point, one thing that I promised I would do, which is a small exercise on storing your prompts. So this is something that you will be doing. I will put in the question here, I will also put it in the chat window, but this is not really a prompt. I'm going to say—three: **Where are your prompts stored?** And I will explain what I mean by this question. My prompts are stored in a folder which I will locate... yeah. I have this page—that this folder—that has prompts for different things. Now I have multiple files with prompts because some of my prompts are really long. For example, actually let's sort by size. My largest—okay, well, this is a meta prompt. Let's take "Analyze call recording," which seems to be large enough a prompt. I'm just doing a quick check to see that I can share it. Yeah, I can share it. **Anand**: [27:23] This is the prompt: "Analyze call—okay, this is what I use it for—analyzing to extract insights. So the prompt is: 'Based on this transcript of mine—there is Anand's conversations—share in simple explain-like-I'm-15 language, the persona of the people who were in this conversation, any insights from this conversation, what did I get wrong or what did they get wrong, what did I miss, what did they miss, what are some next steps, what I should try out next as an experiment,' and a whole bunch of guidelines on how I want it to write this out." I've been using this prompt for such a long time that with this I get a summary of every conversation not just in a way that it is immediately actionable, but serves as the basis for further inspection. **I've gone back to this, for instance, and said, "Now across the last 200 conversations of mine that I have transcripts for, what are the kinds of things that I'm missing and what does that tell you about me?"** "I'm preparing for a conversation with Namit. Based on his personality, what are all the kinds of things that I should be prepared for?" "What are the kinds of experiments that you've suggested that I should try out? Take a look at the experiments that I have tried out—what am I constantly missing?" **Anand**: [28:54] **Effectively, we are building a base through these summaries for even more refined analysis for AI to do.** But that we'll come to later when we talk about how we aggregate these in future sessions. The point that I wanted to make is: I have a folder in which I have prompts. I also have one file called "fragments" in which I load my smaller prompts. So, if here's what I would suggest: **Open one file. Store some prompts in it.** This is how I do it. So, one-line prompt. Anytime I want to put in some analysis notes, I have written down in text: "As you analyze, note any interesting findings in some file." Or take another one: if I want to get a book summary from a book, it's anyway read most of the books that are out there, so I say: "Comprehensively and engagingly summarize and fact-check, writing in Malcolm Gladwell's style, explain like I'm 15, the book," and then I type out the name of the book. **By the way, this is what enabled me to reach my 50-book goal last year, so effectively that this year I made it a 360-book goal for this year. I don't know whether you consider reading summaries as counting towards a book, but I find that I'm getting about as many ideas and actions and takeaways from a summary that is written well...** --- **Anand**: [00:00] ...that is written this way as I am from reading the original book. So, high ROI, and I might actually be on track to finish 360 books this year. Coming back to the point: **Keep one file where you store all of your little prompts. When it gets too unwieldy, put it into a folder—just one place.** Now you say, "Hold on, why does it have to be a file? I could keep it in Google Tasks." Keep it in Google Tasks, or Teams, or To-Do's, or whatever. **Anand**: [00:32] **It should ideally be accessible on your phone.** I will phrase that as strongly; I would recommend that it be some place where you can easily access on your phone. You need to be able to quickly copy-paste from this and put it into ChatGPT, or Claude, or whatever. The question that I would like each of you to decide on is: **Where are you storing your prompts?** For me, it is in `/home/anand/code/blog/prompts`. Actually, this is not a file, this is a folder. So, that's my answer. **Anand**: [01:22] What is your answer? And I suggest you put in at least one small prompt in that file. It could simply be the prompts that you have used here, or here, or here. By the way, my prompts are also published online, so you can take a look at the same folder—this is the folder that I am using. I will hyperlink it from here so you will also be able to click on this and visit and see the full list of all of my prompts. Some of them, like I said, are large; some of them are short. **Anand**: [02:04] Now, okay, 3:45 PM—I'm probably going to struggle to finish this. But it's not that we have a fixed agenda. Let's see. "Are these chat memories confined to specific tools that we use? Is the chat in Claude confined to Claude only?" Yes, Sushruta, **what you chat with Claude stays with Claude.** **Anand**: [02:31] Rajan, tokens are approximately words, Vaishali. Think of AI as using a slightly different language. **As a rough rule of thumb, between English and their internal language, three-fourths of an English word translates to one token.** It's a rough rule of thumb, but their language is different. **Anand**: [02:51] "How do we know which is the best model?" I'll tell you how I figure out which is the best. There are two comparisons. Within ChatGPT, or within Claude, or within Gemini, how do I figure out which is the best model? Simple: **Whatever is the latest, biggest number is usually the best.** Whatever sounds more fancy is probably the best. "Pro" is more fancy than "Flash," which is probably more fancy than "Light." So, that's how you decide. **Anand**: [03:22] Across these, what is the difference? Hard to tell. It keeps changing. It probably doesn't matter too much between Claude, ChatGPT, and Gemini. Right now, Gemini is a bit behind, but there was a time three months ago when I would have said Gemini is actually ahead compared to the other two. It keeps changing. **There is a far more nuanced answer to this that I will send you a link for, which is about the—okay, I'll spend a minute talking about this because this is important.** **Anand**: [03:52] There is a site that measures the quality of models; that is **LMSYS Arena**. I am sending you a link to their leaderboard. If at any point you want a rough rule of thumb saying, "I do want to find out what is the best model," then use this leaderboard. **Keep in mind that 'Model' is different from 'Agent'.** ChatGPT is an application that uses the GPT model. Claude is an application that uses the Sonnet or the Opus models. The applications themselves do a lot of stuff, and that can make a big difference. **Anand**: [04:37] But models-wise, this is a pretty good leaderboard. I've been tracking this leaderboard to figure out roughly what is the level of intelligence of these models. So, there was a time in around 2023 where they were about as good as a high school student. Over time, with GPT-4, they became as smart as a college junior—that was just November '23. And then with time, with O1-preview, they became almost as smart as a Master's student—that was less than a year later in September '24. And with ChatGPT-4o as well as GPT-4.5, they became as smart as a PhD student. **And with Gemini 2.5 Pro, which is almost a year old, they became as smart as a tenured professor. Now, we are talking to models that are smarter than tenured professors.** **Anand**: [05:29] If we look at GPQA, which is OpenAI's study where they asked experts to create tasks that experts do, asked experts to do those tasks, and had AI do those tasks, and had experts verify it and check: where does AI beat humans, human experts? Where do human experts beat AI? Green is where AI on average is doing better. Let's take software developers: 70% of the time O3-mini, an old model, beat humans; today it's 95% plus. Accountants and auditors: at that time Claude was beating humans only 24% of the time. Personal financial advisors: **personal financial advisors are already beaten by Claude, ChatGPT, any of them. So, I'm not going to hire a personal financial advisor anymore.** I will probably hire an auditor, at least for a neck to catch. I will probably hire an industrial engineer if I need to. The thing is, these models are constantly improving in capability. So, the way I think about it is: **rather than hiring a human, I may as well hire an AI for that particular task.** But the link that I shared, which is "LLM Pricing," will give you a sense of the price-quality evolution of AI. **Anand**: [06:54] Next... ah yeah, here we go. We spoke of web search. "Do these tools perform scraping?" Yes, they do, also. Slightly different, and we'll cover it when we get to tools. "ChatGPT does not have access much to prior conversations other than for a couple of key areas." Two possibilities: memory may not have been enabled; memory may have been enabled later. Actually, three possibilities: you may not have the ability to enable memory at all in the first place. **Anand**: [07:29] Avinash asks, "How can I pre-train my model with my historical data—emails, LinkedIn, other social media docs, etc.—that is, create your digital twin?" **"Training a model" is a formal term, Avinash.** What you probably refer to is somehow getting something to be able to do the equivalent of a digital twin. You kind of saw what I was doing on another chat a short while ago, which I'll try and locate here, where I was having it write it in my style. Let's take a look at that and see how far it's come. **Anand**: [08:09] Okay, so it's written a blog post. "I'm running a four-part AI unboxed session... blah blah blah." Here's what survived—not exactly my style, but the structure is very close to what I would have written. Stuff like "Yes, I am fishing for compliments" is exactly the kind of thing—and "ChatGPT is wonderfully obliging with this"—is exactly the kind of thing that I would have written, though I might have worded it as "ChatGPT wonderfully obliged." That's nitpicking. **Yes, today the capability of these is that if you give it a little bit of your context, it is able to copy you well enough that corrections are nitpicks.** **Anand**: [08:58] And therefore, to your specific question: **How do you create a digital twin? The answer right now is precisely this session.** Upload stuff about what you need to give it. Which begs the question: "How am I going to get it all my emails, all my LinkedIn, all my other social media posts, documents, etc.?" Download it. LinkedIn offers exports. Google offers email takeouts. Hopefully your other social media are not such walled gardens that they will not allow you to take it out. Documents created—export those. Method number one. Method number two: **Give them tools to be able to access these.** **Anand**: [09:44] So, let me show you... so one of the tools that, like I mentioned, I created—I'm not going to—I'm just going to show you what is possible, rather than—actually, I'm going to show you what has been created, rather than... let's see. Here I have—yeah, actually, I'm going to show you my meeting prep. What can I show you that would be a good example? Yeah, actually, for AI unboxed, one of the things that I blogged about is relevant here. Is this the one? Yeah. **Anand**: [10:28] Okay, what happened here was... this poster was created by Nayana. And I wasn't even aware that somebody's going about creating posters. What I did do, however, was I asked Claude a question: "**Who are the people in my life that most deserve an unreasonable gesture of thanks, and what would that be?**" Now, this requires a lot of context. It needs to know about my life. Effectively, I could ask my digital twin this question. But randomly asking Claude, how would it even answer this question? **Anand**: [11:07] But it did. And let me show you how it went about it. Remember this local MCP thing that I set up, right? It managed to use that. But in your case, you could run this on your machine and you would be able to figure this out. Where is this? Let's see. Search for "thanks"... "unreasonable gesture"... yes, that is the one. So if you had run this locally—this is the full prompt—I told it to go through my notes at this particular folder, my transcripts at this particular folder, through my emails in this particular folder. **But it also went through my WhatsApp chats.** **Anand**: [12:02] How did it do it? Because I periodically export my WhatsApp chats, whether I read them or not. Actually, there are other ways, and that we'll cover in the tools section. But all you have to do is... in fact, let's export the WhatsApp chat. In fact, yeah, here is an exercise that we should all do. Actually, many of you would have done this in the last session where you go to WhatsApp, and on your mobile, you have an option against the three dots. Let's go to this session, yeah, the AI unboxed group. There is an option here, not out here in the web version, but **on your phone, there will be an option to export the chat.** **Anand**: [12:44] Let me see if I can locate that and show you. Yes. Fingers crossed. Okay, now I have to turn off the camera's background feature and you should be able to see on the top left somewhere, there is an export chat feature. **The way you locate that is by going to any chat and clicking on the three little dots on the top right.** Of course, mine's an Android, yours may be an iPhone; it may vary. And you still won't find it there because you then have to click on the "**More**" at the bottom. When you click on the "More," that's when it will give you an option to export the chat. **Anand**: [13:35] Here's something that I do on a regular basis: export these chats, especially classmate chats, family chats. **Give it to Claude or ChatGPT and say, "Tell me what's been happening. Catch me up. Is there anyone that I should reach out to—some birthday, some celebration, some congratulations? Let me know so that I will be an active member of the community without having to take the pains of reading every single message."** Effectively, I'm creating a secretary to do my WhatsApp management. It can be automated further, but this is a simple one where there is value. **Anand**: [14:16] So, here is another exercise that I'm going to ask each of you to do, and we'll pause here—we're just 10 minutes short of the other Q&A bit. But, and you just have to answer a yes or no: "**Could you export a WhatsApp group and summarize it?**" And the answer can simply be a yes. If it's no, don't bother filling it in. So it's really only one option. **Anand**: [14:53] A gentle nudge: seven of you have filled out Column E—"Where are your prompts stored?" Because this is important, I'm going to wait until we have at least 20 of these. Just figure out one place where you will store your prompts. Doesn't matter where, as long as you remember it, it's fine. **Put in at least one prompt there. Type the location here. It will build your muscle memory. And to take it a step further, if there is only one thing that you learn out of this context engineering session, it is: store your prompts and reuse them.** Frequently, infrequently—anything is okay. **Anand**: [15:44] We have 10 people. Let's give it a few... okay, 14, that's nice. Thank you. Okay, Notes app is a good option, OneDrive folder is a great option, Google Drive folder is a great option—all of these you can access on mobile. Notepad is great; it might be good, Harish, if you had some way of accessing the file from your mobile. As long as you have that, that's fine. "**Memory of the LLM itself.**" Oh, that is clever, Mike; I had not thought of that. Thank you, I will try this. "**WhatsApp group with yourself**" is another good idea. iCloud folder, yeah, "Memory of the AI platform," yes, that's a good one. Okay, "We'll consolidate after this session," sure, fair enough, Rajan. And Ravi, you have it in a folder, that's great. Spreadsheet is another interesting idea. I have not yet cracked being able to copy and paste in and out of spreadsheets on mobile, but on desktop it's actually a great option. OneNote, iCloud, OneDrive—yes, all of these are great. Please just put in prompts, even if you use them infrequently. Go through some of the prompts you've used in the past and say, "Oh, maybe I can use that." It's a good review practice. **Anand**: [17:08] I'm going to pause taking questions because there are at least a couple of things that I wanted to cover before I hand over for Q&A. And that is coming back to our framework. **I mentioned that there are four things that broadly go into context: Prompts**—that's what you type; keep a prompt library, that's where you have maximum value. **Files**—keep your files organized, download stuff, export stuff so that you can quickly upload it. We just saw how WhatsApp would be relevant as a possible upload; you will be able to download emails. With tools, I'll show you how you can automate some of that process further. **Memory**—is just about enabling it and telling it to access the memory. **Tools**—right now we saw search; there are many more tools, just explore the plus icon and see what else you can connect to. You can connect your Google Drive, you can connect your OneDrive, you can connect your Dropbox, and that's probably the closest that you can get to getting to a digital twin, or using the desktop version of the application. **Anand**: [18:14] But this is one part of it. Effectively... prompts, files, memory, and tools are, I would say, the content of what goes into context engineering. **There is the process of what goes into context engineering.** How am I going to capture all of these? Where am I going to get them from? How do I find these? How am I going to structure these? How am I going to store these? And how am I going to improve these? Something to constantly keep thinking about. We've already done a certain amount of this. For instance, prompt storage in a prompt registry is, in my opinion, in this entire grid, the most important from an ROI perspective—not necessarily the highest value, but highest value per unit of effort. **Anand**: [19:12] But there are other places you can store it—okay, I'm not going into this. There are—you might want to consider where am I going to store all my files and how am I going to make it easy to access? If you have files in OneDrive, how easy are your files to access? Are you exporting all of those files? Memory, where exactly is it getting stored—that's a broader topic. Tool registry and all, we'll come to. **Anand**: [19:35] But the other part is shaping these and improving these. And I want to end with the **improvement** part. You've collected files, you've collected prompts, you've collected some memory, you may even have collected some tools. **One of the important things to do is periodically go back and ask AI, "Am I using you well?"** I would ask a vendor for feedback, I would ask an employee for feedback, I would ask them, "How can I leverage you better?" And AI is as smart; I would ask them that kind of question. "Go through my prompts. We have memory enabled. How can I improve the prompts that I'm using? I've fed you a bunch of files. Here are a bunch of additional files. Should I be using them? Check—are these fresh enough? Should I update to my latest set of emails? Are there a whole bunch of CV's that I've not yet put into my folder? Is it actually picking up the files? What are the files that it's picking up, what are the files that it's not picking up?" **Anand**: [20:45] **In short, constantly reviewing the context that you're providing—and when I say constantly, I mean at least once a month or when you remember it.** It's worth doing because these models are improving rapidly. Your context decays, depreciates on a regular basis, and doing this will help you use AI better. **Anand**: [21:07] Which leads me to two meta advices about context engineering. A lot of people ask me, "I don't know how to use AI, how do I figure out? I don't know if I'm using AI well enough, how do I figure out?" **The answer to both of these is: Ask AI.** For figuring out how to use AI, **AI interviews** are a fantastic technique. "I'm feeling lost. Interview me and help me." That's it. It'll figure it out. Now if you have enough wits around you at that emotional state where you are even able to say something as sophisticated as "Interview me like a psychologist and figure out," or "Interview me like a personal coach and figure it out," or "Interview me like an HR professional and figure it out"—even better. You're giving it a role, you're giving it a personality. But effectively you are saying, "**Look, you be the doctor, or you be the professional, and I am seeking your advice.**" And that includes the use of AI as well. AI is excellent at medicine, at computing, at knowledge of how to use AI among other things. **Anand**: [22:24] Interviewing is a very powerful technique; it is one step ahead of meta-prompting. Meta-prompting is where you say, "Blah blah blah, this is what I'm thinking as my prompt, how can I improve my prompt?" It'll give you a good prompt. Interviewing is cascading that and coming up with a really good... well, solution potentially, but certainly a very good problem statement. **Anand**: [22:47] And the second is **continuous improvement**, which is: I call it a **post-mortem**. And the post-mortem applies across anywhere in the ecosystem. Go through your past chats. Ask it, "Where did you do well? Where didn't you do well? Where did you get feedback from me that told you to correct yourself? And in the future, what do I need to do to make sure that you get this right one-shot? Do I need to give you a different prompt? Do I need to give you more files? Don't you have access to my memory? Do I need to give you access to more tools?" And you don't even need to say all of this; it knows this stuff. **Ask it to interview you when you don't know what you want; ask it to run post-mortems so that it can help you improve even further.** **Anand**: [23:35] It's just past 4:00 PM Singapore. I will pause here and maybe just temporarily hand over to Saurabh or Debi—to either of you—if you want to do any, whatever, scene-setting before we dive into questions. **Saurabh**: [23:58] Debi, over to you. **Debi**: [24:01] Uh, first, there are a few outstanding questions on the chat. Do you want to take them up? And then, yeah, if you can, I would suggest we do that first. **Anand**: [24:11] Sure. Though one request, Debi: **Would you be able to talk me through the questions because I'm losing track as I'm jumping across windows?** And I can then just focus on the answers; it'll be easier. **Debi**: [24:22] Okay. So I think one of the things people have said: "When you bounce between Claude and ChatGPT, how do you give both the same and similar context so both can be equally useful?" **Anand**: [24:35] Fair point. **I usually type into ChatGPT, copy, paste it into Claude.** The reverse doesn't work as well, but it's exactly the same thing. I put in a bunch of attachments and type in some stuff into ChatGPT, paste it into Claude. I also do that because ChatGPT is usually set up so that ChatGPT is a little slower, so I give it a little bit of a lead in the race. **Debi**: [24:59] And then I think when you talked about WhatsApp messages—you did talk—I think somebody asked, "How do you give it access to WhatsApp messages?" which I think you showed that if you look at the WhatsApp on your phone, there is an option to export the message. And I think the specific point they said is, "**How to ensure my private content isn't used beyond the analysis that I asked for?**" **Anand**: [25:24] There is an option for that. Let me share my screen, and that reminds me that I forgot that we should take the screenshot. So, yeah, one request, just heads up for everyone: if you could please start coming on video with the AI unboxed background, it'll be nice to take a picture. Let these just start coming on. Thank you, Shijo. Others please just start coming on. **Anand**: [25:56] I'll show you the settings in the meantime. In ChatGPT, I think it's in security where you can—no, not security, privacy—data controls. **ChatGPT has it in data controls where the option to improve the model for everyone is what you need to turn off.** So, I have turned it on saying, "Look, all the stuff that I'm talking about, yes, please take it and train your models with it." If you turn it off, then it will not be used for training. **Anand**: [26:27] The equivalent option for Claude is probably very similarly worded and is likely to be in privacy. Yes, it says "**Help improve Claude.**" If you turn this off, then it will not be used to train others. Let me stop sharing, and thanks, Anand, you're on the nice AI unboxed logo as well. A request: even if you are not there, at least show the logo please. You can move away from the camera if you like, but I'd like to see the logos fill up a screen at least. **Debi**: [27:08] Sure. Some other questions: I think on the WhatsApp, just one more question was: "**Can the chat have to be exported chat by chat, or can you combine it together and export?**" **Anand**: [27:22] If you do it manually, you have to export it chat by chat. I have a tool that exports it in bulk and refreshes it on a daily basis. Let me show you how my tool runs—and I'll send you the link to the tool—but that's something that you won't learn how to build right now. Where's my tool? I should turn this off. Okay. **Anand**: [27:48] So, my tool is called `backup_whatsapp.py`. When I run this, it is unfortunately going to take over my browser and it will go tab by tab. So, yeah, here it is; it's going through and I'm going to hide it because it literally goes through every single WhatsApp message. But yeah, that is what it will do. Where is my Teams gone? Ah. You'll probably notice that it's turned my screen white; that's an indication that it's doing some work in the background. And I will probably show you... yeah, here are the chats. **Anand**: [12:13] **So it's identified 35 chats that it needs to refresh.** In this generative AI chat, it has found 106 new messages in the last 24 hours or so—it's a very active group. Carthage messaged me 11 times; I should check what's happening. And yeah, here are a bunch of other things which it's going through one by one and saving it in a file. **I have no idea what code it wrote.** I do know enough about technology to give it one or two hints to explain how to write that code. But where is the script? Let me show you that if I can locate... uh, scripts... this `backup_whatsapp.py` script is a Python script and you know enough about how to run, not necessarily you don't need to know how to write. WhatsApp script... Shijo, Debi's the master of ceremonies, so she's the one you have to flag. But this is the script, and the prompt that I used to write it is out here. **Audience**: [29:41] I have a question regarding the memory syncing that you did. You basically said that you're going to copy the same message to the different models and get it done, right? Somewhere in the screen you had a very interesting setup where you were able to evaluate the different model responses... [audio cuts] --- **Audience**: [00:00] ...responses and rate it the best answer—red, amber, green, or something like that. **How did you do that?** And the second thing is: typically, when I'm having a conversation, I find value in going to different models, but I usually have a flow of thought and I prefer to stick to one model rather than going to different models. Typically, when I get a response, my next question will also change. So, is there a way... and right now my ChatGPT folder structure is exploding, and if I copy it, it's going to be quite a nightmare—I mean, if I manually try to do that. But you do seem to have done some kind of an automated way to evaluate the responses of the different models, and that sounded very interesting. **Anand**: [00:53] I have a poor short-term memory, so I'll answer the first question first because I will remember that. **It is done manually—very unsexy.** What you saw was Notepad, and I typed out the observations. However, I got a little smarter about it. **I would then ask Claude Opus to judge which of these answers it thought better.** And beyond a point, I just started accepting its opinion, and that's the LLM-as-a-judge paradigm. The reason was because I also learned enough to ask it to give me the reason for doing that. **Anand**: [01:31] That's part A. But before we do that, I must pause to share my screen. I notice that we have enough—almost enough—people to take a screenshot. I'm going to share my screen, and then Shijo will come back to your second question. Good. One, two, three, four, five more people... okay, four more people just turn on... or do I need to do something so that we just have the video backgrounds? I've said "Show videos first." It would be great if... let me do that once again. "Show videos first." If we could have all the IMpact videos in one... hmm. Okay, no, it's not quite... Does anyone else have a better view that has the IMpact logos all together? I'm just going to take a few screenshots as they appear. There we go. And paste it. Third time lucky. Next session, but yes, no, this might be a recurring request, please, that we do some of these screenshots. Shijo, sorry, back to your second question if you could repeat, please. **Audience**: [03:02] So, I kind of hear that you are basically doing a lot of manual work to kind of sync the memories across models. That's the key takeaway. But what I am interested in is: in my workflows, what happens is I have a conversation, and I find it very difficult to copy the same thing into the different models. Usually, when I go to the different models, I'm usually going with a conclusion or something like that, right? So it's not a step-by-step copy into the different models. So, any insights, any observations? **Anand**: [03:41] Just give me one more moment. Let me just clarify that. So what I'm trying to say is, I find the idea of sharing memory very powerful. However, because of the way I'm doing things, the idea of copying each and every step seems a little sub-optimal. So, is there any suggestion? That's the question. **Anand**: [04:09] Got you. Then **the most efficient way for you is, at the end of that conversation, tell it: "I'm going to give what I need to give to Claude. Just write down what I need to copy-paste," and it'll give you what you need to copy-paste.** That's your most efficient option. **Audience**: [04:29] Got it. **Debi**: [04:30] So Anand, there are three sets of questions. I think if you can just take three of them separately. One is obviously, I think there were questions on prompts and broadly they are: **What is your advice on how do you structure prompts well?** Secondly, **how do you use AI to improve your prompts?** Right? So I think that is one set of questions. The second one was that I think you talked about MCP [Model Context Protocol] very briefly, you showed how you use Cloudflare to get there—I think a lot of people just got lost out over there. So people said, "Can you take a little bit of time to talk about MCP?" And third is, I think there was one isolated question on skills. But if you can talk about prompts, MCP, as well as skills, I think that'll be great. **Anand**: [05:21] Cool. How do I structure prompts? **I just write it randomly, and over time, if I'm using it again and again, I just copy it, paste it, take five seconds to look at it and say, "Oh, this is not what I want," change it, change it, change it** in the same folder or file or wherever you are storing your prompts—my equivalent of that. That improves; that's part A. **Anand**: [05:46] How do I use AI to improve my prompts? Using **post-mortems**. After a conversation, whether I'm happy or not, I have AI periodically—like once a week-ish—go through and say, "Here is a longish conversation. What would have improved, or what prompt should I have given to improve it?" And sometimes it goes overboard—effectively doing the equivalent of over-fitting in model engineering. It'll try and tailor the prompt specifically to that conversation. So I tell it, "Go look at half a dozen conversations." Even then, sometimes it over-fits. So I simplify. In short, I glance at it, tweak based on what I find works/doesn't work, ask AI to improve it, and tweak based on what I think will work/doesn't work. **But I don't spend too much time and effort on it because when models change, half these prompts are often not as effective the way they are crafted.** And the models are getting smarter to the point where these days I just say, "**Look, you kind of know me, right? Blow my mind.**" "**Blow my mind**" is an actual phrase that I use in a surprisingly large number of conversations, and I trust it these days to know enough about me. So, this is how I do it. No need to overdo it. **Anand**: [07:05] Question two, which fortunately I've written down: **What is MCP and how do I use it?** **MCP is Model Context Protocol.** Anthropic came up with it as a programmer's way for them to expose programs so that AI agents can use them. So I prompt, and on my behalf, it runs a bunch of programs. That's MCP. How do you use it? We'll get to it in another session; don't worry about it for now because there's just no way we'll be able to cover it in 15 minutes. **Anand**: [07:39] And it's not as powerful as the third question, which is **skill.md**. Skill.md is actually interesting and is unfortunately not as widely supported as I would have liked it to be, otherwise I would have covered it in this session. But I will spend a few minutes talking you through skill.md on Claude. So, **think of skill.md as an on-demand prompt.** In Claude—and so far to the best of my knowledge only in Claude—there is a "Skills" section under "Customize" where you can add, by clicking on this plus button, a new skill. A skill is simply a prompt that is loaded on demand. **Anand**: [08:33] Rather than show you how to create a skill, let me show you some of the skills that I've created first. I've been creating data stories for a decade and a half. If there is one thing that I kind of know well, then it is data storytelling. I've taken everything that I know about it—or even data analysis, I'm not too bad at that—taken everything that I know about data analysis as a process (this is effectively the course that I run on data analysis) and said: "Here is how you analyze data. First, understand the data—what is the structure, what is the quality, what is the distribution? Take a look at all of this. Then figure out who the audience is, what decisions are they going to make, and so on. And now start analyzing the data. Look specifically for unexpected distributions. Where are the patterns breaking? Are there any surprising correlations? Blah blah blah." Very detailed instructions that effectively capture everything that I know about this particular topic and deliver my skill to Anthropic. **Anand**: [09:26] How do we do this? Click on the plus button, you click on "Create a skill," and you can, option A: write skill instructions—which is give it a name, so let us say it is "How not to mess up a workshop." I'm putting this in lowercase hyphens because I got used to it—I don't think it matters. And you describe: "Use this as a guide to when to flag errors—common mistakes when running a workshop." And then you put in all the instructions on how to do this. Now, I would type this in if I really believed I know more than Claude. **There aren't many places where I know more than Claude.** Other people may know more than Claude. If I trust them enough, then I will copy their skill. And you can just do a search for skills on skill.md. I suspect this is my skill.md. A Google search for "skill.md" will land you probably at the site `skill.md` where there will be—I'm sure you'll find a `skill.md` gallery somewhere. And then these will be emerging. **Anand**: [10:42] So if you happen to find a skill that works for you—so here's a marketing skills collection of `skill.md`'s—you can download it, upload it there, or copy-paste. Anthropic themselves have a bunch of skills. Cancel this. An example of a skill that I copied is the "Design Skill." I copied it from Anthropic's blog themselves. I find it kind of okay, not great. **In short, skills are permanent prompts, usable on demand, typically good for hardcore expert advice.** If there is an area where you are looking for expert advice beyond what the models can offer—which is most areas—and you have reason to believe that there is a particular skill that will outperform the model (which is effectively trust in whoever wrote that skill), then you can add it here. My skills are stored in this directory. You can copy-paste... actually I have a whole bunch of skills; there aren't that many that I trust myself to, but for what it's worth, I will paste my list of skills on the chat. Those were the three answers, Debi. **Debi**: [12:07] Got it. And sorry, so once you store the skills, right, which you just showed over here, essentially does that become—I mean, that again becomes context. Is that the case? **Anand**: [12:18] Exactly. **On demand.** So it will look at the chat, figure out when it should use that skill. For instance, I've told it: "Use the data analysis skill to investigate data for surprising and actionable insights." So if I said "Research me," it wouldn't bother reading the skill. **Debi**: [12:35] Okay, okay. So the description is the one which defines in what context this skill should be used. That is what gets matched up. Okay. Got it. **Anand**: [12:47] Yes. A rule of thumb is, it gets a directory of all the skills with the name of the skill and the description so that it knows whether to pick up the details of the skill for a question. **Debi**: [12:59] Got it. Okay. Before I open up, Anand, would you like to just give some kind of, you know, ideas on what people can experiment with before we meet next? **Anand**: [13:12] Hmm. That is a good question and yes, I will do that. **Usually the way I approach these problems is by asking AI,** and I'm just going to do the same. But sometimes I don't tell people that I'm using AI behind the scenes and just read the answer. In this case, I'm just going to tell you this is what I'm doing. Now we have... okay, where did this go? "Context Engineering Techniques." Okay, this is where I wrote the blog post right now. Let's go to ChatGPT and say... this time I'm going to give think-step [thinking mode] and remove the search. **Anand**: [13:50] "I would like to give a series of exercises that the audience can try out by themselves. The thing is, it needs to be: A) very useful and impactful for the majority of the audience; B) I would be more confident that it will be useful if it's something that I've tried and have found it useful. So go through my past conversations, chats, everything that you know about me, blah blah blah, and make sure that you suggest maybe three exercises that are at the intersection." And let it run. Memory is helping. The prompt, because it's verbalized, is allowing me to think as I speak. It is smart enough to figure out, based on what I'm saying, what my intent is. **Anand**: [14:42] What I often do is have little prompt snippets that I add to content—but okay, we'll come to that later. In this case, okay, "Conversation X-Ray." Yes, this is easily one of the most impactful things that you can do. **Take any conversation in any format and analyze this conversation as... try out this prompt asking it: "What was decided? Who promised what? Are there any hidden concerns, tensions, etc.?"** Prompt number one. Why am I doing this? I'm going to take the whole thing and copy this. I will put it on `docs.new` (that's `doc.new`). Paste it into a Google document. This is the best way of pasting it. No, it is not. Hmm. Paste without formatting. Paste it. Yeah, that works much better. I don't need links to my past chats. This is something that I can share publicly. So I will take the same title and "Exercises," save it, make sure that it is viewable by anyone with the link, copy the link, paste it here, and also put it in this as "Exercises." Yeah, these three exercises. Yes, sorry. No, I—and here is the link. **Anand**: [16:35] Ah, I will in all likelihood disable this column before—I mean, in a few hours after this session. So we'll enable it during the next session. Just try it out; it'll be read-only, not read-write for a while. That's all. **Debi**: [16:51] Got it. Okay. No, thank you so much. So sorry, I think these were some of the prompts which came up. So opening up to everybody on the floor for questions. If you can just—I think everybody's mics are on—so if you want to ask Anand anything right now, please feel free to do that. **Anand**: [17:08] Or don't, if you have better things to do. **Audience (Vijay)**: [17:10] Anand, Vijay here. **Just on hallucinations: How do you protect against the model hallucinating—what do you look for?** Part one. Part two is: **Do you find it useful to get Claude or Perplexity to challenge ChatGPT responses and vice versa?** **Anand**: [17:23] I find it extremely useful to have one model challenge another. And the thing I'm learning is: we aren't... we're fairly used to having humans hallucinate all the time—we call it lying—but the thing is, given that there are enough institutions that already do this and do this pretty well, especially where, for instance, the juror is far from an expert in the patent case that they are reviewing, nor is the judge for that matter. The regulator knows far less about banking than the bank, and certainly about the bank. So, we effectively have many of those institutions that we should use as principles. **Anand**: [18:14] And therefore, when I was asking ChatGPT on extensive research: "**How do different professions use this and how can we leverage it?**" Here's what it said: 1. **Ask the model for falsifiable output.** Not "Give me the answer," but "Give me the answer along with evidence, etc." You saw me do that a little while ago. "Give me verbatim the quotations from my transcripts as well as the link to the transcripts," which I will copy-paste and check. And that did a much better job; at the very least, I'm able to verify. 2. **Convert the output into tests—effectively checklists.** If you have these checklists—and you can ask it for a checklist—it's often not trying to intentionally deceive you; it is just making mistakes, and that's what hallucinations are. That's one mental model. 3. **Checklists are a different point.** Use checklists; these are extremely effective. 4. **Testable output.** Almost all kinds of—several things are testable. Code is testable; it will run or it will not run. Once you compile it, you know it works. But surprisingly, insurance contracts are testable. There is a language called InsuREli where you can take an actual contract... let's see... so this is not what I meant to do... where is InsuREli... yeah. Here is a vehicle insurance contract, for instance. Something like this can be converted into a series of rules in the InsuREli language. And when somebody submits a claim, that can be validated almost programmatically against this sort of rule. So, if you can convert into formal tests, then that gives you another means. 5. **Sample.** Why bother reviewing all the answers? I just check a few things and say, "Oh, this is rubbish," or "Oh, this is fantastic." 6. **Independent challenge**, which is passing from A to B—we call it **red teaming** as well; that works. See if it actually works, and if it doesn't... and there are many things you can test out immediately. Have a conversation. If it says "Say X," the person should do "Y." Did they? Did they not? Feed it back. 7. **Maintain a log** like you saw me maintaining on Notepad with that red, amber, green thing. It's a simple way of doing it. 8. **Build memory**, both individually as well as an institution. **Anand**: [20:37] So this I will copy and put into this chat as a reference. Again, like with everything else that I'm sharing with you, this is half AI-generated, mostly ephemeral; take it with a pound of salt. **Audience (Vijay)**: [20:56] Thanks, Anand. **Anand**: [20:59] Rajan, your hand's raised. **Audience (Rajan)**: [21:03] Okay. So hi. Thank you again for your generosity—incredible. So Anand, what I found—and I had the benefit of you supporting me with the data story, which by the way changed the absolute destiny of my business, which I can separately perhaps talk about, but what I'll say is **what Anand did—I'm not using the word lightly—it changed the destiny of my business.** So, what I also found was that when I used that data story and the unbelievable case study that it threw out, there were certain small corrections I needed to do later on—like some data, some, you know, "Don't have this in bold, have this," you know, etc. So it was pretty straightforward stuff. **I found that increasingly as the chat progressed, the model in my view got stupider and stupider.** And the same I have noticed in Gemini image creation—that at first it comes up with something really brilliant, this, that, and you want to make some small changes, you want to say something, etc. And it just... on chat it says, "I got it, this is what you want, this, that, etc.," and then it gives you exactly the same output that it had previously with no change, or even worse. So I'm wondering if I'm doing something wrong, or that's all. **Anand**: [22:24] Known behavior. Claude and Gemini are more susceptible to this than ChatGPT, but we don't necessarily like ChatGPT's style always, so we do want the other models. There are a few solutions, all of which revolve around **somehow starting a new chat**. And the question simply is: what information do we give that new chat? It could be memory; you could say, "There is a past conversation, refer [to] that." It could be copy-paste. It could be "**Summarize this conversation so that I can copy-paste it.**" It could be any combination of these, but **start a new chat is generally the solution.** **Audience (Rajan)**: [23:02] Got you. **Anand**: [23:06] Amita? **Audience (Amita)**: [23:08] Thank you, thank you. Thanks Anand, that was a great session. So my question is—I think you referenced it before—but I just want to say that, you know, obviously over the last several months there have been a lot of prompts on ChatGPT. So at different points in time, I might have spoken... I might have asked ChatGPT about a specific theme, say for example "Donations," but I used different prompts—so sometimes I'll say "Giving," sometimes I'll say "Donations," sometimes I'll say something else. **When I'm trying to triangulate at a later point to remember—by the way, what did I really—because there were great responses, but now I don't have those, I'll have to scroll through all my prompts.** When I'm trying to aggregate and remember what was the response, I... despite the memory being enabled, ChatGPT is unable to retrieve the information. So what might I be doing wrong? **Anand**: [24:05] You most likely aren't doing anything wrong. **The search capability for these models is not great.** Claude seems to have the... okay, so there are two ways in which we can search—one slightly ineffective, the other totally ineffective. So out here there is a "Search chats." This has a combination of explicit keyword search, mostly against the title, plus a little bit of meaning-based deep digging. **It is mostly ineffective, with the sole exception perhaps of Gemini, which seems to get this part right.** So all of them have a little search on the left; that's okay. **Anand**: [24:52] The other is, "Can you give me a link to a past memory?" **Claude does this best, ChatGPT kind of does this okay, Gemini does not do a good job of this.** In short, right now your best bet is twofold/threefold. Number one: **keep a bookmark of your important chats somewhere—manual, painful.** What I do is **rename them**. So for instance, out here there's a "Pranav police complaint preparation discussion." These are not the original titles that it gives; I rename these. And that helps me search for those keywords. So "IIM Alumni Workshop"—it started with a bland "Context Engineering," that's not helpful. Third, there are **Projects**. We did not cover Projects; that is a very useful feature. It's roughly like folders. I don't use Projects much because I haven't learned how to use them effectively, but everyone tells me that it is very powerful. Please do try it out. **Audience (Amita)**: [26:04] Thank you. **Anand**: [26:06] I think Sandeep, you're next. **Audience (Sandeep)**: [26:09] Yeah, thank you, thank you Anand. This was... this was fascinating and fun. Just a question: see, last time when, you know, we did the session in December, there was this Chinese app, Manus, which got acquired by Meta and then got disallowed, but that apart... so they are kind of a **wrapper above the LLMs**, right? They, you know, create their own engines of... **so is that a superior way of using AI functionality of the underlying models, would you think?** **Anand**: [26:43] Yes, they are called **harnesses**, and is now becoming the de-facto way of using models. The idea is that models themselves expose certain raw intelligence; the way in which we orchestrate that intelligence or engineer that intelligence is the job of the harness, which includes for instance: letting it search, letting it write Word documents or PowerPoint presentations, giving it access to Dropbox, etc. And different harnesses are good at different things. **Coding harnesses are the hottest topic these days. Claude Code is a better harness than GitHub Copilot**, even though both of them behind the scenes are using the same Anthropic models. And that's because of how they've been, you know, engineered, whatever. In that sense, yes, **nobody accesses models directly for the majority of their real-life tasks.** They do it through harnesses. ChatGPT is a harness. Claude.ai is a harness to models behind the scenes. **Audience (Sandeep)**: [28:04] And is OpenC similar? Is that a harness as well that you would say? **Anand**: [28:08] In a broad sense, but it is different enough that it probably deserves a category of its own. **It is always on**, and that enables a very different kind of capability unlike almost every other harness which is initiated by us. **Audience (Sandeep)**: [28:29] Okay. Thank you. **Debi**: [28:32] I think Sushmita, did you have a question? Because I saw your camera on earlier. **Audience (Sushmita)**: [28:38] No, I was just testing out the background. Thanks. **Debi**: [28:43] Okay. Anybody any other questions? Otherwise I'm going to bring the session to a close. Guys, anyone? Okay. Thanks so much. So Anand, thank you very, very much for this session, and I think we'll share the recording of this session with everyone. And at the end of it, Anand, I think between the time from now till the session that we are going to have in June, I think you might be sharing something with us. **I also hope that everybody who's attending it is going to try some of the exercises that they have**, and if any of you've got any questions, just shoot it out and then I think Anand can potentially, you know, address it next time we meet or even in the interim. Okay? Thanks so much, guys. Appreciate everyone's time. Thank you. **Audience**: [29:39] Thank you. Thank you, Anand. Thank you, everyone. **Anand**: [29:43] Thank you. Bye.