# Transcript **Sandeep**: [00:01] Yeah, so good afternoon, everyone. Happy Saturday. Welcome to session two of AI Unboxed with resident guru Anand. I'm Sandeep from the IIM Alumni Association. This is part of a four-part series; this is part two of the series that we have planned. The first one was about a month ago, and we have two more planned in July and August towards the end of the month on a Saturday afternoon. **Sandeep**: [00:34] Quick introduction again: Anand is the Head of Innovation at Straive, which is an enterprise AI and data analytics company. He also set up Gramener, which was acquired by Straive a few years ago. He's been kind enough to spare a lot of time for us. As his yearbook says, he is a **"God of all things."** Going back to the last session, Anand, I must say several people have noticed and have been taking cycle rides from police stations to get inspired, the way you did. Hopefully, it wasn't related to any forced visits to police stations. I'll just hand it over to Anand. Anand, you want to just start off? We have about 50 people online as we speak. **Anand**: [01:36] Let's dive in. I'll share my screen and I'll also post the link to a little form which I will be opening with. Let's see if my screen becomes visible. Okay, cool. I thought we could start with a reflection on how we were using AI last month and what was your most memorable use. I'll share mine in a few seconds, but you can either take a photo or do a QR scan of this thing, or I've also posted the link in the chat. **A lot of what we learn is really from how others are using AI**, so one of the things I plan to do is take from your learnings, share in this call, see how we can go deeper into that, sharpen some of what we've been doing, and learn from each other. **Anand**: [02:43] I will be answering this question—and thank you, we have a couple of answers already. For those who may have joined the last talk and seen the link, the link is also available here. You can click on the "last talk's summary," take a look at it while I'm talking, or take a look at it after, whatever. But I'm curious if any of you had seen it before and opened it because that's one of the things: how we went about creating that is something that I will be sharing. [My Most Memorable Anniversary](https://www.s-anand.net/blog/my-most-memorable-anniversary/) **Anand**: [03:20] I'll begin with my best use of AI this month, and the short answer is a little blog post. It was my 24th anniversary this year on 9th June. Around 10:30 PM, fairly typical, I went to ChatGPT and said, "Look, tomorrow's my 24th anniversary. It's a bit late for me to buy anything even online—maybe might work, but whatever. Can you give me some ideas?" **Anand**: [03:46] And I gave it some context: "Look, I uploaded my bank expenses last time, right? That'll give you a sense of the kinds of things I buy, my wife buys; maybe that'll give you a profile. You know us a little bit, do what you want." And it gave a whole bunch of ideas. But here is the thing: some of these ideas were pretty good. One of them was, for instance, effectively coming up with this statement. It said: **"Tell her tomorrow is our 24th anniversary. For the next 24 hours, I'm going to knock off 24 things from your to-do list."** **Anand**: [04:22] She's a to-do list freak. And at the end of the day, she said, "I think it was my most memorable anniversary. Thank you." Clearly, it was my most impactful use of AI. I'm finding that **it is not a powerful use of AI that makes a difference; what makes a difference, however, is the impact that it causes. Sometimes even simple uses have really powerful impacts.** And more practice, more habitualization, tends to lead to these kinds of things. **Anand**: [04:59] So, give it a shot. We have about 14 uses of AI and we will be coming back to this. But I'm going to segue—actually, no, I'm going to spend a few seconds maybe just quickly going through a few of those responses and seeing what we find. I know you can't see this; I'm operating on another screen. I'm just taking a quick look at AI use cases. Let me get rid of all the names for now. I'm just doing a quick filter. Okay, I've got a bunch. Let me put this up here. **Anand**: [06:24] **"Creating my own website."** That is ultra-cool. **"Researching new topics and stocks."** Powerful. **"A Gmail cleanup."** Yes, email management is pretty important. **"Deep dive into company fundamentals"**—fairly similar to the researching topics and stocks. **"Building a personalized web application." "Preparing for a work meeting." "Creating an app to track the World Cup."** That is for conversational purposes, fair enough. I guess then that is probably closer to researching. **"Live analyst preparation," "Work-related summaries"**—that, I guess, would fall under this category. **"Research on stocks," "Medical queries"**—that's more along the lines of personal use. **"Industry deep dives." "Creating a CV for my daughter using her LinkedIn profile."** Actually, we'll do a little more of this sort of thing today, maybe. **"Using Claude to review a redevelopment project."** That is an interesting, powerful use. **"Creating a bond portfolio"** and **"Promo video."** **Anand**: [07:46] That's a very different kind of use case. **"I started in the cab ride to the client—interesting—and were done by the time the elevator door opened."** Yeah, the speed of some of these is really powerful. **"Analysis of a convertible bond."** So, I'm guessing that this class of work that you've been doing—which is working with, let's say, stocks or bonds or doing industry deep dives, effectively some kind of financial analysis—is understandably fairly a common use case for this audience, and it's great you know what others are doing. This is probably not something that we'll do a strong deep dive into. **Anand**: [08:26] But on request from others, I'll just quickly show the QR code again and paste the link to the form for anyone who may have just joined or was not able to see the link earlier. We are collecting the uses of AI last month that were effective. Please do share because what I plan to do would—I mean, I'm just hoping to take all of these, cluster them, and see what we can learn from each other. And as Sonal had mentioned, she'd like to know how to do a Gmail cleanup; let's cover that. And Rajen would like to know whoever made the bond portfolio—would love to know more offline. Let's share that sort of a thing; the knowledge of these use cases really helps. **Anand**: [09:12] Let's go into the summary for the last talk. That is the link that you'd also seen from Sandeep. It's a recap of the previous session. It has the video, it has the audio, it has a detailed write-up in a story form of what we discussed last time, what you can do with it, what are the files that we uploaded, what are the examples that we looked at, what are the tools that we used, and eventually a quick comic summary of this and a series of takeaways that you can get out of this session. [Summmary of the previous talk: What Your AI Doesn't Know About You](https://sanand0.github.io/talks/2026-05-23-ai-unboxed-context-engineering/) **Anand**: [09:56] This entire preparation—meaning creating this story—took about an hour of elapsed time. It didn't take me the hour; it took Claude an hour with me coming in and correcting it and so on. **How does one go about putting together a summary like this and what are the parts involved?** That is what we're going to start with in today's session, which is about tools and workflows. It really should be titled "Workflows and Tools" because we'll be leading with the workflows and that will take us to specific tools. [Google AI studio](https://aistudio.google.com/prompts/new_chat) **Anand**: [10:33] In this particular case, one of the first things that I ended up doing was—let me open Google AI Studio. Now, **Google AI Studio is a pretty powerful tool, partly because it's free.** Anyone can just log in—and I'll put a link in the chat window—and partly because some of the models that Google has are particularly powerful. One of the really powerful Google models or capabilities of the Google models is transcription. I haven't seen many models, at least certainly for that cost, do transcription as well as the Google models can. **Anand**: [11:20] They also do pretty good rapid image generation. For slow image generation, ChatGPT is better, slightly higher quality, but for rapid image generation, it's very hard to beat Imagen or the new Flash models. One of the things about Google AI Studio is that you can put in a prompt. It's not meant for a chat; it is meant for single-use things. **Anand**: [11:43] And there are a series of different kinds of models that you can use on the right-hand side. Some of them are image models, particularly some of them are VEO, which are video models—and we'll come to video generation in a short while, that's a really powerful tooling. Audio is one of the fairly useful ones, both for text-to-speech—that is, synthesizing audio—as well as speech-to-text. But we don't need specialized models for that; the regular Gemini 3.5 Flash, 3.1 Pro, etc., can do a pretty good job of these. **Anand**: [12:22] And well, music—music in particular. Did we cover music last time? Does anyone remember? **Unsure**: [12:27] No, we didn't. **Anand**: [12:29] No, no we didn't. Okay. This is a really powerful one, surprisingly powerful. I did not think that music would have any kind of impact, but I'm just making a note to cover music as a tool today and how we can use it in a corporate environment. **Anand**: [12:47] The tool that we'll be using now is a regular Gemini 3.5 Flash. Now, what I did in order to create this particular session was first take the video and extract the audio out of it. How do you extract audio from video? Ask Gemini or Google or whatever; it'll tell you different tools depending on your operating system. You can download it and use it. That, other than referencing AI, does not require much by way of tooling. **Anand**: [13:17] And then with this audio file, what I did was upload it into Google AI Studio. I have a local recording of the current discussion that we are having right now, for example. So what I'm going to do is show you how I upload this. I'm just going to click on "Upload files," and in this folder somewhere, let me just select the audio file. And that audio file is getting uploaded. **Anand**: [13:52] I could just say "Transcribe." But apart from telling it just to transcribe, I usually have a series of instructions which I've found useful in the past, that if I explain to it, it gives it to me in a format that I find helpful. This is my "Transcribe Talk" prompt. I'll put this in the chat window. You will, of course, be getting all of these links in the form of a story. But this "Transcribe Talk"—and I'm just going to copy-paste this and then we'll read it out. [Transcribe talk: prompt](https://github.com/sanand0/blog/blob/9763887917a006c8aa76b4fc60ed2202548bab6f/pages/prompts/transcribe-talk.md) **Anand**: [14:28] When I run it, it will run in the background. But effectively, it says: **"Transcribe. Do not miss any part of the talk."** I find that—or I found that at least six to eight months ago—it sometimes missed parts of longish conversations. I don't want it capturing all my pauses, verbal tics, etc. Let it smooth over things a little bit so that the transcript is more readable. Correct spelling and grammar to some extent. In case I speak in Hindi or Tamil or any other language, then also translate that into English so that it's accessible for a larger audience. Sometimes it's not clear. **Anand**: [15:08] And you'll get a sense that, based on repeated transcriptions, I've found a certain set of things I don't want, a certain set of things I want, and all of that is what gets captured into this transcription. And AI Studio has now come up with this transcription. So it's saying, for instance: "Here are the initial questions. Sonal, do you have a question?" That was Sandeep asking and Sonal's response was, "Yes, very quick question, Sandeep," etc. Now here, it's putting it in the form of question and answer because I said "Give it in the form of a talk." I have another transcription version where I say "Guess the speaker names." Just telling it to guess the speaker names works well. [Transcribe call recording: prompt. Guesses speaker names](https://github.com/sanand0/blog/blob/9763887917a006c8aa76b4fc60ed2202548bab6f/pages/prompts/transcribe-call-recording.md) **Anand**: [15:47] Now, this is slightly different from the note-taking tools that we have, which are ultra-powerful, and from just using something like Whisper, which is again extremely powerful. I'll tell you how I'm using these tools. And Krishna says, "Can't see the writing very clearly." I'll expand it a little bit, Krishna. I suspect it may be the dark mode/light mode thing, so I'll see if I can switch to light mode on my laptop and maybe that will clear up some of the screens. **Anand**: [16:25] So, **there are three broad classes of tools that one can use for voice.** One is where you are talking and the system is listening: **dictation**. Mac Whisper is a fairly popular tool. You can just go to ChatGPT and dictate; it's an excellent tool. That is my tool of choice; I can use that on mobile. So that's dictation. **Anand**: [16:53] The second is **meeting transcripts**. Meeting transcripts are tools like Fireflies, Otter—Google Meet itself offers transcripts. In fact, on Teams, I can enable captions, and technically this is not transcripts, but close enough. Meaning if I could copy-paste it, I can pull it out. And it is possible to actually hack it so that you can copy-paste it; I've done that for Google Meet, and I can do that for—I'll in fact show you how we can do that for Teams as well and how you can start using it. But this transcription is very helpful when somebody is on a platform that offers transcription. On your WhatsApp calls, on face-to-face calls, that doesn't work. But maybe 70% of conversations these days, at least work-related conversations, are on platforms like these, so that's your second assistant. In that case, you don't need to do this transcription; they do the transcripts themselves. **Anand**: [17:57] The third is **notetakers**, where you can enable them on your phone and have the conversation recorded, or just use your phone's regular recording feature. Record a conversation and then go back to Google AI Studio and have it run. Now, Google AI Studio, you'll notice—and Vikas has in fact commented—it's mentioning a certain number of tokens. Tokens, as I mentioned last time, are roughly the number of words that AI processes, and often models charge on a per-token basis. **Anand**: [18:36] But there is an option here for you to put in a paid API key. However, you don't need to. **Google's deal currently with AI Studio is: "We will use your data for training. If you're okay with that, use it for free."** So if you're transcribing a non-sensitive conversation, go right ahead. The token count doesn't matter. It gives you a sense that if you had used the API, this would have cost you 14 cents for this model—and it's a fairly expensive model, Gemini 3.5 Flash. Earlier versions do an excellent job, and you can get the cost to one-tenth of this. Even if you were paying for it, you'd probably be paying no more than a few cents. **Anand**: [19:23] So, transcription is easy. Now, is it useful? That is a different question, and we will come to that. As we continue, just keep your questions flowing, please. I'll keep covering them and weaving them into the workshop. But before that, let me therefore propose something. Can we share the Google AI Studio thing? Let's see. There is a playground—is there a history to it? Okay, yeah. There is a "Share Prompt" option. Is there a "Share Conversation" option? Share Prompt... okay, maybe it'll work. Let's give it a try. **Anand**: [20:18] So what I'm going to do is share this and make this available. General access: anyone with the link. So anyone should be able to access the transcript of the talk so far and put it on the chat window. Could someone check if they are able to access the transcript output? **Sandeep**: [20:51] It needs access to your Google Drive, I think. Let's see. Ah, okay. Yeah, yeah, we can see it. I can see it. **Anand**: [21:02] Okay, perfect. Great. So let's do a small exercise. What I'd like you to do—and I'm just going to fill out this survey as well so you can have your voice transcript—the question is: **At what link can we see your shared Google AI Studio voice transcription?** And this is going to appear on your live form in a short while. So here's the exercise: please go to aistudio.google.com and select a new chat. I will paste the link in the chat window, and if anyone wants to share their screen and walk along, you're very welcome. Record anything you want. Just click on speech-to-text and speak into it, or upload any transcript that you have from before, however you want to put in some audio. Give it one word of instruction: "Transcribe." And see the result. **Anand**: [22:04] The model that you want to try out is probably Gemini 3.5 Flash. It's free anyway, but any of the Gemini 3 Flash, 3.1 Pro, is also fine. Any of the featured models is what you want to try out. It does not matter what other options you pick. Partly what I'm doing here is introducing Google AI Studio as a tool to you because it has some really powerful capabilities that you can access for free. And these are more niche, more tool-level than agent-level, but still pretty powerful. Let me share my link on the chat and the others—as we get more responses, we'll probably take a look at some of the others. **Anand**: [23:03] Question from Bharat: "AI Studio can do languages outside of English, I assume?" Let's find out. The short answer is yes, but... **Anand**: [23:19] _[In Hindi]_ Abhi main Hindi mein baat karne wala hoon. Aap ise angrezi mein transcribe kar sakte hain? Waise jab kar rahe hain to thoda colloquial angrezi mein kar lijiye. "Bro" waise kuch daal lijiye. _(Translation: Now I'm going to talk in Hindi. Can you transcribe this into English? While doing so, use some colloquial English. Throw in something like "bro.")_ **Anand**: [23:31] This, unfortunately, will just do a direct translation. So actually, to be fair, I shouldn't have told you to click on "Speech to text." You should ideally upload... but okay, ha, no, there is a "Record Audio" option. Let's see if that works. Click on "Record Audio." **Anand**: [23:51] _[In Hindi]_ Main Hindi mein baat karne wala hoon... oops. Main Hindi mein baat kar raha hoon. Ise aap angrezi mein chaap dijiye. _(Translation: I'm going to talk in Hindi... oops. I am talking in Hindi. Print this out in English.)_ **Anand**: [24:01] Okay, that does the recording. Then I can add to prompt. So now the audio that I've recorded gets added to the prompt and I can say "Transcribe." Let's see if it transcribes in Hindi or English, because it's capable of both. And now it's... okay, hmm, oh, okay, fine. Oh, this is three-level. So it's transcribing Hindi, transliterated into English, and translated to English. Okay, that's the best of all worlds, I guess. Certainly useful. --- **Attendee**: [00:06] Yeah, now I was running it in the background using Google AI Studio only. Open Google AI Studio, said "Live transcript," said "Fine, do it." So it was running, and this is what it gave, you know. **Anand**: [00:23] **And that is powerful, right? Because even if you're on a call and the other person says, "Oh, but I don't have permission to record it," okay, I also don't have permission to record it, what do we do? Open Google AI Studio and click on the record button. You're sorted.** This works. And I find that Google tends to have, at least for Indic languages, significantly better language recognition capabilities. Maybe for most languages; I've heard other people say that in general, Google's language capabilities are pretty strong. But for us, it certainly works. **Anand**: [01:02] Yeah, at least one more link has come through. Let me take a look at that link and see what we find. I'm going to open that. That's from Priyadarshini. Let's see. Okay, Priyadarshini, you've pasted the Google AI Studio prompt chat—and I'm sorry, you can't resubmit it—but what we want is after we create a prompt, there is a "Share" link at the top. Once you click on the share link, enable everyone to view the link, and then sharing is what will allow us to see it. **Anand**: [01:45] Do give it a shot. Just click on the—go to the link that I had pasted, click on the plus and record audio. No, it's not letting me do a plus. Let's reload. Yeah, plus... there used to be "Record Audio." Maybe if I remove code execution and a bunch of other things... Oh, sorry, yeah, you have to make sure that the model is one of the featured models. Gemini 3.5 Flash—that will let me record audio. Click on "Record Audio," start the recording, and then ask it to transcribe. **Anand**: [02:27] "Hi, what am I saying?" And add it to the prompt, tell it to transcribe, submit it, and that eventually, after it finishes the transcription, should let you—yeah, eventually—share the prompt. The sharing icon looks like this. When you click on the share icon, it will ask you what access you want to give. Change from "General access: Restricted" to "Anyone with the link," and then you can copy the link and paste it in the survey. **Anand**: [03:11] Vijay's question is: "**What is the temperature setting on the right-hand side?**" Actually, there are a whole bunch of settings on the right-hand side. Let me go through those one by one. These are actually—some of them are pretty useful. Where is the temperature setting? Interesting. Maybe you picked an earlier model that had a temperature setting and the newer one doesn't? Vijay, can you maybe unmute and point me to where you saw the temperature setting? **Vijay**: [03:55] Anand, it is on the Google AI Studio playground. That's what... and on my right-hand side, I can actually see the temperature setting. **Anand**: [04:07] What model? **Vijay**: [04:09] So Anand, when you are on Flash preview, you see it. **Anand**: [04:13] Flash preview. Okay, I see top-p... and I see temperature. Yes, okay. Because I was wondering—given that this is a somewhat outdated setting. Okay. So, **temperature is roughly the amount of randomness that you want to see in the words.** There was a time when people who were using AI were very finicky about the amount of control that they wanted, and with that control, they were sometimes able to get better results, worse results, etc. For example, let me say, "Complete this statement: The world's best..." with five options. Now, I'm going to set the temperature to zero and run it. And let's duplicate this tab; we will run it with the same temperature setting again. And duplicate this tab and run it with the same temperature setting again. **Anand**: [05:35] Now we have three results. It says "World's best food city is Lima, Peru." Second one says "Country to live in, blah blah blah." Third one says... okay, we saw "Country to live in" again here, but "World's best city." So it's kind of picking from a certain kind of responses. Now that doesn't yet tell us what temperature is. Let's set the temperature to a significantly higher value. To be fair, I should have probably picked a better example. Let's see. This one, new chat. Set the temperature to two and ask it the same question. **Anand**: [06:36] Now you'll find a couple of things. Okay, firstly... this is not such a great example; I should just probably tell you. But at the very least, you'll find that the style is different. It ended with "dot dot dot," whereas for the first... okay, no, even that is not true. Fine. Based on this example, I'm not able to tell the difference between temperature lowest value and temperature the highest value. I'll tell you how it works. **Anand**: [07:01] **When an LLM is thinking, it picks the next word.** Not very different from the brain—as we think, we kind of plan for future words, but at any point, we're really picking the next word to say, at least when we're speaking. We probably have a catalog of things—we might go in this direction, that direction, etc. We have maybe a 70% chance we go in this direction, 20% chance we'll go in this direction, 5% in a third direction, and so on. **Temperature is how much of that long tail it picks.** If you said, "I will set the temperature to zero," it will pick its first choice almost every time. If you set a high temperature, like temperature two, it will pick somewhat more uniformly. And therefore, **higher temperature means more creativity; lower temperature means more repeatability.** **Anand**: [07:59] If there's one thing that you need to know about temperature—well, that is probably the second most important thing you need to know about temperature. **The most important thing you need to know about temperature is that you don't need to know about temperature.** The temperature setting has been taken away in the newer models. The model providers are saying, "Look, we've learned enough over the last few years; we kind of know what we're doing. You don't muck around with this; you leave this to us." **Anand**: [08:26] But some of the other settings are important and reasonably useful. **Structured outputs will allow you to get the result in a specific format.** So supposing you always want: speaker, timestamp, quote; speaker, timestamp, quote, etc. Then that's useful. So you could say, for instance, I'll add a speaker, and I'll add a timestamp, and I'll add a quote. Actually, technically, I want a speech array. So this should really be... how do you make... okay, let's call this "Transcript," and inside that... "Transcript" can be an object. Inside that, we want a "speech," which is... okay, sorry, yeah, "Transcript" is really an array which contains a "speaker" and a "timestamp"—which can also be a string—and "speech." And I can get rid of this. **Anand**: [09:41] So if I put in this as the structure and asked for an output saying, "Generate some random transcript" and have it create an output—even though I've just said "Generate some random transcript" and it will generate something—it will generate it in the structure that I wanted, which in this case, as you can see, begins with a transcript: there's a speaker, timestamp, speech; speaker, timestamp, speech, etc. Which is great to then upload into Excel. You can then say, "Here are the columns that I want. Here is the output that I want." If you're taking an earnings transcript and converting it out here, you want it in a specific form. You say, "Get me all of the important announcements along with the numerical values. Add a field for the numerical value." That is pretty useful. **Anand**: [10:32] Next question: **How do you convert this JSON into Excel? The answer, by the way, to any question we have these days is ask AI.** So copy this, put it into AI and say, "Give it to me in Excel." It'll give it to me in Excel. Don't even think twice about these sorts of things. Next is code execution. Code execution lets you run code; we're going to cover that in the fourth session, not now. Function calling is also something that we're going to cover in the fourth session. **Anand**: [11:01] **Grounding with Google Search is exactly what it sounds like. It will—if you tell it to—search online and verify.** So again, let's say we are doing—let's say we upload an earnings transcript and ask it to fact-check each of the statements that the investor relations team has made. There, grounding with Google Search is pretty useful; it'll do a search, check if it is valid, and add a little comment against it. So by now, you're noticing that it's not just documents; it's also voice that we can upload. Of course, we could upload this into Gemini, we could have uploaded this into ChatGPT, etc. This is a tool that is available for free for public data that also happens to have some of these controls. URL context is simply where, if you paste a link, it will also read the link. These are far less useful; I'm going to skip the advanced settings. **Anand**: [11:55] Let's take a quick look if we have... okay, we have one more submission from R D'Silva. Let's take a look at what that was. Okay, looks like I'm not able to access it. That is because the link that was shared is probably a link just to the prompt. Give it a shot. I'm going to move on in the interest of time. **Anand**: [12:30] Remember that what I said was I'd cover how we put together the summary of this talk. **Step one was transcription. And that transcription ended up—and I'll send you a link to the full transcript as well—here. Here is the link to the transcript which I'm going to paste on the chat.** Okay, well, I did type all of the stuff that's above the word "Transcript"—I copy-pasted from my preparation, so ignore that. This is what the transcript mentioned. Now you notice that this transcript is close to production-ready. Meaning it has the timestamps at the right place, it has important stuff that is in bold, it has who said what—what Devi said, what Sourabh said, what... Sorry, it should have been Sandeep. Transcription error or prompting error one way or the other. [Transcript of the previous talk](https://github.com/sanand0/talks/blob/294a1b761b5a7c492826155d2796d36a770f14cb/2026-05-23-ai-unboxed-context-engineering/transcript.md) **Anand**: [13:37] But now that we have this, the next step is then to start producing the story. And I'll come to that, but before that, there was something else that I did, which is ask it to create a comic summary of the same story. Why? Partly for effect, partly those who lack time and want a simple summary can just go through what the takeaways are. And this, as we saw last time, creating images is a pretty easy thing. Creating images in any specific format is also pretty easy. **Anand**: [14:13] Sumit had an interesting point. He spoke something in Swedish—congratulations, you know Swedish—and Google thought that it was Danish. And the translation was correct, though, which is interesting. Danish reminds me, we actually did an engagement on generating Danish podcasts. I'll see if I have... yeah, this is... I'm just going to play this. Yeah, I'm going to send you this link. **This is a collection of various Gemini text-to-speech voices.** The same Google AI Studio that we saw doing text-to-speech, generating in a bunch of different voices, except that these voices are all Dutch. Why? Our client was Dutch-based. They said, "We haven't really heard any good Dutch voice quality." And it turned out that... who was that person they were really happy with? One of these voices—and actually, they were reasonably happy with all of the voices—but one particular voice in this they felt—which was a male, reasonably deep voice, probably Orus, maybe, not sure—was like spot on. It catches the accent perfectly and all of that. [Gemini TTS Voice Catalog - Dutch](https://files.s-anand.net/temp/gemini-dutch/) **Anand**: [15:44] This is four months ago—three months ago, at least—since then text-to-speech has been steadily evolving. Of course, I hope you know that on ElevenLabs or a bunch of other tools, you can just upload a few seconds of your voice and have it say something in your voice. We'll in fact do that in a short while; that's part of the tools that we'll be covering. But it is shockingly easy to be able to reproduce an accent, and there is a long tail, of course. Understanding accents also, there is a long tail, but we're finding that it's steadily getting better and better. **Anand**: [16:20] **What are the implications? One implication is at Straive we said, "You know, we need a podcast from Straive CEO Ankor." He's busy, not available. Fine, we'll create a podcast.** We anyway have a transcript of what he said in meetings, put that together, create the transcript, and it effectively becomes him. Of course, the other far more useful scenario is: find a friend's voice, scam them for some money. That's pretty easy. Anytime you pick up a phone and you hear a familiar voice, you know you've got to be a little more careful. If you hear an official-sounding voice, you know you've got to be a little more careful. But **text-to-speech has progressed a lot.** Maybe we'll cover it today. [ElevenLabs voice cloning](https://elevenlabs.io/voice-cloning) -- PS: This was not covered today. **Anand**: [17:03] But coming back to the comic strip version. All I did was—step two—copied the transcript, pasted it on ChatGPT. Let's see if I can locate that. Yeah, this is the one. And put in—so this is the transcript, this is the prompt that I had provided on top of it. I'll share this prompt with you on the chat. There are a few elements of this that I think are worth highlighting. First, **I am asking it to draw a full comic page out of it.** There's no reason it needs to be a comic page. There are several different styles of illustrations that we can draw. I may have covered this last time, but if not, then let me share a link to at least my catalog of art styles or illustration styles. [ChatGPT chat that drew the comic strip for the previous session](https://chatgpt.com/share/6a362f1d-8e48-83ee-93a0-5000819d8101) [LLM Art Styles Catalog](https://sanand0.github.io/llmartstyle/?category=text2) **Anand**: [18:23] If you have lots of text that you want to convert into an image: **exploded diagrams are a useful format, alluvial flow diagrams can be pretty visually impressive, there are cross-section cutaways that let you dive into a specific area.** You can have wayfinding systems like metro maps, you can have dot chart-like systems. Different systems exist, and a list of prompts to generate these is what I had cataloged, again using a combination of Claude, ChatGPT, etc. And what I do is glance through these, find out which style might be most appropriate for the next session, choose that prompt, and use that as the first section of this prompt. ![Exploded Diagram](https://github.com/sanand0/llmartstyle/releases/download/text2/game.exploded-diagrams.gpt-image-2.png) ![Alluvial / Flow Diagrams](https://github.com/sanand0/llmartstyle/releases/download/text2/game.alluvial-flow-diagrams-as-illustration.gpt-image-2.png) ![Cross-Section Cutaways](https://github.com/sanand0/llmartstyle/releases/download/text2/game.cross-section-cutaways.gpt-image-2.png) ![Wayfinding Systems](https://github.com/sanand0/llmartstyle/releases/download/text2/game.wayfinding-system.gpt-image-2.png) ![Dot Charts](https://github.com/sanand0/llmartstyle/releases/download/text2/game.unit-dot-charts.gpt-image-2.png) **Anand**: [19:09] The other two things that I mentioned, apart from style stuff, is: one, **think about the most important points and structure it as a memorable story.** What we forget is that image generation is a tool call; that is, it is a capability that AI has. **AI these days works in an agentic way. What that means is it thinks, does something, thinks again, does something, and keeps doing stuff until it completes a job.** A couple of years ago, you'd give it a prompt, it would give an answer. That is exactly what AI Studio does as well; you give it a prompt, it comes back with an answer. But Claude, ChatGPT, Gemini, etc., have progressed beyond that, and internally they do a fair bit of thinking, and you would be able to see that thinking in that thinking pane. > **ChatGPT thinking**: I need to make a 3:4 portrait comic page summarizing a workshop on context engineering, featuring full-color diagrams and captions. It could have a generic character, maybe a charismatic presenter with glasses and salt-and-pepper hair, since I don’t have an exact photo. The visuals need to be vibrant and include metaphors. **Anand**: [20:02] In this case, it thought for four—a little over four minutes—before it generated the output. And in order to do that, it did some internal prompting itself. It said, "Okay, I need to make a 3x4 comic page summarizing this. It could have a generic character, maybe a charismatic presenter with glasses." I didn't mention any of this. These are not things that an image model does by itself, but that's what an agent is for. **An agent is a human-level—or smarter than human—reasoner who puts stuff together and then orchestrates and calls the various tools.** Which means that there are two different axes of AI development today. One: the agents are getting smarter and smarter so they know what to do. Two: the tools are getting more and more powerful so that the agents can do more. And that combination is what we're exploring. **Anand**: [20:53] Another lens to think about it is: last call, we spoke of prompts, which was about how we get more from the intelligence of the model. **This session is a little more about how we get more from the capability of tools by broadening the set of things we can do.** And in that combinatorial explosion is where most of the power lies. **Anand**: [21:18] But coming back. Because we're asking it to think through what it should cover and how it should cover it, etc., think about the most important points—it manages to do a better job than just a naive "Here's the transcript, create a story," in which case it might have just created a story that was sequential. But in any case, we have the comic story. And I'm not going to ask you to generate this because last time we generated an image, that's easy. And then you export it, you compress it—there are several tools to compress it. **Anand**: [21:59] **Just FYI, when you get an image from AI—actually when you get an image from anywhere—they are massive.** You may be sending it as an email. Now that email, let us say it is 2 MB. This one, for instance, is almost certainly more than 2 megabytes. Let me download it and I'll see how large it is... Gemini image, it is... yeah, 2.8 megabytes. Now if I send this 2.8 megabytes to 1,000 people, that is 2.8 gigabytes sitting in their inboxes. Every week if I send one of these, by the end of the year, that is probably 100 gigabytes. Our IT teams will thank us if we were to compress images. So **in general, if you're sending images, please compress them or put them on a server.** **Anand**: [22:54] And **one of the best tools for image compression is Squoosh.** I'll paste a link here. This is not really AI-related, but I think is worth using. And you can just paste an image—or if I were to take and upload this image—you can choose from a variety of different compression formats. **The best compression format today is AVIF.** AVIF. With regard to images, probably the only thing that you need to know is: **compress using AVIF on Squoosh.** You can see the difference between the original, which is on the left, and the new on the right. And update the quality. So now what it's done is compressed it by 96%. I can say, "Okay, I want better quality," get it to 52% quality. Now my compression will fall to 91%, but it's still more than 10 times better, and honestly, I can't tell the difference between the original and the new. So just let people know that compressing images is a good idea. **Anand**: [24:01] Step three: **How do we go about creating a story? This is actually easy.** All I did was—and I will share the prompt that I used on the chat and I will also put that out here—**I have a "Talk Story" skill.** Last time I briefly touched upon skills. I'll mention them again in a sentence, but today we will be going a little more into how you create and use skills. **A skill is effectively telling an agent, "Here's how I do stuff."** I'm framing that carefully. It's not how _to_ do stuff; there are many ways of doing stuff. It knows what the world does generally. "This is how _I_ do stuff," and therefore indicates a preference. Secondly, a skill is something that it can automatically pick up if you put it in the right place. And I'll show you where you might be able to put in a skill. --- [Talk story skill](https://github.com/sanand0/talks/blob/294a1b761b5a7c492826155d2796d36a770f14cb/.claude/skills/talk-story/SKILL.md) **Anand**: [00:00] But this skill says, if I call it this particular name or that, step one: find the top directory, make sure that there is a transcript, look at all of the stuff that's there—usually these are the kinds of things that I put in there. I might have put in some PDF slides, if so, convert them into images. Check for the context. Is this already there? Then use it. If not, then screenshots. In case you don't have screenshots of web pages, go take screenshots. Here’s how you take the screenshots. Here's how you read all the content. Make sure you create the HTML in chunks because Claude will get stuck and I've seen that kind of a problem. Update the final README and so on. And **this is how all of my talks end up getting converted.** So for instance, since our last talk, I've added four others, and the link is just one level above the earlier one. And they all follow exactly the same skill. **Anand**: [01:04] I'm actually going to wrap this up in a minute. But Sandeep shared a very interesting image, which does not at all look like a comic. Sandeep, could you share the prompt you used for this one? **Sandeep**: [01:16] Actually, I did. **I just used [NotebookLM](https://notebooklm.google.com/). So I just copy-pasted whatever that raw format was, the data, and then in NotebookLM you have the infographic button. I just pressed that.** Then I tried all other things, and it kind of gave me rubbish answers, so this was the closest. **Anand**: [01:33] Right. No, firstly, this looks nice. And yes, it's not quite a comic format; this is an infographic format. It's a great infographic-level thing and interestingly doesn't seem to have any obvious spelling mistakes that I can spot, which means that the Gemini models are getting slightly better. **Nano Banana 2 was good, but usually with this much of text it would have made one error, and I'm not able to instantly spot it.** **Anand**: [01:59] But here's what I suggest: you could take the earlier comic prompt that I had provided and try it on [NotebookLM](https://notebooklm.google.com/) itself and see where it gets you. It might work; it'll certainly work on Gemini. I find that **ChatGPT has the best image output quality today.** This wasn't true three months ago—Nano Banana 2 was the best at that time—and it keeps toggling, but it's a marginal difference. Gemini is much faster. **Anand**: [02:49] So, how can you use some of these techniques? For me, if I’m giving a talk and summarizing it and sending it, yeah, that could be useful. But how is that going to be useful for you? Let me share what happened day before yesterday. We had an education services client come over. And we were in Mumbai having a long discussion that was talking about what kinds of AI they might use. At the end of the session—which was recorded by our note-takers—I took exactly that process. Took the audio, converted it to a transcript, converted it to a comic strip, converted it to a story using exactly the same format with all of the hyperlinks, and sent it to them. **Anand**: [03:39] They were, obviously apart from being happy that they got a summary of the conversation, blown away by the richness and the intelligence of what was captured. And clearly, it also had enough steps as a follow-up. So **from a business development perspective, if we're saying attention is a scarce commodity, at least for the next few months, maybe a year, new formats like this are a strong attention mechanism.** **Anand**: [03:59] Whether it's comics as a summarization method, whether it is infographics as a visual medium, whether it's data stories as a way of summarizing—we had a sales meetup last Thursday in New Jersey, and one of the things the leadership team decided there was: no presentations. Only HTML. There is a gentle shift. **HTML output, data stories like these are easier for AI to create and for us to consume. We may as well shift. So I expect we'll see a lot less of PowerPoint going forward.** **Anand**: [04:53] Let me share the screen and see Sonal's output. Okay, that is probably the same prompt rerun, Sonal. Nice. Clearly a different comic style. **Sonal**: [05:04] Yeah, same prompt. It took time. It took about three minutes plus. I thought it was going to collapse on me, but it finally came out with something. **Anand**: [05:13] ChatGPT tends to get there, but yeah, mine took as much as four minutes. Question from Rohit: "You just uploaded the text and asked it to create the comic story?" Yes, Rohit, exactly that. Though there is probably one other thing that I should mention I did—I'm not sure if this is optional—but in the "plus" icon, there is a "Create Image," and I chose that. You can also do that by pressing "@" and you get the "Create Image," and then you can mention it. **Anand**: [05:54] But let's try what happens if I don't say that. "Draw me a comic. Any comic. Be quick about it." And I'll change the thinking to instant. So let's see if this is smart enough to... okay, yeah, I think it's probably going to create a comic. It's still... okay, "Creating Image," that was fast. So it might work. Ignore... in case it does not create a comic even when you tell it to, then select "Create Image," but otherwise you don't even need to do that. You just upload the text and ask it to create a comic story. [ChatGPT conversation: "Draw me a comic. Any comic. Be quick about it."](https://chatgpt.com/share/6a375036-781c-83ee-939a-53d485f7517f) **Anand**: [06:37] Sandeep asks: "Can you paste the prompt here, Sonal?" Yeah, it's the same prompt, Sandeep, that we had in the earlier chat that I had sent. And yes, Sonal has shared it. **Sandeep**: [06:43] No, no, go on please. I just couldn't find the chat. **Anand**: [06:51] Ah, sure. And why is this different from what I did? Any time you ask an LLM for an output, it gives you a slightly different answer. We saw that a short while ago. Now, actually, it's a deep question as to why that happens. And the interesting—or rather, there are levels of reasons why that happens. **Anand**: [07:07] **One level is the temperature that I mentioned. LLMs are explicitly told to kind of randomly pick slightly different stuff because that is how it seems to match what humans say, write, read, etc. But even if you set the temperature to zero, in theory, you should have perfect reproducibility.** The trouble, however, is that at the GPU level, there is a certain amount of non-reproducibility because of floating-point errors. **Anand**: [07:44] What does that mean? The way we store large numbers involves truncating decimals. So if I take, let us say, 0.1, round it off to zero, and I have a hundred 0.1s all rounded off to zero—when I don't round them off, they will add up to 10. When I round them down, they will add up to zero. When I group them in different ways, depending on whether I group them to more than half or less than half, they will add up in different ways. **Anand**: [08:14] **A neural network is a way of grouping numbers and, by the nature of the way in which numbers are stored in CPUs, there is always a certain rounding that happens. And every time you ask, depending on how the rounding ends up getting grouped, we get different answers.** In other words, we haven't yet figured out a way of making sure that the same input gives the same output almost no matter what, unless we're running it on exactly the same chip at the same temperature and all kinds of crazy things. And, yeah, here's a styled 1940s DC Comic on the chat, okay, that's interesting. **Anand**: [08:53] Now, with this, let's move on to our next step. So far we've talked about how we can use audio, convert that to text, how we can use comics as part of a workflow, how that can help in real-life situations, create a data story out of it, etc. There's one part that I didn't cover, which is: how do you actually get it to create a data story like the one that I shared? And how do you publish it? That we will probably cover in the last session, or maybe even the next session. Actually analyzing data or transcripts or anything and coming up with an output. **Anand**: [09:43] But **one of the other classic workflow issues that we tend to have is: there are errors in workflows. How do we deal with that? How do we deal with hallucinations?** Let's, in fact, do a quick survey on that. And I'll frame the question as follows. In fact, let me just dictate that. **Anand**: [10:13] **Hallucinations are actually not very different from people lying or making mistakes.** Sometimes we don't know enough about what they've done or enough about their fields of expertise to know and judge. How have you seen this sort of situation being dealt with effectively in the past? In other words, **how do you verify humans? And what are the most effective methods?** Let's frame that as a question and run that as a quick survey. **Anand**: [10:52] Hallucinations. Because what I'm about to share will be particularly useful once you've formed a hypothesis on this. So, Question number four, which will appear here... oh okay, I need to go back... Questions again. And yeah. If you could take a shot at how you deal with hallucinations or errors by people—your team members, your colleagues, your vendors, your boss. Anyone comes back and says something, how do you tell? How do you verify? How do you fact-check? What's been most effective for you? **Anand**: [11:47] For me, **one approach is I ask other people to cross-check.** That's helpful. I ask ChatGPT; that's also been helpful for me in cross-verifying. And a third thing that I've found pretty useful is: check if specific data points I know are correct. These would probably be my top three responses. Once we have maybe three or four, I'll take a look and we'll share what are some of the popular workflows for validation. **Anand**: [12:50] Okay, I will share the link to the form and the QR code is right on top of the screen. Okay, we have six. Okay, let's go through some of the responses that we have so far. I'm just going to read them out from my other screen. **Anand**: [13:19] Okay, **"Ask Claude or GPT to cross-check."** Yes, that’s from Sandeep, and that's a good one. Arvind suggests **asking the same LLM to check itself again.** Claude most of the time says it missed something. Yes, one of my classmates used to do this to me. Every time I'd say something, he just would ask, "Are you sure?" And then I'd think for a little while and say, "Okay, yeah, maybe I got this wrong, maybe I got this wrong." **Anand**: [13:45] I’m trying to remember who it was, this diplomat who was really famous for saying this. He’d always ask, "Is this your best work?" And the person would say, "Okay, maybe not," and improve and come back. And he’d ask the same question, "Is this your best work?" And they’d iterate, and he’d wait until they said, "Yes, this is my best work, I really can't do better than this," at which point he would review it. What is the point of reviewing something that is not somebody's best work? So yeah, that's a good technique. **Anand**: [14:22] Rajen says, **"Cross-check with what I knew."** Yes, a dipstick check on data points that we are familiar with; that's powerful. And then **"Waking up till 2:15 like yesterday to try AI and correct myself."** Yes, totally. Sleep over it and sometimes you get a different point of view. Just bang your head against it and you get a point of view as well. **"Doing a quick random recheck."** From Anand—that's another useful technique. Pramod suggests **people tend not to lie, but it's more the information they are aware of.** Okay, we're getting a reasonable number of responses. **Anand**: [15:01] Now, let's do a synthesis on this itself. What I mean is, between us, we have a series of responses. Let me consolidate all of these. All right, I have the responses out here. And the question is out here. Let's ask. **Anand**: [15:55] I asked the question below and got a series of responses from people. **Could you cluster them and tell me which are the most common techniques that people were able to apply, and which are the biggest gaps in techniques that therefore they could learn from?** [ChatGPT conversation: Verification Techniques for Hallucinations](https://chatgpt.com/share/6a375149-92d8-83e8-b8dd-80e275973fd1) **Anand**: [16:18] You'll notice that **this is the kind of thing that we can do in the middle of meetings. Transcribe what somebody's saying, put it out there in the chat, ask for a response, share the response, make it look like you're ultra-smart. It works.** I do this often enough that I just ask people to give me a few seconds of thinking, whereas in reality, _it_ is thinking. **Anand**: [16:47] And increasingly this will become a powerful technique. I would strongly suggest if you have a team where somebody's in the middle of it saying, "I'm thinking, give me a minute or two," and actually take a minute or two, they're probably using AI. Please promote them; they will likely do a good job of at least leveraging a free team member. **Anand**: [17:03] Okay, so it’s saying **cross-checking with another LLM is the most prevalent technique.** Conscious that there may have been a few responses since... yeah, okay. **Comparing against known facts** is another equally popular technique. Spot check is popular. Logic or anomaly detection is popular. **Asking for sources**—this is actually a very powerful one. Where did you get this from? Give me the link. I will click on it and check if it is actually there. That is a very useful one. And asking other people to verify. **Anand**: [17:38] So, LLM triangulation, known fact anchoring, etc. What are the biggest gaps? **Source hierarchy.** And it's saying that source hierarchy is powerful beyond just asking for a specific source. But also, yeah, just asking for sources we mentioned once. **Track record checking.** Yeah, this is a good one. If I find that a model tends to give a certain class of answer right and I've checked it three or four times, the fifth time I usually say, "Yeah, this is probably going to be right." Like, **I no longer check ChatGPT's images for spelling mistakes. I've checked it half a dozen times, I have not been able to find any. I do still check Gemini's image generation for spelling mistakes, because one small mistake I tend to find every day.** **Anand**: [18:31] **Incentive checking.** Okay, that is—yeah, is there an incentive for the human to say what they said? With AI, I don't know how effective that would be. **Expertise boundary detection** is an important one. Does it know enough about the field? If the field is, let's say, bond market research, I'd say yeah, it probably knows as much as many experts—may not be the best expert, but close enough. Does it know enough about what my wife would want for her anniversary? Absolutely not; you are not the expert by any stretch of imagination. I may have to take that back, but anyway. **Anand**: [19:18] **Counter-example search is a very powerful one. Asking it to find a counter-example, and agents are very diligent.** If you tell it, "Find me errors in your logic," it will. Reproducibility, yeah, multiply... and I'll share this. **Anand**: [19:39] Now here's the thing. This is a bunch of responses. Not all of them are equally effective. Also, not all of them are equally easy. **What I want you to do is give me a table that has listed the techniques along with two columns: how impactful is it when applied with AI, and how easy is it to use with AI?** Finally, also give me a crisp summary of the top three techniques that I should take away when verifying AI. Effectively, quick little prompt fragments which I can just add at the end of any conversation—ideally no more than a sentence—which will have the highest effectiveness and highest probability of success for detecting hallucinations and making sure that the answer is correct. Let's give that a shot. **Anand**: [20:29] And again, we're just using the same prompting engineering techniques that we spoke of last time. **Be lazy. Let it do the work.** You tell it what you need, don't spend too much time verifying it. And we're meta-applying that principle here by saying, "I want verification, not just effectively, but also with ease." **Anand**: [20:51] And asking for sources or evidence requires some amount of effort from our side. But asking it, "Look, tell me what you're sure of, not sure of, and mark it, maybe red-amber-green or high confidence, low confidence"—very easy for me to verify. I'll just filter out all the high confidence ones and drop the rest. Claim-by-claim verification, etc. **Anand**: [21:11] So, we have a list which I'm going to share and this is by no means exhaustive. I do have some better approaches. But it's saying if I had... okay, here are three techniques. But if there's only one suffix that you can add to prompts: **"Break this into key claims, mark certainty, identify the five highest risk claims, and tell exactly how to verify or falsify them."** Great. Let's apply it to this conversation. Apply that to this conversation. And I'll have the... I'm not going to continue sharing that. **Anand**: [21:54] But one thing that clearly emerged is: **we are using LLMs as a judge, and that in itself has a certain class of benefits.** But it went on slightly beyond just using LLMs as a judge because we're asking for source, we're asking for through-citations, we're asking for confidence levels, and that brings a certain structure in. A few other things that are happening from a verification perspective. **Anand**: [22:21] But the workflow that we spoke of here, that is: **when you do a task, after that task, just make sure you automatically, as a habit—and perhaps even as a skill—verify it.** That's probably one of the most critical workflows that we find. **Anand**: [22:42] Something that I may have shared last time, but because this is important enough I will share again: we did an experiment. **We asked different models to classify different statements**, like "I want to update the address." Now, does this, when GPT-4o mini classifies it, did it put it into the right bucket? One model said, "I will call it 'Change Shipping Address'," but it should have been the "Set Up Shipping Address" category. [LLM double-checking](https://sanand0.github.io/llmevals/double-checking/) **Anand**: [23:17] So, when I wanted it to categorize a chat message into the right bucket, some models got it wrong and right at different levels of accuracy. The interesting thing is that **when we got models to cross-check each other—that is, when one model was checking, on average, it made a 14% error. When two models cross-checked and both of them agreed, the error was only 3.7%. When five models cross-checked and all of them agreed, the error was only 0.7%.** These are the clear wins. **Anand**: [23:54] Now, that means that when they disagree we have to still manually verify or do something that is slightly more expensive. But that happened to be only 28% of the time. So **I'm getting 99.3% accuracy with a 72% reduction in effort for negligible cost.** These are trivially cheap models. That's a decent way. So double-checking, triple-checking, quadruple-checking, etc., is a pretty good way of automating in production verification. **Sandeep**: [24:30] Question from Sandeep: **"Can we create a Claude skill for this, which runs after every workflow?"** **Anand**: [24:37] Yes, we can. So now let's create a Claude skill. In fact, let us specifically create a Claude skill for this purpose. Now, those of you who are using ChatGPT or Gemini, unfortunately, there is no default mechanism by which the agent will pick up the skill unless... --- **Anand**: [00:00] ...unless you're on the Pro or Enterprise account with ChatGPT, where there is a skill catalog. Claude even on the $20 version has a skill catalog. You might remember from last time I was suggesting keeping a prompt library of copy-pasteable stuff. Worst case, just keep it as copy-pasteable. **Anand**: [00:18] So, I'll suggest a prompt that we can all try out. **This prompt will help us create a verification skill.** And that verification skill will get added to the skills repository, or you can add it to the skills repository. But first, what are skills at the next level and where are they stored? **Anand**: [00:41] Anthropic came up with the concept of skills. **Skills are stored in Claude under "Customize".** Any other agent other than Claude, you'll just have to maintain them in your notes. **Anand**: [00:54] These skills—so here are some of the skills that I have. I have an "Expert Lens" skill, I have an "Anand Writing Style" skill, I have an "Ideation Protocol". I'll go through some of these. Each of these skills has three parts, apart from a name. So let's take "Anand Writing Skill". This is a skill very specific to me, how I write. And it has a description which says "Write in Anand's style in blog posts, talks, summaries, etc.", and the details of how to do it. It says, "Make it easy to read, jump straight in, blah blah blah, I am human and flawed, keep the voice curious, style, message, content, format." All of that is captured based on my past iterations. **Anand**: [01:41] Anthropic by default also includes a "Skill Creator" skill. And the Skill Creator skill is pretty useful in, as you might have guessed, creating skills. So **creating a skill sometimes is as simple as just telling it, "Create a skill for X."** **Anand**: [01:58] So, let's take Sandeep's question. "Can we create a Claude skill for this which runs after every workflow?" I'm going to paste that almost verbatim and say, "Create a Claude skill for verifying answers which runs after every workflow." [Claude chat: Automated answer verification skill](https://claude.ai/share/7f7d657a-42af-46b3-b73c-d3533fa99416) **Anand**: [02:20] Now I'm going to burn some tokens on this. I'm going to put the effort to focus 4.8 effort max, let it—and thinking is enabled to basically go the whole hog. Now, a lot of my conversations are already on Claude. I have already done some cross-checking. It probably knows how I cross-check. It is smart; it knows how other people cross-check. It should be able to search online and find out the latest best practices on this, etc. **Anand**: [02:53] But though the prompt is the same, the context is completely different and therefore the prompt which I just typed also on the chat window, if you were to try it, is likely to come up with a different answer. You can try this on anything—Gemini, ChatGPT, Claude, whatever. The only difference is what it provides in Claude, I can just directly add it into that skill, whereas here we will not be able to. **Anand**: [03:19] Here is the request: copy-paste this, modify it however you want, and please share your chat output. And I'm going to do the same. Let me add that as a question in the survey that we can do, sharing on... **Rohit**: [03:41] Anand, your screen is not shared right now. **Anand**: [03:45] Oh, I'm so sorry. Thank you for flagging that. So, sorry, this is, yeah, all I typed—same thing that I put in the chat window. And now I understand the question Rohit that you had asked. You just go to any agent and type in this prompt. But **in Claude, the list of skills are under "Customize", and under "Customize" you will find "Skills".** These are my skills. Sorry for that, I'll just keep my screen shared. **Anand**: [04:26] Question from Rajen: "Do we really need to set up a skill file in Claude or will it build it up on repeated interactions?" It operates at two levels. All of the agents operate at two levels. One, they have access to past memory unless you tell it not to look at past conversations. So even if you didn't say anything, it's automatically building up context and if you had asked it to verify in certain ways in the past, it might pick it up. **Anand**: [04:58] **A skill is something where it will go through all of the descriptions in every conversation and decide whether it should pick it up.** So I'd say it will—if it's in a skill, it will probably pick it up. If you want it to definitely pick it up, put it in the prompt. **Anand**: [05:16] And ChatGPT suggests adding a short verification footer to prompts and that is a good idea, which is exactly how you would think about a skill as well—you're just copy-pasting it from somewhere. In Claude, it is automatic, so even if you forget, it does it for you. **Anand**: [05:33] Madhu asks, "Do you update your user instructions to Claude to shape how it responds?" Yes, I used to, and my current user instructions have only one line. I will show you the evolution of this, Madhu. Let me take maybe a minute, no more, to summarize. **Anand**: [05:54] So, in settings—whether it's Claude or ChatGPT or Gemini—there is a box somewhere where you can give these agents a general instruction which is sent to it every time. I used to have a long, detailed list of instructions: "I am so-and-so, this is what I'm interested in, this is how you should respond," and so on. **Over time I realized when I reviewed those, it was doing that even without me having to tell it. The models are steadily improving.** The applications around the model are steadily improving. [ChatGPT Custom Instructions](https://www.s-anand.net/blog/prompts/chatgpt-custom-instructions/) **Anand**: [06:27] Therefore, earlier I used to put in stuff here to correct their deficiencies. I have stopped doing that as much as possible. Today in Claude there is only one deficiency that I see, and I don't see it as a deficiency, more as a stylistic disagreement. It prefers putting in HTML visuals in the output; I prefer Markdown that I can copy-paste and share—purely stylistic, that is it. > **Claude instructions**: Answer directly in Markdown rather than artifacts (embedding HTML, visuals) unless requested. **Anand**: [06:53] There are some corrections and some personalized context that I give, and I provide that more to ChatGPT because ChatGPT is where I'm sharing it a little more of... I just feel I need it a little more in ChatGPT. And that is in the personalization section where I have given it a little more about me. This really I should delete because by now it knows this and a whole lot more; this is really not required. **Anand**: [07:23] I did have "Goals" as a more detailed prompt and I crafted this yesterday, I'm still trying to figure out if I should put it as a skill because it doesn't often align with my goals. So maybe that is the one section that I will keep. In short, I'm reducing the size of my custom instructions. **Anand**: [07:44] Okay, longish comment from Bikram which I will come to in a minute. Oh wait, I had a Claude prompt somewhere, did I delete it? Thought I... let's go back and see if this conversation is done. So it's done a lot of thinking and it has created this skill, which I guess I would need to... okay, fine, finished. Let me first share this and you will have a copy of the skill that has learned from my past conversations. I'm going to put this as the last question in the form, which is the "Verification Skill" question: "Please share the public link to your verification skill." **Anand**: [08:52] And that in my case is out here. Do remember that if it has the word "/share" in it—if it is Claude, or even ChatGPT it should have the word "share"—only then it will work. You have to click on the share link on the top right and then copy the link and submit. Otherwise, just copying it from the top where it says "/chat" will not allow us to learn from each other. **Anand**: [09:19] Now, what I'm going to do is take this, copy the whole thing. If you are using Claude, paid or free, you can go to **"Customize" and click on "Skills". There's a plus that lets you create a skill by writing the skill instructions. Paste it.** Now, a few things you'll have to do manually, like moving the skill name to the top, moving the skill description to the description section, and getting rid of whatever looks like rubbish. And that will give you a new skill. **Anand**: [10:02] How good is the skill? You'll know only after using it a few times. How do you know if it is using it? It will tell you: "I'm reading this skill." Great. Take a look at the output. Take a look at the skill. See when you find something relevant. **Rohit**: [10:17] Rohit says, "Can't see the plus sign when I go to skills." This is where it should be. Maybe it does not appear on the free version and you have a free version, Rohit? Do check, I'm not sure. **Attendee**: [10:33] It's a paid version. **Anand**: [10:35] It's a paid version? Okay. It should absolutely be there. Do share your screen, I would be curious what version is this that doesn't have that. **Rohit**: [10:42] I'm just doing this on the computer whereas I'm on my iPad right now, so I'm on a separate screen. **Anand**: [10:49] Ah, okay, okay, fair enough. Yeah, possibly. Let me share my screen again. And there was a question from Rajen: "You pasted the verification skill to your Google Form, can I paste it here?" Good point, I thought I had done that but I could be wrong. Oh, it is pasted, Rajen, just a few chats above. I'll paste it again. **Rohit**: [11:15] Also, Anand, are skills available only on Claude desktop or they are available on the normal Claude as well? **Anand**: [11:24] On all formats of Claude. Mobile, desktop... Claude code has skills, Claude Cowork has skills. Practically anything Claude has skills. **Rohit**: [11:36] Anand, where is the public link to your verification skill? **Anand**: [11:41] The one that I just pasted now? Has it not come through? Or is it not... okay, yeah, yeah, it is there. **Anand**: [11:51] Okay. And so, the thing inside this Markdown box is the verification skill. Now, I have absolutely no intention of using this unmodified. **Skills, like the instructions, impact a large number of chats. That is a high productivity gain and loss. So I intend going through every single word here** and making sure ... --- **Anand**: [00:00] ...and making sure that this is something that I agreed, disagree, use chats. And I will spend three, four hours on a skill before I let it get in. And I will also spend a few hours a week—maybe just after a few conversations where I use it, spend 10 minutes checking—and then saying, "I don't quite agree with this. I don't quite find that this is useful," and telling it to correct my skill. [00:29] **Skills, in other words, are your assets.** They are the things that you are building up over time based on your experience—the stuff that can't be replicated. And when I say "you," I mean as an individual, as a team, as an organization, as a family, however you want to structure the group. But they are assets. **These are assets that will probably last a reasonable amount of time.** These days, a reasonable amount of time is probably a few years. And yeah, they'll probably be worth building because the productivity improvement that they will give over a few years is probably worth it. [01:13] Sriram comments that he can't spot the customized option on the iPad. Yeah, maybe it's available only on the desktop or maybe it's somewhere else, but once set up, you will be able to use it across... or it automatically applies itself across all devices. **Participant 1**: [01:28] So, Anand, would it... would it run the skills in every single time you run Claude, by default? **Anand**: [01:34] Kind of. What it does is, the skills have descriptions. It takes the name of the skill and the description of the skill and adds it to every chat, saying, "Feel free to use any of these as relevant." Therefore, the description can be compact; the actual skill can be long. If it feels it is relevant, it will apply it—it will load the full skill and then apply it. [01:57] So let's take an example where I might have used a skill recently—some blog post, perhaps. Okay, this is a good one. So, I was on a session where we were using Google Meet, and it has only 250 people. So, I logged off and logged back in because my screen sharing wasn't working. I ended up not being able to get into my own session, which was a fun thing. And this is literally what I dictated, and it wrote a certain post. [02:35] But then I said, "Look, this is not quite in my writing style, and therefore, I want you to write in my style." Unfortunately, that may be somewhere I can't find. Okay, let's do one thing. We will just create a new chat where I'll say, "Write an article about some esoterically funny topic in my style, and verify." Okay. Brainstorm... **Participant 2**: [03:19] Anand, I think the screen sharing is gone again. **Participant 3**: [03:21] No, no, we can see it. **Participant 4**: [03:22] We can see it, yeah. **Participant 2**: [03:24] Oh, okay. My bad. **Anand**: [03:25] Thanks. But thanks for flagging. And I'm going to change this back to just regular high effort, and submit. Now, this, if my guess is right, should trigger three skills. I have an ideation skill, which the word "brainstorming" might trigger; "my style," which... yeah, so it's reading the ideation protocol skill for the brainstorming step. And I'll show you this because that is the next topic. And then it should read my writing style, and it should read the verification skill that we just saw. [Claude chat: Esoteric humor article brainstorm](https://claude.ai/share/c557d691-ef8b-4fd8-9774-d3f8ac131fe5) [03:59] So, let me click on this. It's reading the ideation protocol skill. When I click on that, it shows the ideation protocol that I'm using, which roughly says, **"Think in different ways, come up with wide ideas, and then pick the best amongst those."** Took a few iterations to get it to this form, but this is my single most used skill. I will share a link to this on the chat. Ideation protocol, and I'll also show you how it might work. [Ideation protocol SKILL.md](https://github.com/sanand0/scripts/blob/d9fde4f5f169c1d465603e0cd6095db3fcdabad3/agents/ideation-protocol/SKILL.md) [Anand writing style SKILL.md](https://github.com/sanand0/scripts/blob/d9fde4f5f169c1d465603e0cd6095db3fcdabad3/agents/anand-writing-style/SKILL.md) [04:38] This worked, and then it, secondly, also read my writing style skill. This is public—it's very similar to the URL that you have there, but might not be of use to you because this is my writing style. And **this is the one that I keep editing every week, saying, "No, no, no, this part—you must avoid LLM smells."** Every time I read an output and it looks like an LLM, I take that snippet and give it an example, saying, "Look, don't say 'X is the Y of Z,' 'not just X, but Y,' 'honest or genuine X,'" the word "load-bearing," the word "landing." A catalog that—at least a reasonably compact catalog that will not make it sound too much like an LLM and hopefully a little more like me. In any case, I edit manually, but this saves me the time and effort. [05:31] The third skill that it automatically read is the verification skill that we just saw. So clearly, in this particular case, it's picking it up automatically, and it usually does, but in other cases, you can always copy-paste from a skill. Let's see what it's... okay, it's still thinking and will come up with something. When it does, I'll share. [05:52] But okay, we have some public verification skills that others have shared. I'd love to take at least one of those and put it on the screen out here. Right, it's created the SKILL.md, which... okay, I can see... no, I can't see, okay, which is a pity. But Arvind's does a five-check audit run, scaled by stakes, tune to your standards, and yeah, that is useful because this is how you seem to be verifying it, which is nice. Let's take another one that was shared. Bikram's. [06:49] And that has already read some of the skills that exist. Skill writing. Okay, this is fine. This is a regular skill creator skill. Okay, and it asked you a couple of questions, which is great: "What should it focus on, and how heavy should it be?" Which is great, and it created a _numeric_ sanity check skill. Okay, which is niche and pretty focused, I guess because this is your strength, yes. [07:21] And yeah, okay, the script is not exposed, but good to get a sense of how it came about creating that skill. What's interesting is, Vikas, you created a skill with Perplexity, and Perplexity by itself does not use skills, but certainly can create skills. Oh, it can use skills? I did not know this. Learning something. [07:43] And this is an interesting one. It says, "When to use the skill: to cross-check financial figures." And there's a standalone and an embedded mode, okay. "You identify what kind of input, you get all the data points out, and you do a series of consistency checks, cross-period checks, unit scale checks, source traceability." Yeah, this is good, right? **This, if we could get it to do this kind of a numeric cross-check every single time, can be pretty powerful.** [08:22] And yeah, mark what is red, what is amber, with examples. Yes, this... let me paste this skill in the chat. And that's, yeah, a nice one, Vikas. Something that now, now that we have this skill, it's something that we can go through and say, "Hmm, have I seen this kind of an issue before? Should I ask my agent to automatically cross-check in this particular way?" And so on. [08:58] Question from Rohit: "Can you remind how skills can be created in ChatGPT?" Just paste the prompt or tell it to create a skill, and it will create a skill, Rohit. It understands what a skill is. "How do you use a skill?" is what is different in ChatGPT, and the answer is store the skill somewhere and copy-paste from it. Unless you have the ultra-expensive ChatGPT version, which might automatically let you paste the skill and apply it automatically. [09:31] Okay. The last thing that I wanted to cover was... let's go... oh, wait, before that. **Participant (Female)**: [09:41] Anand, when you said that, are you suggesting for ChatGPT we just store these skills as prompts? **Anand**: [09:49] In ChatGPT, you could store these skills wherever you store your prompts. **Participant (Female)**: [09:56] Okay. **Anand**: [09:58] ChatGPT by itself, that I know of, does not have a place where you can store prompts. **Participant (Female)**: [10:02] Oh, prompts? No, you need a separate file. Okay. All right. **Anand**: [10:06] I just store it in Notepad or whatever is the closest equivalent. **Participant (Female)**: [10:11] Okay. **Anand**: [10:14] Okay, here's some story that it wrote—an esoteric story, apparently in my style, which says, "Tamil used to cost 18 times more than English when talking to a chatbot. And today, it's 1.7x." Okay, that actually is not a bad story. **The cost of non-English languages from a token cost perspective used to be pretty high.** And that was a big theme, saying, "Oh, there is a lot of preference towards English" and so on. [10:51] And okay, and it did the verification: "Is it esoteric? Is it funny?" Well, let me put it this way: even without me going through the verification, it makes me a little more confident that it's done a verification pass, and I will feel a little more comfortable about this story. I might actually publish this. Why not? It's genuine news of the kind that my audience is interested in. [11:22] The last topic that I wanted to cover was ideation or brainstorming. One of the things that we believe... oh, actually, sorry, there are two things that I want to cover. I'll come to the last one, which is schedules, and I'll try and cover this quickly. [11:47] **Brainstorming is something that LLMs are great at because they hallucinate. Hallucination is creativity, creativity is helpful for brainstorming, and so on.** But because they've been trained on the average of the internet, they tend to go in a certain direction. We sometimes need to steer them in different ways. I have two approaches, both of which have proven reasonably effective as a workflow. [Ideation Protocol](https://github.com/sanand0/scripts/blob/d9fde4f5f169c1d465603e0cd6095db3fcdabad3/agents/ideation-protocol/SKILL.md) [12:13] And let's see, did I paste this ideation protocol? One of them is the ideation method that I've added as a skill. Here's what it looks like. It says, "Ideate when you are trying to brainstorm or do something creative. Two steps: one..." --- **Anand**: [00:00] One, first, find new stuff—don't worry about whether it's possible or not—and here are the ways in which you might do it. Find three people who will see this differently and guess what they might think about it. Find two completely different domains and take three ideas from there. Find the five most obvious ideas and eliminate, completely ban, those ideas and similar ideas. And then, if you have any duplicates, put them all together, cluster them, etc., and now you have a catalog of interesting ideas. And then, critique each one of these. Score them: is this going to be high impact, is this really new, how fast can somebody execute it? And then recommend the best ideas in whatever shape or form. Feel free to tweak it. But what I've found is that consistently, so far, when I've run this—either because the prompt is helpful or the models themselves have gotten so good that even without this skill they would have gotten the job done—I'm getting good ideas out of brainstorming exercises and therefore I use them in workshops quite liberally. Put in a question "How do I solve this kind of a problem?", it usually comes up with one thing that I can share well. **Anand**: [01:21] Another approach to brainstorming or ideation that I've been using is an "Ideator". The premise is that you can take two ideas and combine them to get an interesting new idea. Now, let's say we want to create a new web application or a process automation or literally anything else. My goal is to create an experiment for a workshop. And I want to combine two threads. What two threads? I'm not fussed. I've taken a bunch of sources—all of my learnings which I've jotted down, or creative ideas, or a bunch of stuff—and I can randomly pick from these, maybe filter the slightly more recent ones, whatever, and yeah, not even worry about it. [Ideator Tool](https://tools.s-anand.net/ideator/) **Anand**: [02:18] So on the left side it says, "Comic-Con has toys, games, figurines." On the right side it's "Amara's Law: we tend to overestimate the effect of a technology in the short run and underestimate in the long run." Can we combine these completely unrelated ideas and come up with an experiment for a workshop? The buttons here will copy these as a prompt into whatever—let me pick ChatGPT, no particular reason, any of these would work just fine. And the prompt that it's been given is "You are a radical concept synthesizer, heart of a storm, even expert"—I'm not sure how much of that really helps—but the idea is **generate a big, useful, non-obvious idea for a given concept.** And the concept is basically an experiment for a workshop. That's the goal. And the concepts are something to do with Comic-Con and something to do with Amara's Law. [ChatGPT chat: FutureCon Workshop Idea](https://chatgpt.com/share/6a37598b-3920-83ee-92bf-4f9295f2f5e3) **Anand**: [03:26] And it has a rough version of the brainstorming ideation protocol that I had shared. It comes up with a series of ideas and eventually, what is the idea? What is the experiment that I could run in the workshop? Oh, okay, this was the very first one. "Run the workshop as Comic-Con from 10 years in the future." So instead of asking people to predict the future of AI in technology, make them build toys from a future where the technology has become boring, normal, and everywhere. My daughter studies at SUTD. This is exactly the kind of project she brings home. Actually, this is in fact almost the same as a project that she brought home: "Create a future world where blah, blah, blah." So all of that IB-style education seems to have also gone into it. But this is just a good, effective way... This is a good exercise. And for an audience where we're trying to evangelize AI and tell them, "Look, why don't you use more AI?", this is a simple, fun way of creating an experiment, an exercise where you say AI has become boring technology. But my aim here is not to tell you that this is a useful exercise. **My aim here is that you can create out of completely random ideas something that will align to any goal that you have, and that makes the power of brainstorming quite powerful.** **Anand**: [05:01] Comment from Madhu, "Could you please share how you manage your notes, reading, tasks, etc., a second brain essentially?" A very short answer to that, Madhu, is I keep them as notes, Markdown files. Obsidian, Evernote, OneNote, whatever—anything that you can type in plain text and you can share across mobile and laptop would work for me. In my case, it is literally files like this. My Obsidian or Evernote is Visual Studio Code. In this case I had made a list of about 100 tools that I wanted to show you, out of which I've probably covered 20. Here is the responses, my notes from other areas—all of that stays in Visual Studio Code. And all of these are in a... or are saved on Dropbox, synchronized with my phone as Markdown files under a "Notes" directory. I have several such directories: one for my talks, one for my readings, one for my tasks, etc. **Anand**: [06:08] Bharat has a different question: "Amazon Quick has a version for non-enterprise use. How does it compare to Claude or ChatGPT?" No idea, Bharat. I'm hearing about Amazon Quick for the first time and I will check it out, I would certainly love to find out how this works. I'll tell you. **Anand**: [06:33] The last thing that I wanted to cover is schedules. Does ChatGPT have schedules? Does Claude have schedules? Let's try something. Schedules are a way of running something on a schedule, nothing complicated about that. ChatGPT definitely has schedules, and you'll see that I don't have any schedules out here. I have other mechanisms where I run schedules. But what I can do is ask it to—on a daily, weekly, monthly, any basis—automatically run a prompt and get me the output. This can have various uses. For example, I could say, "**Go through all the new regulations that have appeared related to financial services in Singapore and do this on a weekly basis. Research exhaustively but summarize specifically what this means to my portfolio.**" Now, the "my portfolio" part assumes that you have shared your portfolio with it in some past chats, and it will do it. **Anand**: [07:53] Or let's take a different example. Every week I want you to go through the financial markets and tell me how exactly these markets are likely to move over the course of the next week and why. In other words, I want you to make concrete, actionable predictions with evidence and reason that I can evaluate. Put this in a concise form that I can review within two minutes. Another example that might work. Let's take a third thing that... let me show you what this might look like if I were to run it. [ChatGPT chat: Weekly Market Outlook](https://chatgpt.com/share/6a3759e1-2718-83ee-9b3f-7add3f0bbbc3) **Anand**: [08:31] Now, you notice that I went to "Scheduled" and typed this, but it doesn't matter. You can just go anywhere on ChatGPT and tell it, "Every certain frequency, I want you to do a certain task." And it will create a scheduled task. In this particular case, maybe all it will do is say, "Okay, I have added a scheduled task." It is, as you can see, creating the task... and yeah, set up a weekly blah, blah, blah. And this task I can look at—this is what it's decided to do. I can edit it if I wanted to, I can pause it, I can change all kinds of things. Now, **this is powerful because we don't have to remember to use AI. It just applies its knowledge, it has our context—whatever we provided—it has the access to external tools, and that means we can do all kinds of interesting automations with it.** **Anand**: [09:43] One of the automations that I've been exploring is "Who can I...?" So here is my thesis: **AI is going to make humanity more important in some sense. Meaning, we will value human relationships even more.** AI is going to make a lot of technology go behind a layer. And if, therefore, in the AI era human relationships will have higher value, I better start being nice to people. And since I have zero practice in that, I may as well use AI's help to do that. So one of the prompts that I schedule—like I said, I run my schedules locally, but it is this: "**Who in my professional life right now deserves an unreasonable gesture and what would that be?**" And I've told it, "You can go through my transcripts, you can go through my emails, you can go through my WhatsApp conversations, look at content in the last two weeks, and just tell me who are the five people that I should make an unreasonable gesture to." And I'm trying to, over the last month or month and a half, practice this on a regular basis and it works. Good thing is I'm not expecting anything... obviously I'm expecting anything [word?] relationship capital, but at the moment I'm not particularly expecting anything and it's harmless. But this is exactly the kind of thing that I will forget to do. And **a smart assistant that reminds me to do something, whether it's personality development, whether it's an investment or whatever, or even just give me the latest news—whether it's regulations, is there some new due diligence risk that might come up to my business—is the sort of thing that we can create.** [Unreasonable gesture prompt](https://www.s-anand.net/blog/prompts/unreasonable-gesture/) **Anand**: [11:29] So the last exercise that I will request of you is of course, please run this—and we have an input from Vijay that Claude also allows you to run tasks regularly. Let's think about this for a minute: what would be a useful schedule for you to run? Something that if you get on a daily, weekly, monthly, whatever basis, will help you in some shape or form? **Anand**: [11:59] While you answer that, I'll take the question from Arvind, which is "Is there any way for LLM portability? I'm using Claude currently, all my memory is there, I want to move to Gemini, can I do that?" Yes and no. No in the sense that you can't export the chats themselves... sorry, you can't _import_ the chats themselves. You can export them—I'll show you how to export. But what you can do is, as we've seen in the past, just tell Claude, "**This is my situation, how would I solve it?**" And many AI agents have instructions on how to do this. Let me put it across. "A friend asked whether there is a way to move all the conversations and memory that Claude has into Gemini. Is there?" And you notice that unlike ChatGPT, the transcription here is not good. Let it run. Most likely it will say, "Have Claude summarize everything that it knows about you, put that into Gemini." And yeah, there isn't a one-click way. **Anand**: [13:19] But there is an "Export Data" option. Actually, yeah, this is an important one. You can go click on "Settings"—ChatGPT has the same—and somewhere in "Privacy", you can click on "Export Data". That will let you export the last 30, 90, custom, or all conversations. And this will send you an email which you can download. It's a large file. You will be able to read it, process it, and in theory, you could give this to Gemini but it'd be too huge. In practice, yeah, actually you can't do much with it until you learn a little bit of coding—white coding—which we'll cover in the fourth workshop. So, not yet. **Anand**: [14:06] Question from Rajen: "Can we do a task, a scheduled task, in Claude?" Vijay said we can, so let's find out how can I do this. Can you run scheduled tasks automatically? The news on this is... I just forgot. Okay, and Vijay is saying "Yes it can." And yeah, I would love to see where it appears. Oh, it's under "CoWork". Got it. CoWork is the desktop application and you can download it by clicking on this download button for Claude desktop for macOS or Windows—if you're running Linux like I am, you won't get it, but you probably aren't—and when you download the Mac or Windows desktop app, there'll be three tabs: one is Chat, one is Code, and the middle one is CoWork. Apparently that lets you run regular tasks. [Claude chat: Weekly Market Outlook](https://claude.ai/share/80adbad7-1128-4841-bad0-01b021bea8d3) **Mukul**: [15:19] Is it possible to do daily trading on Claude in the stock market? **Anand**: [15:23] I am so tempted right now based on your comment, Mukul, to actually do a trade. I don't have a trading account. If you have a trading account and you have a Claude account—others can drop off, but you can share your screen and I can tell you how you can make a trade. I'll leave it to you. **Mukul**: [15:46] Sure. I will connect with you later. **Anand**: [15:51] Cool, yeah, fine. Or if anyone wants to try it right after this session, that's cool too. But since we are talking about that, let me show you how you would go about doing that. Let's say I am on LinkedIn and I'm visiting... I've recently gone to, let's say London. And I would like to know who are the interesting people that I'm one-hop away from, or two-hops but with lots of mutual connects, that I can connect with? So, I will dictate to ChatGPT based in Claude. "**I'll probably be in London in August and I want to find out who are people that are direct connections of me that are in LinkedIn or one-hop away but tightly connected with multiple mutual connects that I can probably easily reach out to them. Give me a list and you know me, so tell me the people that I would be most interested in connecting with. Scan LinkedIn as you wish to get that information.**" **Anand**: [16:59] Now, Claude and ChatGPT have a plugin for browsers, which you can activate by installing the plugin and clicking on this little button—in my case it's Claude. You can choose the model—this is not a very complicated task, I'm just going to choose Sonnet 3.5 which is midway—and I'm going to tell it, yeah, it can do whatever it wants with my LinkedIn account, I couldn't care less, and paste it. And it looks like I can teach Claude some stuff as well by clicking on it. I've never tried this before, but I'm just going to submit. Usually, I just let it run. At the end of it, it will have the answer. I'm quite confident this will take at least 10 minutes, maybe 15 minutes. So, but right now, I'm hands-free with it going through my LinkedIn profile, and that orange cursor is it's so-on, and it's probably reading the screen to figure out where to click and all of that. It will get the job done. It might burn a few more tokens than I might have been happy with, but **if it can do this, there's nothing that stops it from making a trade, there's nothing that stops it from filling out my CAPTCHA, my OTP... if I can connect it to my phone and get it done. I've certainly filled expense reports using this entirely.** Saying, "Look, go through my email, take whatever you find, go through whatever websites, take what you find and update my expenses." **Anand**: [18:35] While we watch Claude do this, I'll just check if there are... okay, a bunch of 10 useful schedules. I'll wrap up with a quick read-out of the kinds of things people are finding useful on the schedules side. Okay: - A prompt to remind me of social appointments weekly. Good one. - Portfolio health checks related... yeah, checks. - Professional networking reminders. - News tasks reminders. - Next things on the market to suggest and focus on. - Weekly reviews of portfolio. - Being prepared for meetings. This is something that I do on actually a daily basis: **every day I say, "Go through my calendar, tell me how I should prepare for each of these meetings," and yeah, including points from previous meetings.** - A daily morning market brief. - Learning a new skill—okay, interesting—keep me on track. - And to go through my calendar to find patterns. Okay, that would be a useful prompt if your patterns change on a weekly basis, this would be a good scheduled task. If your patterns may not change that often, you might want to run this maybe annually on a schedule as well. **Anand**: [19:59] But yeah, give this a shot. I'll just wrap up by sharing what we covered today and what you might want to explore. We started with how the talk summary slides... let me stop sharing the screen. We started by looking at how we can use Google AI Studio at no cost to transcribe and create sketch notes or comics or whatever using ChatGPT, compress them, and generate a story out of it. We looked at verification workflows, how we can validate experts who are smarter than us, how we can make that verification automated through skills, and how LLM-as-a-judge (multiple checks) can reduce the errors. You built skills, and skills are a way of automatically applying the knowledge that we have—yet another tool in our arsenal. We looked at tools for brainstorming, how we can use a combination of different ideas and synthesize them into a vastly new set of ideas or just have the LLM create a wide set of ideas and come up with new stuff. And we ended with schedules as a workflow where again, we don't need to trigger most of what we're doing. **Anand**: [21:30] **The big part of the theme here is: the work that we did somewhat manually last time would have led to a certain amount of extra work for us. We will have to verify it, we'll have to do it regularly, we will have to copy-paste the prompts, we'll have to convert it into new formats. This session was largely around how do we alleviate that extra burden using AI as the theme.** In the next session, we're going to take a completely different tangent: analysis. How do we analyze data? How do we analyze content? What can that lead us to? And it's a very different dimension altogether. So let's take it up then. Thank you everyone for joining in. If anyone has questions, I'm happy to take that up, but otherwise we are good. Session stopped. Thank you. **Sandeep**: [22:18] Thank you, Anand. You've been two hours and 10 minutes. And thank you, and you know, the summaries will be provided to everyone with the recording, and Anand will cook up a nice story with the cartoons as always. And the next session is on the 25th of July, so we'll get in touch with all the details a few days before the event. Thanks once again, Anand. Thank you so much. This has been absolute... **Attendee**: [22:45] Thank you. **Attendee**: [22:45] Thanks a lot. **Attendee**: [22:46] Thank you.