The Workshop That
Made Itself

The page you are reading was not, strictly speaking, written by a human. Nor was the one from last month's workshop — the recap with the video, the audio, the comic, the story and the neat list of takeaways. That summary took about an hour of elapsed time to produce. It did not take Anand the hour; it took Claude the hour, with Anand stepping in to correct it. And the obvious question — the one every person in this workshop was secretly asking — is: how?

It was Saturday, 20 June 2026. About fifty professionals — bankers, fund managers, consultants, founders — had given up another weekend afternoon for the second of four AI Unboxed workshops run by the IIM Alumni Association Singapore. The first session had been about context — how to get more out of the model's intelligence. This one was about tools and workflows — how to get more out of the model's capabilities. As Anand put it, the title really should have been the other way around: workflows first, because the workflow is what leads you to the tool.

The host, Anand, is Head of Innovation at Straive and former co-founder of Gramener. His colleagues call him an "LLM Psychologist" — someone who studies not just what AI does but how it thinks, where it lies, and how to build a working relationship with it. His IIM Bangalore yearbook, co-organiser Sandeep reminded everyone, called him "God of all things." Sandeep also reported, deadpan, that after last month's session several alumni had started taking cycle rides from police stations to get inspired, "hopefully not related to any forced visits."

The Most Memorable Anniversary

Anand opened, as he likes to, not with theory but with a story about himself. His best use of AI that month had nothing to do with work. On 9 June — his 24th wedding anniversary — at around 10:30 at night, far too late to buy anything, he went to ChatGPT and asked for ideas. He gave it context: "I uploaded my bank expenses last time. That'll give you a sense of the things we buy. You know us a little bit. Do what you want."

One idea stood out. ChatGPT suggested he tell his wife: "Tomorrow is our 24th anniversary. For the next 24 hours, I'm going to knock off 24 things from your to-do list." She is, Anand confessed, a to-do-list freak. At the end of the day she said it was her most memorable anniversary.

"It is not a powerful use of AI that makes a difference. What makes a difference is the impact. Sometimes even simple uses have really powerful impacts."

— Anand, on his 24-things anniversary

This was not an aside. It was the thesis. The workshops, he has said before, add value less through content than through the simple fact that you have decided to spend time on AI — and that more practice, more habit, leads to exactly these small-but-mighty moments.

So he asked the room to do what they'd done last time: open a live survey and share their best use of AI that month. The answers scrolled in, and they were a portrait of this particular audience.

The cluster was unmistakable. "Researching new topics and stocks." "Deep dive into company fundamentals." "Used Claude to build a bond portfolio." "Analysis of a complex convertible bond offering by an Australian company." One person had "consolidated my financials from different countries, in different formats and places, into a single Excel sheet." Another had "used Claude to review a redevelopment project — JDA, POA, other documents — to see if there was anything untoward."

But the funniest and most telling answers were the ordinary-life ones. A Gmail cleanup. "Found the one email that contained a critical bank ID." "I asked AI to help create a CV for my daughter using her LinkedIn profile." A medical query with sub-topics. And the one Anand clearly loved:

"Created a promo video and every business proposal with it. Some started in the cab ride to the client and were done by the time the elevator door opened."

— a workshop participant, on the new speed of work

The point of collecting these wasn't a popularity contest. "A lot of what we learn is really from how others are using AI," Anand said. So when Sonal asked how to do a Gmail cleanup, and Rajen wanted to meet whoever built the bond portfolio, the room had quietly become a peer-learning network. Then Anand pivoted to the question that had been hanging in the air since the first slide: how, exactly, was last month's beautiful summary made?

Step One: Turn Talk Into Text (For Free)

The answer began somewhere most people never look: Google AI Studio. Not Gemini the chatbot — the studio, a stripped-down playground for single-shot tasks. Anand's pitch for it was blunt and practical: it's powerful partly because it's free, and partly because some of Google's models are genuinely best-in-class.

"I haven't seen many models — certainly not at that cost — do transcription as well as the Google models can."

— Anand, on why he reaches for AI Studio

The workflow that produced last month's summary started by extracting the audio from the recording ("How do you extract audio from video? Ask Gemini; it'll tell you the tool for your operating system"), then uploading that audio into AI Studio. He could have typed a single word — "Transcribe" — and it would have worked. Instead he uses a refined, battle-tested prompt he calls Transcribe Talk.

Every line of that prompt is a scar from a past mistake. "DO NOT MISS ANY PART" exists because, six to eight months ago, models would silently drop chunks of long conversations. The "drop verbal tics" line exists because nobody wants their ums immortalised. The translation line exists because Anand routinely slips into Hindi or Tamil. The prompt is a memory of everything that went wrong, encoded so it never goes wrong again. A sibling prompt, Transcribe Call Recording, even guesses the speakers' names — and, remarkably, just telling it to guess works.

Then Vikas noticed the token counter, and Anand seized the teachable moment. AI Studio shows that the two-hour transcription would have cost about 14 cents on the paid API — and that's with an expensive model; older ones cost a tenth of that. But you needn't pay at all:

"Google's deal 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."

— Anand, on the AI Studio bargain

When Bharat asked whether AI Studio handled languages beyond English, Anand answered live — by switching to Hindi mid-sentence and asking it to transcribe in colloquial English, "throw in a 'bro.'" The model obliged, and on a second pass did something even better: it gave the Hindi, a transliteration, and a translation. "Best of all worlds," he shrugged. (When Sumit slipped in some Swedish, Google guessed Danish — but translated it correctly. Close enough.)

One attendee had quietly been running AI Studio's live transcript in the background the whole time, capturing the workshop itself. That, Anand said, is the real unlock — and it loops straight back to the bigger point: the documents aren't the only thing you can feed an AI. Voice is now an input too.

The Knob You Don't Need To Touch

Vijay spotted a slider on the right of the screen — temperature — and asked what it did. It is one of those settings that AI obsessives once fussed over endlessly, so Anand ran a live experiment. He asked several copies of the model to "complete this statement with 5 options: The world's best…" at temperature zero, then again at temperature two. The outputs were variations on a theme — "the world's best food city is Lima, Peru"; "the world's best airport is Singapore Changi"; "the world's best country is Switzerland" — and the honest result was that he couldn't reliably tell the low- and high-temperature runs apart. "To be fair, I should have picked a better example."

But the underlying idea is worth keeping. When a model writes, it picks the next word from a distribution of candidates — maybe 70% chance one way, 20% another, 5% a third.

"Temperature is how much of that long tail it picks. Temperature zero picks its first choice almost every time. Higher temperature picks more uniformly. So higher temperature means more creativity; lower means more repeatability."

— Anand, on what the dial actually does

And then the punchline, which is very Anand:

"The most important thing you need to know about temperature is that you don't need to know about temperature. The newer models have taken the setting away. The providers are saying: we've learned enough, leave this to us."

— Anand

The settings that do earn their keep are the unglamorous ones. Structured output forces the answer into a shape you define — say, an array of {speaker, timestamp, speech} — which is exactly what you want when turning an earnings call into a spreadsheet. ("How do you convert this JSON into Excel? The answer to any question these days is: ask AI.") Grounding with Google Search lets the model check claims against the live web — perfect for fact-checking an investor-relations transcript line by line. Code execution and function calling he deferred to the fourth session.

Step Two: Make It A Comic

With the transcript in hand, Anand did something that looks frivolous and turns out to be strategic: he pasted it into ChatGPT and asked for a full comic page. There's nothing sacred about comics — it's one of many ways to turn dense text into something glanceable. Anand keeps a whole catalog of illustration styles and picks one to suit each session.

Two other lines in the comic prompt did the heavy lifting: "think about the most important points and structure it as a memorable story." That instruction matters because image generation is now a tool call, one capability among many that a reasoning model can decide to use.

"AI these days works in an agentic way. It thinks, does something, thinks again, does something — and keeps going until the job is done. An agent is a human-level — or smarter-than-human — reasoner who orchestrates and calls the various tools."

— Anand, on the two axes of AI progress

You could watch it happen. For last month's comic, ChatGPT thought for a little over four minutes before drawing a single line, writing its own internal brief along the way:

This, Anand noted, is the through-line of the whole session. Two axes of AI development are racing in parallel: the agents are getting smarter, so they know what to do; the tools are getting more powerful, so they can do more. The real power is in the combinatorial explosion between them.

An aside that saves your IT team

When you get an image out of any AI, it is enormous. The comic was 2.8 MB. "If I send 2.8 MB to 1,000 people, that's 2.8 gigabytes sitting in inboxes. One a week, and by year-end that's 100 GB. Our IT teams will thank us if we compress." His tool of choice is Squoosh, and his one-line rule is worth taping to your monitor:

"With images, probably the only thing you need to know is: compress using AVIF on Squoosh."

— Anand. The comic on this page is an AVIF, compressed ~90%.

Mid-demo, Sandeep dropped a striking image into the chat that, Anand observed, didn't look like a comic at all. Sandeep had used NotebookLM: paste the raw transcript, press the infographic button, done. It came out clean — no spelling mistakes, which is itself a milestone. "Nano Banana 2 was good, but usually with this much text it would have made one error, and I can't instantly spot it." His current verdict on image quality: ChatGPT has the edge today; Gemini is much faster; the lead keeps toggling every few months.

Step Three: Skills — "Here's How I Do Stuff"

Transcript, then comic. The third step — turning it all into the story you're reading — was almost anticlimactic. Anand didn't write a long prompt. He invoked a Talk Story skill. This is the moment the session quietly turned its biggest corner, from tools to 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. It knows what the world does generally. This is how I do it. It indicates a preference."

— Anand, defining a skill

His Talk Story skill is a recipe written once and reused forever: find the talk directory, make sure there's a transcript, convert any PDF slides to images, check for context, screenshot the web pages that are linked, read everything, generate the HTML in chunks because Claude gets stuck otherwise — Anand had clearly been burned by that — and update the README at the end. Every one of his talk pages, including this one, is produced by running exactly that skill. Since last month he'd added four more talks, all from the same recipe.

And here's why a finance audience should care. Two days before the workshop, an education-services client visited Straive in Mumbai for a long, recorded discussion about where they might use AI. Afterwards, Anand ran the identical pipeline — audio → transcript → comic → story, hyperlinks and all — and sent it over.

"They were blown away by the richness and intelligence of what was captured. From a business-development perspective — if attention is a scarce commodity, then new formats like this are a strong attention mechanism."

— Anand, on the real ROI of the workflow

He sees a broader shift coming. At a Straive sales meetup in New Jersey the previous Thursday, the leadership made a striking call: no presentations — only HTML.

"HTML output, data stories like these, are easier for AI to create and for us to consume. We may as well shift. I expect we'll see a lot less PowerPoint going forward."

— Anand, calling the slow death of the slide deck

A detour worth keeping: Anand also played a catalog of Gemini text-to-speech voices reading Dutch, built for a Dutch client who'd never heard convincing local voices — one deep male voice nailed the accent so well they were thrilled. Which leads somewhere darker, too: "Find a friend's voice, scam them for money — that's pretty easy now." Anytime a familiar or official-sounding voice calls, he warned, be a little more careful. (Voice cloning on tools like ElevenLabs needs only a few seconds of audio.)

Why You Never Get The Same Answer Twice

When Sonal re-ran the very same comic prompt, she got a different — and rather lovely — result. Rohit asked why. Anand's answer went two levels deep, and the second level is genuinely surprising.

The first level is temperature: models are deliberately a little random because that's how human language sounds. But even at temperature zero, you still won't get perfect reproducibility — and the reason lives down in the silicon.

"Take 0.1, round it to zero. A hundred of those, rounded, add to zero. Un-rounded, they add to 10. Depending on how you group the rounding, you get different answers. A neural network is just a way of grouping numbers — so the same input can give different outputs."

— Anand, on floating-point non-determinism

It's a neat intuition for a real, recently-understood problem. Researchers at Thinking Machines Lab showed in 2025 ▸ that the true culprit usually isn't randomness at all but batch invariance — the same request gets numerically different results depending on how many other requests share the GPU at that instant. Tiny differences flip the model's next-word pick, and the outputs diverge. The deeper lesson for the workshop: don't expect determinism from AI. Expect to verify. Which is exactly where the session went next.

How Do You Verify Something Smarter Than You?

Anand reframed hallucinations in a way that defused the usual panic:

"Hallucinations are not very different from people lying or making mistakes. Sometimes we don't know enough about their field to judge. So: how do you verify humans? What's been most effective?"

— Anand, turning a tech problem into a management problem

He ran it as a survey, and the room's collective wisdom was excellent. The most common answer was to cross-check with another LLM — Sandeep's "ask Claude or GPT to cross-check," and Arvind's lovely variant: "ask the same LLM to check itself again — Claude most of the time says it missed something." Anand recognised the move from his own student days, when a classmate would simply ask "Are you sure?" and watch him reconsider. It reminded him of a famous diplomat's trick:

"He'd always ask, 'Is this your best work?' The person would go improve it and come back. He'd ask again. Only when they finally said 'yes, I genuinely can't do better' would he review it. Why review something that isn't someone's best work?"

— Anand, on the cost of premature reviewing

Other responses anchored on what people already knew: Rajen's "cross-check with what I knew," a dipstick on familiar data points. Pramod's wise note that "people tend not to lie; it's more that the information they're aware of is limited." And one gloriously human confession: "Waking up till 2:15 a.m., like yesterday, to chide AI and correct myself." Then Anand fed all the responses back into ChatGPT to cluster them and find the gaps — a move he was performing as he taught it:

"This is the kind of thing you can do in the middle of meetings. Transcribe what someone's saying, ask for a response, share it, and look ultra-smart. I just ask people to give me a few seconds of thinking — when in reality, it is thinking."

— Anand, with a wink

His genuinely useful corollary: if a team member says "give me a minute to think" and actually takes it, "they're probably using AI. Please promote them."

The clustering surfaced the techniques people under-use: asking for sources ("where did you get this? Give me the link, I'll check it"); track-record checking (Anand no longer checks ChatGPT images for spelling — he's failed to find an error in a dozen tries — but still checks Gemini's, where he finds one most days); expertise-boundary detection (does it actually know this field?); and the powerful counter-example search — "tell it to find errors in its own logic, and it will. Agents are very diligent."

Then he pushed the model one step further: rank every technique by impact and ease, and boil it down to a single sentence you can paste at the end of any prompt. The model's own impact/ease map put "ask another model" surprisingly low ("models often share the same blind spots") and crowned three winners instead — claim-by-claim verification, uncertainty marking, and targeted spot-checking. Distilled to one paste-able line:

This is LLM-as-a-judge, dressed up with sources, citations and confidence levels. And Anand had the numbers to prove it works at scale.

In a controlled experiment, Anand's team had different models classify customer messages ("I want to update the address") into buckets. A single cheap model erred 14% of the time. Get two to cross-check and only accept when they agree, and error drops to 3.7%. Five models in agreement: 0.7%. They disagreed — and so needed a human — only 28% of the time.

"I'm getting 99.3% accuracy with a 72% reduction in effort, for negligible cost. These are trivially cheap models. Double-, triple-, quadruple-checking is a great way to automate verification in production."

— Anand, on cross-checking economics

Which set up the most consequential question of the afternoon. Sandeep asked: "Can we create a Claude skill for this — one that runs after every workflow?"

Building An Asset, One Skill At A Time

Yes. And so they built one, live. Anthropic introduced Agent Skills ▸ as folders — a SKILL.md file with instructions, plus optional scripts — that an agent discovers and loads only when relevant. In Claude they live under Customize → Skills. ChatGPT and Gemini have no automatic equivalent unless you're on their most expensive tiers; there, Anand's advice is humble and practical: "store the skill wherever you store your prompts — even Notepad — and paste it in."

Anand typed Sandeep's question almost verbatim — "Create a Claude skill for verifying answers which runs after every workflow" — set Claude to maximum effort with thinking on, and let it draw on years of his own cross-checking habits stored in memory. Crucially, he wasn't going to use the result as-is:

"Skills, like instructions, impact a huge number of chats — that's a big productivity gain and loss. So I'll go through every single word. I'll spend three or four hours on a skill before I let it in, and ten minutes a week correcting it."

— Anand, on why skills deserve real time

"Skills are your assets. The things you build up over time, based on experience, that can't be replicated — as an individual, a team, an organisation, a family. They'll last a few years. They're worth building."

— Anand, the line that reframed the whole session

What Claude actually produced was sharper than a naïve "check your work" checklist — because, as it noted, naïve self-checking is known to backfire (models rationalise their original answer, or "correct" right answers to wrong ones). Its verification-gate skill is built on four hard rules worth stealing wholesale:

Participants shared their own creations — and you can read each one in full: Arvind's five-check audit "scaled by stakes", Bikram's focused numeric sanity-check, and — to Anand's surprise — Vikas's financial cross-check skill built in Perplexity, which doesn't run skills but happily writes them. "It can use skills? I did not know this. Learning something."

To show skills firing together, Anand typed a deliberately loaded prompt: "Write an article about some esoterically funny topic — brainstorm first — in my style, and verify." If his hunch was right, it would trigger three skills at once. It did. You can watch them light up in the shared chat: the Ideation Protocol (on "brainstorm"), the Anand Writing Style (on "my style"), and the freshly-built verification skill.

The writing-style skill is the one he edits most. Every week he reads an output, spots something that smells of a machine, and adds a banned phrase: "don't say 'X is the Y of Z,' or 'not just X, but Y,' or 'honest/genuine X.'" The words "load-bearing" and "landing." A growing catalog that makes the writing sound a little less like an LLM and a little more like Anand.

Two related questions came up. Madhu asked whether Anand updates his Claude instructions to shape responses. His answer was a small revelation: his custom instructions used to be long — and now they're essentially one line.

"I used to put in stuff to correct their deficiencies. The models kept improving, doing it anyway — so I stopped. Today there's only one thing left, and it's a stylistic disagreement: 'Answer directly in Markdown rather than artifacts, unless I ask.'"

— Anand, on shrinking his custom instructions

Rajen asked whether you even need a skill file, or whether Claude learns from repeated chats. Both, Anand said: agents quietly build memory from your past conversations, but a skill is something the agent scans the description of in every conversation and loads when relevant. Memory is passive; a skill is a deliberate, reusable asset. Want it to definitely apply? Put it in the prompt.

Brainstorming: Hallucination As A Feature

Here Anand flipped the morning's villain into a hero:

"Brainstorming is something LLMs are great at because they hallucinate. Hallucination is creativity, and creativity helps brainstorming. But they're trained on the average of the internet, so they drift toward the obvious. We need to steer them."

— Anand

His Ideation Protocol — his single most-used skill — steers in two moves. First, diverge: find three people who'd see this differently and guess their views; borrow three ideas from two unrelated domains; and, cleverly, list the five most obvious ideas and ban them. Then, converge: cluster, critique, and score each idea for impact, novelty and speed-to-execute before recommending the best.

His second technique is a little web app, the Ideator, built on the premise that any two ideas can be smashed together into a third. Live, it served up "Comic-Con has toys, games, figurines" on the left and Amara's Law on the right, and asked for a workshop experiment fusing them.

Schedules: So You Don't Have To Remember

The final workflow removes the last bit of human effort — remembering to ask. Both ChatGPT and Claude can run a prompt on a schedule. You needn't even visit a settings page; just tell it, anywhere, "every week, do this." Anand's examples were aimed squarely at the room:

Or a weekly market outlook. Here's the actual exchange — the entire setup was one sentence:

Notice ChatGPT turned a vague ask into a scheduled task with a falsifiability clause — exactly the verification discipline from earlier, applied automatically. Pop the ChatGPT chat  ·  the same idea on Claude.

The survey filled with the room's own wishes — and they read like a portrait of busy professionals trying to stay sharp:

• "A daily morning market briefing on US / India / ME markets" — one attendee had already built it on Claude CoWork, emailing himself a draft each morning.
• "Weekly review of my stock portfolio — gains, losses, red flags from last week's news, stocks to watch."
• "Being prepared for my meetings — a list for the day and week, points to discuss, notes from previous meetings, something about the people."
• "Go through my calendar and find patterns — are they aligning with my goals?"
• "A prompt to remind me to set up social appointments weekly :)"

That last one points at Anand's favourite schedule — and his most contrarian idea of the day:

"AI is going to make humanity more important. It pushes technology behind a layer, so human relationships gain value. And if relationships matter more, I'd better start being nice to people. Since I have zero practice, I may as well use AI's help."

— Anand, on the human dividend of AI

So one of his standing prompts is the unreasonable gesture: "Who in my professional life right now deserves an unreasonable gesture, and what would it be?" — fed by his transcripts, emails and WhatsApp from the last fortnight. He's been practising it for a month. "It's the kind of thing I will forget to do. A smart assistant that reminds me is exactly what we can build." Daily, he also runs: "Go through my calendar and tell me how to prepare for each meeting, including points from previous ones."

The Browser That Works While You Watch

Mukul asked, half-joking, whether Claude could trade stocks. Anand didn't have a trading account to risk live — but he showed the engine that could. He opened the Claude browser extension, dictated a request, and let it loose on his own LinkedIn:

The orange cursor began moving on its own, reading the screen, deciding where to click. It would take ten or fifteen minutes and burn some tokens, but it would get there hands-free. And then the line that made the room go quiet:

"If it can do this, there's nothing stopping it from making a trade, filling out my CAPTCHA, my OTP. I've certainly filled expense reports entirely this way — 'go through my email, take what you find, update my expenses.'"

— Anand, on where agentic browsing leads

A practical coda for Arvind, who wanted to move his memory from Claude to Gemini: there's no one-click migration, but Settings → Privacy → Export Data gives you everything, and you can always ask the AI itself how to make the move. The honest answer for now: until you learn a little "vibe coding" — coming in session four — there isn't much you can do with that giant export file yet.

The Theme Beneath The Tools

Anand closed by naming what had quietly tied two hours of demos together. Last month's manual work — the prompting, the context — created a debt: you still had to verify it, repeat it, copy-paste it, reformat it.

"This session was largely about how we alleviate that extra burden using AI. Transcription, comics, stories, verification, skills, brainstorming, schedules — every one removes a trigger we'd otherwise have to remember to pull."

— Anand, closing Session 2

Sandeep, signing off after two hours and ten minutes, promised the recording and — inevitably — "a nice story with the cartoons, as always." You're reading it. Which means the workshop, in the most literal sense, made itself. Session three turns to a different dimension entirely: how to analyse data and content, and where that can take you.