The Problem

My wife and I are parents of two school-going kids. And like most parents, we were drowning in school emails and WhatsApp messages— homework reminders, event notifications, registrations to be done, teacher updates — all buried, all mixed in with work email, all easy to miss. This led to lots of missed school deadlines and lots of heartburn!

As a product leader with 25+ years of experience and as a founder of a company (LivKidz), whose main aim is to simplify the lives of working parents, I knew that I had to build a solution to solve this pain point which was so real, for so many parents.

What I Built: LivKidz School Planner

LivKidz School Planner is a parent dashboard that connects to your Gmail, reads your school emails, and automatically organizes them into four categories — Child To-Dos, Parent To-Dos, School Calendar, and General Updates — per child. No manual tagging. No copy-pasting. Sign in with Google, tell the app which senders are your school’s, and it does the rest.

How did i build this?

To build this, I used Claude Code — Anthropic’s AI coding agent — as my entire engineering function. Two weeks. Roughly two hours a day. A production-grade app with authentication, a database, third-party API integrations, and AI at its core.

Tech stack: Next.js, Supabase, Gmail API, Google OAuth, Claude AI, Vercel

Overall, I was very pleased with the result. Spending few hours every day, late night after work, building with Claude Code - it really took me back to my engineering roots and I really enjoyed the process of building on my own :-)

Here are my learnings from the experience.

Learnings from Building “LivKidz School Planner” with AI

1. Claude Code > Code

Claude Code is not just your Developer. It’s Your Entire Team. Going in, I thought of Claude Code as a coding assistant. I was wrong.

Over two weeks, it played every role on the team and I encourage anyone using it, to leverage it that way too.

• Product Manager: When I described the dashboard feature, Claude Code asked clarifying questions that I hadn’t thought through — should marking an item done for Child A affect Child B? What if the same school email is relevant to both kids? These are PM questions. It was working as a PM in my team and helping me refine and drive clarity on the requirements.

• Architect: It made all the big structural decisions. It recommended the right stack — Next.js (full-stack React, no separate backend needed), Supabase (for database), NextAuth (OAuth out of the box), Vercel (one-click deployment). When I pushed it to improve latency, it consolidated three separate API calls into a single parallel endpoint for better performance.

• Engineer: It built everything — Gmail API integration, OAuth flow, email sync engine, Claude AI parsing pipeline, email dedupe logic. And it handled all project setup: installing dependencies, configuring libraries, setting up the database schema, connecting third-party services. End-to-end.

• Designer: It came up with a simple, modern UX for the dashboard including onboarding, dashboard setup and the final dashboard of course. It proposed things I never asked for — a user avatar with initials in the top-right corner, a loading bar across the top during data fetches, a “Gmail disconnected” reconnect banner for when OAuth tokens expire, fade-out animations when items are dismissed. It knows what good product design looks like.

• Project Manager: It helped me track overall progress throughout the build — what had been built, what was pending, what needed testing. I didn’t need a separate tracker. Claude Code was it.

• Thought Partner: When I described features loosely, it pushed back and asked for more clarification When I was going in circles on a logic problem, it reframed it.

I had no developer, no designer, no architect, no tech agency. I had Claude Code — and I was able to ship a product just using Claude Code.

Stop thinking of Claude Code as merely a developer. Think of it as your entire team.

2. You Still Need to Own the Product. That Part Doesn’t Change.

Here’s the flip side of Learning #1.

Claude Code is a brilliant team member. It is not the product manager. That’s you. And if you forget that, you’ll feel it.

Early in the build, I described the dashboard loosely — something like “show school items organized by type.” Claude Code made a reasonable assumption: a single list, sorted by category. Functional. Logical. Wrong. What I actually envisioned separate cards — one per category — with items organized per child. That’s a fundamentally different UX. The assumption wasn’t unreasonable. I just hadn’t been specific enough.

Some of the product decisions made while building LivKidz School Planner came from me, as the builder and not Claude Code.

• Four separate cards (Child To-Dos, Parent To-Dos, School Calendar, General Updates) — not one list. And what goes where

• Onboarding process and steps involved. Supporting multiple children, different schools, how to define email senders and guiding parents in this process, are other examples of decisions taken

• Landing page structure and key message to highlight. Separating “Request an Invite” from “Already Invited? Sign In” on the landing page — two different user states, two different CTAs

• Lookback window, auto refresh or not and many other decisions where I took the call as the PM.

Claude Code executed all of these brilliantly once I defined them. It could not have made these calls on its own. I had to.

AI lowers the cost of building to near zero, but it does not lower the cost of not knowing what to build. You should own that part.

3. Let AI Handle the Engineering Complexity — That’s Its Superpower

Although I am an engineer by background, but I haven’t coded in a while.

I didn’t know many technical details (e.g. how OAuth token refresh works).

It didn’t matter. Claude Code handled all of it — and the depth genuinely surprised me. Some specific examples:

• Email deduplication: An email_log table tracks processed Gmail message IDs — already-parsed emails are skipped on every subsequent sync. I didn’t ask for this. It just built it the right way.

• Per-child state isolation: The item_dismissals table uses user ID + item ID + child ID as a three-part unique key — so marking done for Child A never affects Child B.

• API performance: Three separate database calls on dashboard load, consolidated into one parallel endpoint. Faster, cleaner.

None of these were in my requirements. Claude Code brought them proactively (for the most part). That’s what I mean by letting it handle the engineering complexity.

4. Is MVP Still Relevant in the Age of AI? Short Answer: YES

Midway through the build, the core product was working — sync, dashboard, per-child state, the full flow.

And, then I started to add more features (e.g. notifications, shared view between parents, ..and more)

Why? Because I could, and no one was pushing back !

None of these were bad ideas. But, they were resulting in scope creep for the MVP and it would have resulted in a feature-rich, bloated product which the user may have found too complex to use. I made a deliberate decision to stop.

LivKidz School Planner does one thing: reads school emails and organizes them per child. Everything else belongs to a future version, post MVP validation with real users.

MVP isn’t about what you can build. It’s about what you should build first. That judgment becomes more valuable, not less, when building gets cheaper.

5. AI Makes Mistakes. Stay in the Loop to catch them.

AI is powerful, but it is not perfect. Here’s what I ran into - some examples.

Getting the four cards right took many iterations. Deciding what belongs in each category sounds simple. It isn’t. In one case, a homework reminder email was appearing in both Child To-Dos and School Calendar — two cards simultaneously. The root cause: the AI was producing multiple items from a single email at different priority levels.

The fix required building an explicit priority-based classification system, a decision tree e.g. deadline beats to-do beats event beats update, a hard rule that a single email can only produce one item at its highest applicable priority. It took several rounds of testing outputs, identifying wrong classifications, and refining the parsing prompt before it was consistently correct.

Context window truncation. Claude has a context window limit. When school emails were long — full newsletters, multi-section updates — the parsing would silently truncate content. Due dates buried at the bottom of a long email weren’t being extracted, because that text never made it into the prompt. It took multiple rounds of “why isn’t this showing up?” before we diagnosed it. I had to be involved and push Claude Code, to uncover this flaw in the design.

Email formats Some school systems like Veracross use a different email structure, which resulted in those items not appearing on the dashboard. I caught it during my testing, AI didn’t flag it. It required extensive debugging to figure out the root cause - that the parsing logic was built around a standard email format and simply didn’t handle Veracross’ structure. Once diagnosed, the fix was straightforward. But getting there wasn’t.

Claude Code does have limitations (some examples - my wishlist for Anthropic)

1. No image generation. I wanted custom illustrations for the landing page. Claude Code can’t generate images. I had to use other AI toolsfor this (Side note: the new chatGPT image generation is super awesome! give it a try!)

2. No screen sharing or pasting screenshots. During setup — finding an API key, locating a project URL, navigating a third-party settings panel — I had to describe in words what I was seeing on my screen for Claude to guide me. Being able to paste a screenshot, or share your screen with Claude, would dramatically reduce this kind of overhead.

Review AI output like you would any other feature! Stay in the loop. Dive Deep

6. Build in Small Steps. That’s Where the Magic Happens.

I’ve seen what happens with one-shot prompting. You describe the full vision, the AI generates something that looks right, and then it breaks in ways you can’t trace — because nothing was built incrementally.

With LivKidz School Planner, I built it feature by feature and it worked quite well.

1. Landing page and Google sign-in

2. Onboarding (Children details)

3. Dashboard / Email sender setup

4. Claude AI parsing engine and email categorization

5. Dashboard UI Including 4 cards

6. Dashboard Refresh

7. Profile page

8. Avoid Clutter - Mark item as done / dismiss an item

9. UX Enhancements (e.g. “Overdue / Due today / Due tomorrow” labels, New item badges and first-time sync polling)

10. Gmail reconnect banner (e.g. if the user changes their Gmail password)

Each step was reviewed before the next started. When something broke, the blast radius was small — I always knew exactly what had changed.

Build incrementally. Polish incrementally. The product gets better in layers — with AI just as much as without it.

7. Always Ask AI for a Plain English Explanation

Every time Claude Code proposed a significant change, I asked it to explain what it was doing in plain English — as if briefing a non-technical product manager.

This resulted in two benefits over time.

a) First, I stayed genuinely in the loop — I knew why each change existed and what it might affect elsewhere.

b) Second, it caught misalignments before they became bugs.

Example: When Claude Code explained the email parsing logic, I realized categorization wasn’t working as I’d intended. I fixed it afterwards.

There were moments where the explanation itself was the debugging — something felt off, I’d ask, and the act of articulating what had been built would surface the flaw.

Don’t treat AI outputs as a black box. Treat them as proposals to be reviewed and approved by you

8. It Felt Like Having a Reliable Partner — Not Just a Tool

This one is harder to quantify, but it might be the most important thing I want to leave you with.

Throughout the build, I never felt alone. When bugs appeared — and they did — I wasn’t staring at a stack trace with no idea where to start. Claude Code diagnosed, proposed a fix, and explained the reasoning. When a deployment failed, it walked me through it.

What this gave me was a sense of control throughout. With so much code written, I couldn’t go in and debug it myself. But, I always had the confidence that Claude Code would be able to troubleshoot it, and the technical side of things were taken care. It gave me the belief that I will be able to ship the product eventually and this kept me going.

It also brought back something I hadn’t felt in a while: the genuine joy of building. As someone who started in engineering before moving into product management, this process reconnected me with why I got into tech in the first place. Seeing something go from a rough idea to a LIVE product being used by real parents — that feeling doesn’t get old.

If you’ve been sitting on an idea, but waiting for the right team: now is the time to build

**Brief Update - Post Launch**

9. Shipping is where building actually begins

Few observations below post the launch

A few days after launch, I opened my own dashboard. It said “Synced just now” — the green banner, everything looking healthy.

But something felt off. Emails from the last few days weren’t showing up.

I dug deeper.

The sync was running on every page load, completing, showing success — but nothing was actually being fetched. Turns out, Google OAuth tokens issued during testing expire silently every 7 days and mine had expired.

1. I reconnected Gmail to fix it — but the reconnection itself silently failed, issuing a token without the right read permissions.

Also, it required 2 reconnects to get the dashboard to work again. Root Cause: When a Gmail token expired, clicking “Reconnect Gmail” appeared to work — but the dashboard asked to reconnect again. A two-step database operation where the second call silently did nothing if the first returned no data.

Two layers of quiet failure, zero indication anything was wrong

The fix: make failures loud. The app now immediately shows a “Reconnect Gmail” prompt the moment auth breaks, and collapsing two dependent database calls into one atomic operation

2. Few days after, a beta user reported their dashboard was hanging and showing nothing. Same app — but a completely different experience. Root cause: Their first sync had to process weeks of emails and was silently timing out before finishing.

The fix: the first-time sync window was tightened to stay within platform limits.

3. The third issue was caught from my daughter’s iPad. She uses the planner in Chrome signed in with my Gmail, but her Gmail app is signed into her own account. Clicking the email preview opened her inbox, not mine. The school email was invisible to her.

The fix — instead of linking out to Gmail, I built the email preview inside the planner, fetching content directly using stored credentials. Whatever Gmail app is on the device became irrelevant.

I fixed the issues. But the real lesson was harder: the app had no way to surface its own failures. It looked healthy while being completely broken. Build in a way that lets you see when things break — not just when they work.

Monitoring shouldn’t be an afterthought, it should be a priority

Final Thoughts

Building LivKidz School Planner was one of the most satisfying things I’ve done in a long time. Modern AI makes it genuinely possible for a non-coder to go from idea to production-grade product in weeks — but only if you bring the domain expertise, own the requirements, dive deep enough, and stay engaged throughout.

If you’re a PM, product leader, or founder sitting on a problem you’ve always wanted to solve — start. The barrier is lower than you think.

Anyways, these are my learnings. Your experience may vary — would love to hear yours in the comments!

Parents: Try it Yourself!

If you’re a parent exhausted by school email overload, here is how the planner works

• Connects to your Gmail securely — read-only, no email content stored on our servers

• You tell it which senders are your school or teacher

• It automatically extracts what matters: deadlines, to-dos, events, and updates — organized per child

• Done items can be marked off or dismissed — separately for each child

P.S. This post is the 4th post in the series “How I AI with LivKidz”. Through this series, my goal is to share my practical, hands-on learnings in building with AI as part of my own venture. Follow me on LinkedIn, if you would like to be updated about future posts.

At LivKidz, we often get the following question from parents.

“What activities and classes should my child do?”

The answer to this question is rarely simple. It depends on the child’s age, interests, and available classes nearby. For many parents, discovering the right activities becomes a time-consuming process of browsing websites, WhatsApp groups, and recommendations from other parents.

So, I turned to AI. Could we build a chatbot that asks parents a few simple questions and recommends relevant activities, while leveraging the highly curated and verified LivKidz activity database.

That experiment led to the creation of LivKidz Activity Buddy — an AI chatbot that recommends kids’ activities based on a child’s age and interests.

In this post, I’ll share practical lessons from building the chatbot, including:

• why data format matters more than you think

• how prompt engineering drastically improves results

• what broke repeatedly during training

• and how I finally got reliable recommendations

A quick word on LivKidz first and what our mission is.

The Goal of the chatbot

The goal was simple: Create a chatbot where a parent could say something like:

“My child is 8 years old and loves sports and creative activities.”

And the chatbot would recommend relevant options from LivKidz activities database.

To prototype this quickly, I experimented with two AI tools:

• Chatbase

• Chatling

Both tools allow you to build AI chatbots using your own data. After testing both, I found the experience with Chatbase to be smoother and more flexible, so the rest of this post focuses on that experience.

Lesson 1: Getting Started is easier than you think

Initial setup was farily fast. I just answered a few questions and provided the LivKidz website link, and it was able to build a basic chatbot v1.

After signing up on Chatbase, I was able to:

• Specify my website URL

• The type of agent I wanted (General/Customer Support/Sales).

• Provide my own training data (e.g. docs/text snippet/notion doc). Don’t worry, you can skip this step and come back to it.

One interesting feature is that Chatbase automatically:

• crawls your website and extracts content

• picks up your branding elements such as logos

Customization options were also quite good. You can configure:

• chatbot name and logo

• color scheme

• default greeting and questions upon chat initiation

• choose the underlying AI model

This meant a working v1 prototype could be created very quickly, which was professional looking and aligned with the LivKidz brand.

But that was the easy part. Read on for more.. :-)

Lesson 2: Data format matters as much as quality

“Garbage in, garbage out.” We all have heard this phrase and this applies here too. But what surprised me during this process was something slightly different:

The format of your data matters as much as the data quality itself.

Quality of your training data matters. Of course! But what was surprising was not just what data, but how it is presented to the model can influence your outcome.

What do I mean by this? Read on..

Attempt #1: Uploading a PDF Export of the Sheet

LivKidz activity information was available in Google Sheets. Since Chatbase couldn’t directly access Google Sheets, I exported the sheet as a PDF and uploaded it as training data.

This did not work well!

The AI struggled to understand the tabular structure with (70 columns), and the recommendations often mixed up fields like:

• activity name

• description

• age group

• booking links

Clearly the model was not parsing the table correctly.

Attempt #2: Extracting Data into a Document

Next, I wrote a small script (with help from AI 🙂) that:

• extracted specific columns (just the ones we needed) &

• added the info into a document (google doc)

This improved the output. However, another issue appeared!

Because the document was exported as a PDF, pagination sometimes split activity details across pages. For example:

• descriptions and links could be separated on different pages

• activity blocks were broken mid-way

This confused the model and reduced accuracy.

Attempt #3: Structured Text Blocks (Best Result)

The best results came from using plain text blocks with a clear structure. Pasting the data from the Google Doc into the Text Snippet and training the model on it, did the trick!

The key improvement was introducing a separator string such as

<END OF BLOCK>

This ensured the AI did not accidentally merge data from multiple activities.

Once the data was structured this way, recommendation quality improved significantly.

Activity block Data Structure

"ID": 3

"Location Link": https://www.google.com/maps/place/Pooja'z+Artisitik+Avenue/@28.3973162,77.0968071,17z/data=!3m1!4b1!4m6!3m5!1s0x390d233775e39ee3:0x58b7b7397a0c3e7b!8m2!3d28.3973115!4d77.099382!16s%2Fg%2F11l2k4qyqv?entry=ttu&g_ep=EgoyMDI0MTAwNS4wIKXMDSoASAFQAw%3D%3D

"Headline": Feel the Rhythm, Own the Stage: Poojaz Artistik Avenues!

"Short Description": Pooja’s Artistic Avenue is a music, dance, drama, art and fitness studio. We also offer services like wedding choreography, school and corporate events. The organization works with the aim to discover, train and promote upcoming artists in the field of Performing Arts. We also offer at home classes

Home classes available, Enquire for more details!

"Category Name": Dance, Music & Theater

"Sub Categories": Music, Dance, Kathak, Keyboard, Guitar, Hip Hop, Freestyle, Fitness, Vocal, Semi classical, Folk Dance, Singing

"Min Age": 4

"Max Age": 

"Facility Type": Indoor

"Deep Link":  https://prod.livkidz.com/Screens/details/3/Index

"Image Link": https://livkidz.wordpress.com/wp-content/uploads/2026/03/pooja-artistik.jpg

<END OF BLOCK>

Lesson 3: System Prompt - This is where the magic happens

Chatbase automatically generates a default system prompt, which defines how the chatbot should respond. This default prompt includes things like Context, Role, Constraints, Tone and Instructions.

But, you need to refine the default system prompt to get the output you want!

Be prepared to spend most of your time with the system prompt, as it will require multiple iterations

Not surprisingly, the more specific and structured the prompt, the better the output.

Here are a few specific things that helped significantly.

A) Explaining the Training Data Structure

I explicitly described:

• what fields exist in the data

• the order of the fields

• how activity blocks are separated

This helped the model interpret the data correctly.

B) Defining the Output Format

I wanted the chatbot responses to look clean and structured.

For example, instead of showing raw URLs like:

 https://prod.livkidz.com/Screens/details/2/Index

The chatbot was instructed to display them as:

Learn More/Book

Small formatting instructions like this made the responses much more user-friendly.

C) Reducing Hallucinations

To avoid the chatbot inventing activities, I added strict step-by-step instructions and directed it to use only training data for giving recommendations.

ONLY recommend activities present in the training data. Do not invent new activities.

This significantly improved accuracy.

D) Highlighting Critical Instructions

Another trick that helped was emphasizing important instructions using markers like Hash and ALL CAPS letters.

### CRITICAL RULE
Use ONLY the training data when recommending activities.

This helped the model prioritize key instructions.

Lesson 4: Real fun starts - Expect Multiple Iterations

Even after setting up data and prompts, the chatbot did not work perfectly right away.

Most of the effort went into iterating and testing repeatedly.

Some of the issues I encountered included:

Activity data merging

Without clear separators, the AI sometimes combined information from multiple activities. Having activity separators improved the output accuracy.

Image handling challenges

Initially, I tried using image links stored on Google Drive.

However, Chatbase could not reliably access those links.

The solution was to upload images to WordPress and use those URLs instead.

Even then, training sometimes failed when image URLs were included in the dataset.

Troubleshooting issues

One feature I found surprisingly useful was the Chatbase support chatbot, which appears on their interface.

It helped answer many configuration and troubleshooting questions during the process. It helps to get quick answers / troubleshooting help as you are going through these iterations

Lesson 5: Deployment is Easy, Evals not so much!

Once the chatbot was working well enough, deployment was straightforward.

Chatbase provides multiple deployment options, including:

• iframe embedding or floating chat button on your website

• standalone chatbot links

• integrations with other platforms like Slack, WhatsApp

I embedded the chatbot directly on the website and had it as a floating chat bubble for users to access.

LivKidz Chatbot in Action - Hello World!

However, deployment is not the final step.

The real challenge is evaluation and improvement.

Chatbase allows you to:

• review past chat logs

• see how users interacted with the bot

• see confidence scores

• revise incorrect answers, which are then stored as QnA

• there are some additional paid features too such as chat analytics, sentiment analysis and most discussed topics

One feature I wished existed was the ability to explain in general terms what went wrong with an answer, instead of having to manually rewrite the correct response. That would make systematic improvement easier.

Now, Evals is a entirely separate topic on its own. More on that later :-)

Final Thoughts

Building the first version of LivKidz Activity Buddy was a fascinating experience. Modern AI technology makes it happen. What took months, now takes days. What took x-functional teams, now can be done by 1 person!

But a few things became very clear during the process:

1. Data structure matters enormously
2. Prompt engineering is a core part of the product
3. Iteration is unavoidable. God lives in the Details!

For founders like me, tools like Chatbase make it possible to go from idea to working version very quickly. But getting the system to produce reliable and useful outputs still requires careful work.

Anyways, these are just some of my learnings. Your experience may vary. Feel free to share your learnings through comments below.

Try it yourself

You can check out the LIVE LivKidz Activity Buddy here . Give it a try, would love to hear your feedback!

P.S. This post is the 3rd post in the series “How I AI with LivKidz”. Through this series, my goal is to share my practical, hands-on learnings in building with AI as part of my own venture. Follow me on LinkedIn, if you would like to bet updated about future posts.

Question: What happens when you mix a College reunion yearbook, LinkedIn and AI?
Answer: A Fun Weekend AI Project!

Recently, I attended my 25th year reunion at Indian Institute of Technology, Delhi (which was amazing, BTW!). Like most reunions, it left me wanting to reconnect with many batchmates on LinkedIn.

We do have a Class of 2000 yearbook (physical copy), but it doesn’t have a LinkedIn profile link/QR code. There was no easy way to connect with batchmates, other than manually searching for them on LinkedIn, one at a time. Slow and tedious process!

Enter AI!

I vibe-coded a simple app that:

1. Scans a yearbook page
2. Uses OCR to extract text containing names of people and companies they worked at
3. Lets you search directly on LinkedIn with a single tap

Scan → Tap → and done!

I found this to be a pretty simple, but powerful use case of AI. Something that would earlier take days of coding came together in just a few hours (including iterations).

AI is cool 🙂!

hashtag#AI hashtag#AInewusecase hashtag#reunion hashtag#IITDelhi hashtag#Reconnecting hashtag#Classmates hashtag#POC

P.S. This post is the 1st post in the series “How I AI with LivKidz”. Through this series, my goal is to share my practical, hands-on learnings in building with AI as part of my own venture. Follow me on LinkedIn, if you would like to bet updated about future posts.

P.S. Here is the link to the app. Give it a spin and let me know what you think. It is configurable so you can use it for your own college yearbook). The app was built using Replit

I am the founder of LivKidz. At LivKidz, our mission is to help parents engage their children during after-school hours in constructive activities by making it easy for parents to find and book these activities. We are currently LIVE in the Indian market. One of the new offerings we are working on launch is “LivKidz at Home”, which as the name suggests, enables parents to book verified kids’ instructors for home classes.

Goal: I wanted to use AI tools to build the prototype for LivKidz at Home, so that I can use it for sharing with colleagues, getting customer feedback and handoff to the tech teams for development. I tried 4 AI tools to build the prototype - Figma Make, Google Stitch, Galileo AI and Uizard.

My approach: I tried building the same prototype, using text prompt. My starting point was the 1-pager for LivKidz at Home, and then subsequent prompting and iterations to get the prototype ,the way I wanted. Also, I used the code editor and “point & edit” functionality, to make specific changes instead of prompting.

Takeaways

1. These tools will help you go a long way, but not all the way. After several iterations, I could get to probably 70% there in terms of functionality and flow, but not all the way in terms of getting to hi-fidelity mocks. It lacks the final finish and polishing, which you need before you can give it to developers to start building. You will require a designer to do the last 30% and get it production ready.

2. Weeks to days: Get to insights faster. The biggest advantage of using AI here was that I get could get to a working prototype faster, all by myself in a few days, which otherwise might have taken few weeks. Just with the right prompts, you can get to a working prototype which you can use to socialize with colleagues to get feedback and do customer usability testing with

3. Getting imagery right is hard. Be prepared for compromises. For example, as the hero image on the landing page, I wanted to show an instructor engaging a child at home in an Indian urban setting. I couldn’t get to the right image despite numerous iterations (20+). It kept messing up (e.g. showed non Indian subjects, or classroom settings, or random art images)

4. Be ready for an iterative process: It took me several days and 30+ iterations to get to a decent working prototype. So you have to be patient and your prompting skills determine how fast you can get to the right output.

   

5. Tools are quite smart and they are getting smarter by the day. With simple prompts, the tools could implement the right business logic in the prototype . As an example, it doesn’t have payment screen in the trial session booking, unlike paid session. It also shows lower pricing per child, if the parent chooses group booking. I didn’t have to tell AI explicitly to implement these logic in the prototype

6. Hallucinations are real! AI sometimes lies (and pretty confidently too!): Models make up things/Incorrect responses are a reality. Many times, AI will tell you (and pretty confidently too) that it has fixed the issue based on your feedback, but it has not! So do cross check and verify!

7. Figma Make was the winner. I was able to get to the output that I wanted with Figma Make. With other tools, I couldn’t get there. Figma make provided the best output and the most granular control (e.g. ability to go into code, point to a specific section and make minor changes) to get the output you want. Also, it shows you its reasoning steps when you prompt, so that you understand how it is building based on your instructions, which can be useful in the iterations further.

Above are some of my learnings. Your experience may vary. Feel free to share your learnings through comments below.

I have published the prototype i generated using Figma Make here. Check it out!

P.S. This post is the 1st post in the series “How I AI with LivKidz”. Through this series, my goal is to share my practical, hands-on learnings in building with AI as part of my own venture. Follow me on LinkedIn, if you would like to bet updated about future posts.