How I built a viral religious app with 100% AI-generated code & content

Back in March, I wrote an article that went viral, breaking down exactly how I built a SaaS product only using AI generated code.

Due to its popularity, I decided I’d write another article, this time about Hindu Pray, an AI-powered WhatsApp bot I built, built with not only 100% AI-generated code, but also AI-generated content.

Hindu Pray - The #1 app for Hindus across the world

Hindu Pray as a project had almost the entire stack being powered by AI. In this article, I’ll break down each step in the process, from ideation to launch to TikTok virality.

I shared my article across various social platforms last time, with Reddit getting plenty of traction (and disagreement). With almost a decade of product-building experience in my career, I’ve never been more bullish about the space than I am today. The lines between traditional tech roles are being destroyed in front of our very eyes. Hindu Pray is living proof of this, and there are far more talented people than me who are building spectacular things.

The rulebook is being written in front of our very eyes.

Enjoy the ride.

ChatGPT’s depiction of me building Hindu Pray

As Nassim Taleb says, the sign of a great piece of writing is to be able to jump to any part and get value.

Here are the article chapters:

1. Intersecting AI, Hinduism & Family Whatsapp Groups
2. Crafting the funnel
3. Building with Meta’s API
4. Creating PRDs and Landing Pages
5. Midjourney prompting
6. I love edge functions
7. Insufficient MVP
8. Automating image prayers
9. Cursor with Claude 3.7 Sonnet
10. Debugging with MCPs
11. Launch & Internal feedback
12. ‘Warming up’ TikTok & Runway ML
13. Viral TikTok formats
14. Testing Reelfarm & Storyshort.ai
15. Future plans for Hindu Pray

Hindu Pray was born from an idea to intersect religion, AI and South Asian family WhatsApp groups.

At a high level, this is how it was set up:

The plan

1. Create viral TikTok videos about AI gods & prayer
2. Use TikTok link in bio to drive traffic to a landing page
3. Capture the user’s prayer preferences in a form
4. Offer free or premium image prayer subscriptions
5. Pay with Stripe
6. Start interacting with WhatsApp bot
7. Get your first prayer
8. Get sent daily prayers by replying with the 🙏 emoji

The stack

1. IDE — Cursor with Claude 3.7
2. Database & Functions — Supabase, Edge functions
3. Chatbot— WhatsApp/Meta
4. Landing page — Vercel template, Nextjs, Shadcn
5. Deployment — Vercel & Github
6. Domain — GoDaddy
7. Screen recordings — Hotjar
8. Analytics — Hotjar / Vercel
9. Product Manager — GPT 4o

The costs ($100-150/month)

1. Cursor — $16/month (annual plan)
2. Meta / Whatsapp — $5/month
3. Anthropic API — $2/month (LLM)
4. GoDaddy — $18/year (Domain)
5. Google Workspace — $8/mo (Emails)
6. TikTok followers — $25 (one-off)
7. Reelfarm — $50/month (AI content generator)
8. Storyshort — $50/month (replaced Reelfarm)
9. Supabase — $0/month (free tier, <50K users, Database)
10. Vercel — $0/month (Hosting)
11. Hotjar — $0/month (Analytics)

Disclaimer: many of the $0/month products do start charging when you get to a certain # of users, which is fine as you’ll be making money by then.

This project was slightly more expensive than my previous project Vehicle Expiry Tracker, mainly due to the marketing channel being TikTok instead of Google Ads. Whilst TikTok is ‘free’ traffic, there are costs associated with creating faceless content.

Let’s get into it!

1. Intersecting AI, Hinduism & Family Whatsapp Groups

One of the most classic behaviours in the South Asian community is participating in a family WhatsApp group. It’s almost a rite of passage at this point.

Another important part of South Asian culture is prayer. A holy act, I’ve seen my mother pray almost every day for her entire life. Prayer is a powerful vehicle for progress, perseverance and strength. I’m personally a big believer in affirmations and prayer, having benefited from them greatly over the last few years.

My mother often puts YouTube videos of chants on in the background at home, and when you take a peek at the comments in these videos, they often praise how the chants helped them overcome some obstacle or physical ailment.

With this, I quickly thought of a type of ‘prayer app’. But instead of creating an entire iOS app (and going through what at the time was still a 30% tax from Apple), I decided to keep the experience centred around the behaviour that already existed for the South Asian & Hindu community — sharing things in WhatsApp groups.

A conversation with a South Asian friend of mine

The WhatsApp group chat was key for me in this project. I’ve spent a large part of my career trying to create habits, and have found a lot of resistance in adopting new behaviours, simply because existing habits are directly competing with any new habit you’re trying to encourage.

I also hadn’t seen many AI-powered prayer/religious products for the Hindu market (which I see as being greatly underserved to this day). So the idea of something that would access all the holy texts and give you holy texts and guidance seemed like a powerful sell and easily possible with large language models (LLMs).

Documenting the structure of holy texts in Hinduism

2. Crafting the funnel

With the popularity of TikTok amongst the religious and South Asian community, I decided I’d create faceless content to drive traffic to the landing page. Then there would be a WhatsApp bot to start the conversation with and receive prayers from.

I initially looked at third-party solutions for hosting my WhatsApp bot. I stumbled onto Webots, which does exactly this. They have some powerful functionality, especially around LLM integrations. But there were some restrictions, especially with the fact I wanted to customise and create a workflow around my prompt where a specific prayer would be sent to users, and ran into some limitations.

WeBots - DIY WhatsApp & AI Chatbots

I decided to, therefore, go ahead with the direct Meta route, which I was advised would be hard and is reserved for developers…

Honourable mention — Before & afters

At the time I was building Hindu Pray, I was reading ‘’ by Jay Abraham. In there, he talks about purchases that a client makes directly before or directly after your product/service.

Getting Everything You Can Out of All You've Got: 21 Ways You Can Out-Think, Out-Perform, and…

And it made me think about what religious people might be doing during the act of planning their prayers. I thought about ornament shops, which would sell statues. Perhaps those who are physically praying may be interested in digitally praying at the same time…

I parked this idea for now as I wanted to own my customer acquisition channel to begin with.

Now, onto getting set up with Meta.

3. Building with Meta’s API

Reading through Meta’s general & WhatsApp API docs made me realise something about software development.

A developer is a skilled middleman who provides a more human access point, instead of highly unintuitive documentation that other developers have created.

It’s like a self-fulfilling prophecy.

You need developers to build programs. They write incomprehensible documents that only other people like them can understand, thus preventing access to non-developers.

The only way to get in is to become a developer, so…

I decided to read through plenty of Meta’s documents and learnt about what would be required to make Hindu Pray work effectively on WhatsApp.

I’d need to upload my beautiful AI-generated media through a hyper-secure server, I’d need to set up pre-approved message templates, set up some webhooks and ensure that my phone number wasn’t already connected to WhatsApp.

One of the nuances of Meta’s API is that you can’t just blindly send messages to users without their permission. If you haven’t received a message from a user in the last 24 hours, you can only send them one of the pre-approved ‘message templates’. I spent a lot of time actually crafting out these messages based on the key steps in the user experience, such as asking people for their prayer preferences, sending them their prayer trigger and error messaging.

There was also the further complication that to upload images to your prayer template, and I find out that you need to upload your image to Meta, then return an ID for that media item and then send that ID in your message, even if it is already included in the templated message. These nuances, when it came to different systems, became the norm rather than the exception as I started to build more things, and it gave me an appreciation of how it seems to always take more time than you might think to build out some of this functionality, especially when you have to rely on external APIs.

The first image a user gets after talking to Hindu Pray

My overall take on building apps within the Meta ecosystem?

Buckle up for a hell of a (unintuitive) ride.

3. Creating PRDs and Landing Pages

I created a PRD to outline the overall architecture of Hindu Pray, and the inter-connectivity between:

Landing Page → Supabase → Stripe →Whatsapp → Supabase → Claude → Back to Whatsapp

By capturing the true end-to-end flow of a project, I’ve always found Cursor to be far more effective at building things the right way, especially for creating copy in any customer-facing part of the website.

I originally started with the Landing page on the left, which was pretty much a one-shot prompt with Cursor, and then some finer tweaks like the gradient hero and italicised underline style.

I created a semi-A/B test by creating a TikTok-specific version, which had slightly different copy, and had testimonials immediately, rather than after the explanation of how the product worked.

Cursor is remarkably good at creating best-practice landing pages that contain the necessary sections to help you convert visitors.

What a time to be alive.

5. Midjourney

For one of the sections of the landing page, I wanted to explain to the user that there would be various Gods that you could pray to, and to facilitate a more vibrant experience, I leveraged Midjourney to create a set of images for various gods in Hinduism.

Below are examples of prompts that I used to create these. The output was surprisingly high, and initially broke my assumption that Midjourney may only have been trained on Western imagery.

To help create these Midjourney prompts, I first used ChatGPT to come up with a few formats and tested those. Then I took the best format and then ran that up with other gods and scenes.

With AI media generation, always test early on different formats before driving at scale

I also used prompts about Hindu/South Asian scenes to craft background images that would eventually be used for image prayers. Below was my GPT output:

And then my Midjourney output. Some truly spectacular images!

6. I love edge functions

I won’t say too much about this, as I went into detail in the last article about how wonderful edge functions are.

Serverless, Lambda, Edge functions, whatever you want to call them, are just wonderful ways to make one app share something to another one. Supabase handled my WhatsApp message triggers, Claude LLM prayer generation, and the image transformation with Imgix.

7. Insufficient MVP

I had initially thought that text-only prayers would be a viable MVP for launch. I had initially made this pay with a free trial. As the landing page started to take shape and I tested the flows end to end, something felt off, and I felt there was something critical missing for launch, even as an MVP.

I looked at Innertune affirmations and Pray.ai as two examples of products that were already successfully working for the Western/Christian markets, and one of the things they both did quite well was visuals.

They had a lot of imagery around their product, whether it be the actual affirmation on a nice image background or just having images throughout the landing page, there was plenty of visual flair on show.

I decided to pause the launch until I automated the image prayer generation flow, which would be the premium & charged product, leaving the text-only prayer as the free entry point.

A key hypothesis that was developed was that image prayers would create a network effect inside WhatsApp groups, and with a clear watermark ‘Created by Hindu Pray’ it would drive people into searching for us.

I proceeded to create a large set of backgrounds that my edge functions could choose from when sending image prayers.

Background templates created using Midjourney

8. Automating image prayers

The technical concept of image prayer was quite straightforward.

Take the text generated by Claude and overlay it onto any one of the background templates (in future, smartly matching the prayer text with the relevant image type).

However, this was near-impossible to do within the Supabase edge function. I tried many different paths, and it was one of the only times I genuinely asked my engineering friends what I should, and even they couldn’t crack it within Supabase.

Eventually, I decided to pursue an alternative path and use an image transformation API, Imgix. This worked a charm after a few tweaks, and after finding the right documentation, I fed Cursor the links, and it generated the code for me to use within Supabase.

It led me to create a new edge function, which was not optimal. But with how important this piece of functionality was, I deemed it necessary to add this complexity.

Below are some examples of the image prayers in action!

9. Cursor with Claude 3.7 Sonnet

I could write an entire article about how great Claude’s new 3.7 Sonnet model is. Initially, when I used it, I found it went off to the races too fast, but I learnt that when I gave it a proper implementation plan (you know, like a real developer) that it worked through problems in a more controlled manner.

Here’s an example of it cranking through the time zone implementation in the landing page:

The remarkable moment (highlighted in red) was when Claude detected that the WhatsApp webhook code also needed to be updated to stay aligned with what the landing page was becoming.

It was remarkable because I didn’t initially list out this requirement in the PRD, and I realised after I submitted the message, at which point I saw it had already realised the same thing I did!

This level of thoroughness from an AI assistant creates one of those moments where you really wonder how many human brains you need alongside you to build in this day and age.

10. Debugging with MCPs

Thanks to the functionality of Hindu Pray not having much of a frontend that I controlled, and WhatsApp shouldering a lot of the load, I didn’t run into many issues with repetitive bugs.

This coincided with the launch of the official Supabase MCP, which solved one of the biggest pet peeves I had using Cursor. Explaining my schema every single time (amongst other things). With the MCP, Cursor was able to read my Supabase tables, and with recent launches, it's now able to read logs from edge functions as well, saving not only time but, more importantly, those tedious clicks that you wish could be automated.

There have been some security issues around MCP and Cursor generally, but at this point, getting hacked is part of the fun.

If you haven’t already, make sure to set up your MCP.

Model context protocol (MCP) | Supabase Docs

11. Launch & Internal feedback

With the app ready, I did a friends and family launch and shared it across targeted group chats. The feedback didn’t get the adoption that I was hoping for, and got plenty of ‘that’s a nice idea’.

One common response was that deeply devoted people who already pray regularly didn’t see much added value, which I knew might be the case, but even my mum said that for her it wasn’t a value add as she’d just come up with prayer in her head and she knew enough of the holy texts.

Trust emerged as another concern — people within South Asian communities can be hesitant to share personal life details and goals, as these are sensitive subjects, and I thought the privacy element was particularly interesting as a finding. Pricing was also a factor, given the frugal tendencies of South Asian culture. Still, I had hypothesised that when it comes to religion, people would be willing to spend money on something they believe would improve their lives.

12. ‘Warming up’ TikTok & Runway ML videos

Alongside the family and friends launch, I began preparing for the mass marketing campaign that would happen through a dedicated Hindu Pray TikTok account.

TikTok - Make Your Day

When creating a new TikTok account, it helps to do a ‘warm up’, which gets your algorithm in the mode of targeting the exact user that you want to target.

To ‘warm up’ the Hindu Pray TikTok, I consumed Hindu content and discarded everything else on my feed for a good 40 minutes. Once my feed became only Hindu content, it was time to buy some followers and get that all-important link in bio.

I’m personally not a fan of buying followers, but I needed to get the bio link to have any hope of driving traffic, and I didn’t want to have to wait for the first proper viral video to start getting traffic and giving me valuable insight (and early users).

I then went on to RunwayML to begin creating striking AI-generated gods. Initially it seemed promising, but ultimately ended up not being a fruitful exercise.

RunwayML’s functionalities were some of the hardest AI tooling I’ve had to work with, with its own prompt structuring required, and so I sought out simpler solutions to begin with.

RunwayML examples with Ganesh & Lakshmi

13. Viral TikTok formats

Having built a TikTok account with a few viral formats in the past, I knew the algorithmic game was quite scientific.

Finding videos from accounts that have done disproportionately well, relative to their followers, was they to finding video formats that would succeed.

I landed on a style titled ‘What people think gods look like vs. How they’re described in the holy texts’ as a captivating and provocative format.

It turned out to be right and led to two videos, which have around 100K views so far.

These brought over a steady stream of users to the site, and there were plenty that dropped off at the form stage, around 60% dropoff occured from form questions to sign up (for any package).

This gave me plenty of insight to chew on and iterate using.

14. Testing Reelfarm & Storyshort.ai

I’d seen a lot about Reelfarm on Twitter, but I found out quickly that it was not suited for audiences outside of Gen Z. My target audience was firmly on the opposite end of the Gen Z spectrum. So anything referring to things like ‘Run don’t walk to Hindu Pray’ or ‘Omg why did no one tell me about Hindu Pray’ didn’t make sense, as hilarious as it was.

Reelfarm output videos for Hindu Pray

Storyshort.ai, however, was far more fruitful and had one of the best AI tool experiences I’ve had. They’ve done a great job of adapting the common timeline view in video editors and converting it into a vertical timeline.

Storyshort.ai’s timeline functionality on a video about Lord Ram

I used GPT to create scripts of famous Hindu stories, such as how Lord Ganesh got his head, and let Storyshot run with it and come up with videos. Below is an example of a story about Lord Ram:

15. Future plans for Hindu Pray

For now, the platform user base is growing slowly and is not yet profitable. The TikTok account gets decent engagement but hasn’t had an ultra-viral hit yet.

I’ll continue to drive traffic to the site using faceless TikTok content, monitor Hotjar and user feedback on the prayers in the coming weeks and months.

If you’d like to get help on getting set up with vibe-coding or want to work on a product you’re building, drop me an email at pegasus.dco@gmail.com, to discuss more.

If you found any part of this article valuable, make sure to highlight and clap!

Sakky B