I work as a VP at Citi, building AI products that streamline workflows across the organisation. One of the platforms I helped build is used by over 200,000 users, and working at that scale has shaped my thinking around AI-assisted development. I’ve been using GitHub Copilot since it was little more than an autocomplete engine, and today I rely on four different AI coding tools daily. Many more are quietly embedded in our pipelines, making AI feel less like a novelty and more like part of the engineering substrate. AI Coding Tools AI Coding Tools With the rise of Cursor, Devin, Copilot Agents, Claude Code, Codex, and countless other tools, building software has fundamentally changed. If you can describe your requirements well and understand that most engineering problems have multiple valid solutions AI can carry you a long way. My colleagues and I use multiple agents to accelerate development dramatically. Most of these tools are public, many are free, and even their baseline capabilities are enough for an individual to ship a working prototype sometimes even a production-ready application. Some developers dislike this shift. They argue that anyone can now build and deploy software without necessarily understanding scalability, security, or architecture. That’s true but it’s also the point. The friction from idea to execution is collapsing. Innovation increases when the cost of experimentation falls. Reality Check Reality Check Given all this, you would expect everyone to be taking advantage of these tools. They aren’t. During a mandatory block leave, I was scrolling through WhatsApp statuses and noticed a friend had launched a website for his fruit-subscription business. He had worked hard to build his brand. When I opened the site, the enthusiasm quickly faded. I build web products for a living and what I saw was painful: misaligned layouts, incorrect image resolutions, broken icons, and a clunky flow. In today’s digital landscape, a brand’s website is often its first impression. A poor experience can destroy conversion. Let’s Build I decided to fix it myself. Using ChatGPT, GitHub Copilot, Google Antigravity, and Vercel, I built a fully working app in six hours — a project that would normally take a small team 3–4 weeks. ChatGPT, GitHub Copilot, Google Antigravity, and Vercel six hours 3–4 weeks The delivered features included: E-commerce
POS (offline point-of-sale)
Bulk orders
Refund workflows
Inventory management & forecasting
Analytics
Subscriptions
Warehouse & delivery fleet management
Customer reviews
Admin portal
Delivery partner portal E-commerce POS (offline point-of-sale) Bulk orders Refund workflows Inventory management & forecasting Analytics Subscriptions Warehouse & delivery fleet management Customer reviews Admin portal Delivery partner portal Everything was functional, styled, deployed, and ready for real usage. Learnings I recently came across a quote by a founder: Why should companies pay for SaaS (HR/CRM/ERP/etc.) when they could just vibe code them? Why should companies pay for SaaS (HR/CRM/ERP/etc.) when they could just vibe code them? Why should companies pay for SaaS (HR/CRM/ERP/etc.) when they could just vibe code them? My interpretation is simple: SaaS products are built to serve broad markets, which means they must be generic. Generic solutions always come with trade-offs. They’re like ready-made clothes you can wear them, but they don’t always fit. If organisations can build internal software that is tailored to their exact workflows, and if AI can accelerate that development enough to make it economical, then traditional SaaS becomes unnecessary overhead. Most companies use barely 20–30% of the features of large SaaS platforms anyway. Process Here’s the workflow I followed end-to-end. Setting the context I opened a new ChatGPT thread and established the entire requirement context there: existing website link
business model
customer flow
product categories
screenshots of the current UI existing website link business model customer flow product categories screenshots of the current UI GPT-5 refined its understanding through a few clarifications. Choosing the tech stack I went with: Next.js
TailwindCSS
Shadcn for component UI
Prisma ORM
Neon for the database
Vercel for deployment Next.js Next.js TailwindCSS TailwindCSS Shadcn for component UI Shadcn Prisma ORM Prisma Neon for the database Neon Vercel for deployment Vercel Prompt I asked ChatGPT to generate a Copilot Agent-ready prompt tailored for VSCode. After a few iterations, it produced a detailed, structured prompt defining the full project architecture, both customer-facing and admin-facing features, data models, routing, component design, and deployment instructions. Copilot Agent-ready prompt You are a senior full-stack dev. Build a modern fruit-subscription web-app + admin panel for a local fruit store, using:

- Next.js + TypeScript  
- shadcn/ui + Tailwind CSS for UI/UX & components  
- Database (Postgres / MySQL / SQLite / serverless) for products, customers, subscriptions, orders, inventory  
- Hosting on Vercel, code in GitHub  

Requirements / Features:

## Public Store (Customer­-facing)  
- Responsive, accessible UI built with shadcn/ui components (cards, buttons, forms, dialogs, modals, etc.)  
- Product catalog with images, description, price, “fresh/seasonal/organic/exotic” tagging  
- Subscription module: choose recurring fruit-box (weekly/biweekly/monthly), configurable preferences (e.g. mixed/seasonal/custom), quantity, delivery address  
- One-time order option besides subscription  
- User authentication (signup/login), user profile (address, preferences), order history, manage subscriptions (skip/pause/resume)  
- Clean checkout UI (can stub payment initially), order confirmation, simple order summary  

## Admin Panel (Vendor / Shop Owner)  
- Sidebar + dashboard layout using shadcn/ui (nav menu, responsive, collapsible)  
- Views/tables for: products (add/edit/delete, inventory/stock levels), subscriptions (list active, upcoming deliveries), orders (one-time & subscription-based), customers (with order history, address, subscription status)  
- CRM-style interface: ability to filter/search customers/orders, export CSV/JSON for marketing, view churn, recurring-revenue metrics  
- Analytics dashboard: key metrics — active subscriptions count, monthly recurring revenue (MRR), order volume, retention rate, seasonal demand patterns, stock usage / inventory warnings  
- Basic admin UI: add new fruits, update prices / availability, set seasonal availability, manage deliveries  

## Project Setup & Deployment  
- Initialize Next.js + TypeScript + Tailwind + shadcn/ui using the shadcn CLI (`npx shadcn@latest init`)  
- Maintain all components in `components/ui/` so UI is fully owned and customizable  
- Use API routes or server-components for backend logic (CRUD, subscription scheduling, order management)  
- Store code in GitHub repo, enable automatic Vercel deployments on push  
- Include README with setup steps, env vars, database setup, deployment instructions  

## Code Hygiene & Extendability  
- Use TypeScript types/interfaces for products, orders, subscriptions, customers  
- Structure project for maintainability — clear separation: UI components, pages/routes, API routes/business logic, data layer / ORM  
- Build admin & store UI using composable shadcn UI components so customizing design/theme later is easy  
- Stub payment gateway integration initially; design code so it can be plugged in later (e.g. Stripe, Razorpay) You are a senior full-stack dev. Build a modern fruit-subscription web-app + admin panel for a local fruit store, using:

- Next.js + TypeScript  
- shadcn/ui + Tailwind CSS for UI/UX & components  
- Database (Postgres / MySQL / SQLite / serverless) for products, customers, subscriptions, orders, inventory  
- Hosting on Vercel, code in GitHub  

Requirements / Features:

## Public Store (Customer­-facing)  
- Responsive, accessible UI built with shadcn/ui components (cards, buttons, forms, dialogs, modals, etc.)  
- Product catalog with images, description, price, “fresh/seasonal/organic/exotic” tagging  
- Subscription module: choose recurring fruit-box (weekly/biweekly/monthly), configurable preferences (e.g. mixed/seasonal/custom), quantity, delivery address  
- One-time order option besides subscription  
- User authentication (signup/login), user profile (address, preferences), order history, manage subscriptions (skip/pause/resume)  
- Clean checkout UI (can stub payment initially), order confirmation, simple order summary  

## Admin Panel (Vendor / Shop Owner)  
- Sidebar + dashboard layout using shadcn/ui (nav menu, responsive, collapsible)  
- Views/tables for: products (add/edit/delete, inventory/stock levels), subscriptions (list active, upcoming deliveries), orders (one-time & subscription-based), customers (with order history, address, subscription status)  
- CRM-style interface: ability to filter/search customers/orders, export CSV/JSON for marketing, view churn, recurring-revenue metrics  
- Analytics dashboard: key metrics — active subscriptions count, monthly recurring revenue (MRR), order volume, retention rate, seasonal demand patterns, stock usage / inventory warnings  
- Basic admin UI: add new fruits, update prices / availability, set seasonal availability, manage deliveries  

## Project Setup & Deployment  
- Initialize Next.js + TypeScript + Tailwind + shadcn/ui using the shadcn CLI (`npx shadcn@latest init`)  
- Maintain all components in `components/ui/` so UI is fully owned and customizable  
- Use API routes or server-components for backend logic (CRUD, subscription scheduling, order management)  
- Store code in GitHub repo, enable automatic Vercel deployments on push  
- Include README with setup steps, env vars, database setup, deployment instructions  

## Code Hygiene & Extendability  
- Use TypeScript types/interfaces for products, orders, subscriptions, customers  
- Structure project for maintainability — clear separation: UI components, pages/routes, API routes/business logic, data layer / ORM  
- Build admin & store UI using composable shadcn UI components so customizing design/theme later is easy  
- Stub payment gateway integration initially; design code so it can be plugged in later (e.g. Stripe, Razorpay) I pasted it into VSCode and Copilot began scaffolding the entire project layouts, routes, UI components, Prisma schemas, and API routes. Within minutes, I had a functional first version deployed on Vercel. Iterating with Agents (From Phone!) Once the initial build was live, I asked GPT-5 to help ideate additional advanced features. I got a list of nine, each with a structured prompt. For example: Feature: Customer reviews & ratings for products

Context:
- We have Users and Products (Prisma + PostgreSQL)
- Next.js + TypeScript + shadcn/ui frontend

Requirements:
- Extend Prisma schema: create Review model:
    id (uuid), productId, userId, rating (int 1–5), title (optional), comment (text), createdAt, updatedAt, verifiedPurchase (boolean), status (enum: PENDING | APPROVED | REJECTED)

- API endpoints:
    POST /api/products/[id]/reviews — allow authenticated customers who have purchased the product to submit review (store as PENDING by default).
    GET /api/products/[id]/reviews — return list of APPROVED reviews for that product, include user name (or anonymized), rating, comment, date.
    Admin endpoints under /api/admin/reviews — list, approve, reject reviews.

- Frontend:
   On product detail page: show average rating (stars), number of reviews, list of recent reviews (rating + comment + date).
   On customer's order history page: after delivery (or on order detail page), show a “Leave a review” button.

- Admin UI: page under /admin/reviews to moderate reviews, approve/reject, view flagged reviews.

- Validation & restrictions:
   - Only users who have purchased the product can review (check order history).
   - Prevent duplicate reviews by same user for same order/product.
   - Input validation (rating bounds, comment length).

Please generate:
1. Prisma model for Review.
2. API routes (customer and admin) TypeScript code.
3. Frontend components for submitting review + displaying reviews (shadcn/ui).
4. Admin review moderation page skeleton. Feature: Customer reviews & ratings for products

Context:
- We have Users and Products (Prisma + PostgreSQL)
- Next.js + TypeScript + shadcn/ui frontend

Requirements:
- Extend Prisma schema: create Review model:
    id (uuid), productId, userId, rating (int 1–5), title (optional), comment (text), createdAt, updatedAt, verifiedPurchase (boolean), status (enum: PENDING | APPROVED | REJECTED)

- API endpoints:
    POST /api/products/[id]/reviews — allow authenticated customers who have purchased the product to submit review (store as PENDING by default).
    GET /api/products/[id]/reviews — return list of APPROVED reviews for that product, include user name (or anonymized), rating, comment, date.
    Admin endpoints under /api/admin/reviews — list, approve, reject reviews.

- Frontend:
   On product detail page: show average rating (stars), number of reviews, list of recent reviews (rating + comment + date).
   On customer's order history page: after delivery (or on order detail page), show a “Leave a review” button.

- Admin UI: page under /admin/reviews to moderate reviews, approve/reject, view flagged reviews.

- Validation & restrictions:
   - Only users who have purchased the product can review (check order history).
   - Prevent duplicate reviews by same user for same order/product.
   - Input validation (rating bounds, comment length).

Please generate:
1. Prisma model for Review.
2. API routes (customer and admin) TypeScript code.
3. Frontend components for submitting review + displaying reviews (shadcn/ui).
4. Admin review moderation page skeleton. I triggered these prompts in GitHub Agent Modefrom my phone.Nine agent runs → seven completed PRs → each with a preview link from Vercel. GitHub Agent Mode I reviewed code, tested previews, and merged PRs all from my phone. The workflow felt futuristic. The ability to architect, iterate, and deploy a full application from a phone would have been unimaginable a year ago. ⭐ Final Thoughts This experience reinforced a simple truth:We are entering an era where a single engineer, paired with AI, can outperform entire teams. We are entering an era where a single engineer, paired with AI, can outperform entire teams. Not because AI replaces expertise, but because it amplifies it.Not because engineering becomes trivial, but because execution becomes frictionless. The companies that learn to build tailored internal tools at AI speed will outpace those locked into bloated SaaS ecosystems. The future of software development is not about writing more code; it’s about orchestrating intelligence.

Walkthroughs, tutorials, guides, and tips. This story will teach you how to do something new or how to do something better.

Public

dashboard

Vercel

Google

The Death of SaaS? When Internal Tools Can Be Built Faster Than You Can Buy Them

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

Pushing the Limits: Running Local LLMs and a 24/7 Personal News Curator on 4GB of RAM

The Noonification: How Often Do NFTs Pass The Howey Test? (1/13/2023)

Darwin's Hybrid Intelligence to Align AI & Human Goals for Startups & VCs

The Noonification: White Man (11/26/2022)

The Noonification: The Metaverse is a Sh*tshow (11/2/2022)

100 Days of AI Day 1: From Newsletter to Podcast, Leveraging AI for Audio Transformation

Pushing the Limits: Running Local LLMs and a 24/7 Personal News Curator on 4GB of RAM

The Noonification: How Often Do NFTs Pass The Howey Test? (1/13/2023)

Darwin's Hybrid Intelligence to Align AI & Human Goals for Startups & VCs

The Noonification: White Man (11/26/2022)

The Noonification: The Metaverse is a Sh*tshow (11/2/2022)

100 Days of AI Day 1: From Newsletter to Podcast, Leveraging AI for Audio Transformation

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps