The 'First Dollar' Stack

This post was originally published on my Substack publication as The 'First Dollar' Stack.

I spent a few hours yesterday wrestling with an AI that confidently suggested seven different ways to implement a toast notification system, and all of them broke.

Picture yourself driving through an unmapped wilderness at midnight, where you have a fast car and a full tank of gas but absolutely no roads exist to guide you. Every turn demands a guess based on intuition rather than data, and you hit dead ends constantly as you reverse and try another route that might not exist either.

This frustrating experience matches exactly what happens when you build a V1 with obscure tech stacks, because you ask an AI to write a feature using a niche database or a complex deployment setup, and the AI hallucinates fixes because it simply lacks the training data to know better.

Now picture the exact opposite scenario where you drive through a heavily mapped city with GPS that knows every street, every pothole, and every shortcut that locals have discovered over decades.

React, Supabase, and Cloudflare make up what I call the First Dollar stack, and the reason this combination works so well is that LLMs have ingested millions of tokens of public code and documentation for these specific tools.

The AI knows every pothole and fast lane because thousands of developers have already driven these roads and published their maps on GitHub, and you stop fighting the tools and start shipping features that actually work.

We need to start with the reality of production debugging instead of abstract definitions, because code is a liability that grows heavier with every line you add. Migrations cost time that you could spend acquiring users, and AI blind spots create infinite loops of broken code that waste days you will never get back.

When an AI tool does not know a stack perfectly, it guesses confidently and produces code that looks correct but fails in edge cases you only discover at 2 AM on a Saturday.

You end up wasting three days fixing authentication bugs instead of acquiring users, and three weeks pass while you remain trapped in the setup phase wondering where all your momentum went.

The First Dollar stack prevents this entire downward spiral because the AI writes correct code on the first try, and the models know React and Supabase better than any human developer ever could since they have studied every public implementation ever published.

I want to share a model I call AI Training Depth that helps you evaluate tech choices more effectively.

This model makes you break your thinking into three clear buckets that show why giving the AI familiar tools speeds up development dramatically.

The first bucket holds what I call The Well-Lit Roads, and these are tools with massive amounts of public documentation (like React). The AI can write React components flawlessly because it has seen every variation on GitHub and understands the patterns that work in production versus the ones that create subtle bugs.

The second bucket contains The Standard Connectors, and these are predictable API patterns like Supabase that follow conventions the AI already knows. The AI knows exactly how to query the database without complex custom logic, because Supabase designed their API to match what developers already expect from PostgreSQL.

The third bucket holds The Invisible Delivery, and these are deployment methods that remove manual configuration entirely like Cloudflare Pages / Workers. The AI does not need to guess your server setup because there is no server to set up, and you simply push your code and watch it go live without touching a single configuration file.

Here is the execution plan for prompting an AI to build this stack, and you must enforce these boundaries early before the AI starts suggesting alternatives that sound appealing but will cost you weeks of debugging.

First, you define the strict stack constraints in your initial prompt so the AI understands the boundaries from the very beginning. It's also great to start with a framework supported by Cloudflare so you get a complete boilerplate app to start with.

Second, you instruct the AI to generate the Supabase database schema first, because a clear data model makes everything else flow naturally.

Third, you tell the AI to build the React frontend connecting to that schema, and fourth, you deploy directly to Cloudflare Workers automatically without manual steps.

If the AI suggests a different database or proposes an alternative hosting solution, you force it back to the First Dollar stack immediately without negotiation.

Consistency keeps the AI on the well-mapped city streets where it can drive confidently, and stepping off those roads invites the hallucinations and broken code that kill your momentum.

We have all lost a weekend to server configuration errors and authentication flows that worked in development but broke the moment real users touched them. Take a breath and appreciate that this stack eliminates an entire category of late-night debugging sessions.

I want to share the Stack Enforcer prompt template with you, because this simple prompt forces the AI to stay within the boundaries of its deepest training data. You literally constrain the AI to the tools it knows best, and the mechanism works because explicit constraints prevent the AI from wandering into territory where it starts guessing.

Setting these boundaries stops the AI from importing random libraries that seem helpful but create dependency nightmares you discover months later, and you get clean code that works on the first copy-paste because the AI stays on familiar ground.

You will act as my senior developer. We are building a V1 application.
You will strictly use React, Supabase, and Cloudflare.
Do not suggest custom databases, alternative hosting, or extra libraries.
Write the Supabase SQL schema first.

I started spending a ton of time on X lately and so its only natural that this week's discovery comes from there. Emre built markdown.new, a FREE tool that can convert any URL to clean Markdown using Cloudflare's native text/markdown content type. I append it to any URL I send to my AI agents to keep the costs down and context short and sweet.