The stack is simple; the workflow is the product
The basic home-project setup is GitHub for code, Vercel for deployment, Supabase for database and real-time behavior, and Codex for planning, implementation, testing, and iteration. Free tiers can cover a small prototype. The AI subscription is not a magic builder button; it is a teammate that performs better when the job has boundaries.
What each service does
- GitHub: version history, backup, issues, and collaboration.
- Vercel: builds the web application and gives it a public URL.
- Supabase: stores shared data, handles authentication, and can stream real-time changes.
- Codex: reads the workspace, proposes plans, changes code, runs checks, and collects proof.
The five-step build loop
- Set up a real repository and connect deployment before the app becomes complicated.
- Ask planning questions and sketch the user flow.
- Divide the product into minimum viable slices.
- Build one slice, test it, inspect it as a user, and only then continue.
- Test refreshes, mobile layouts, late joins, duplicate actions, empty states, and failure states.
The MVP prompt contract
Every build request should include requirements, acceptance criteria, tests, and proof. This forces “done” to mean something visible. It also makes it easier to stop an agent from polishing the wrong thing.
Costs and limits
Free tiers change. Check current limits before promising a free production service. Add spending alerts, protect server keys, use row-level security in Supabase, and do not treat a public prototype like an audited enterprise system.
Software, links, and setup
Prompt pack
Planning prompt
Before writing code, interview me about the user, problem, critical journey, data that must persist, privacy concerns, failure modes, and what proof will demonstrate value. Then propose no more than three MVP slices.
MVP build prompt
Build only MVP slice [name]. Requirements: [list]. Acceptance criteria: [observable outcomes]. Tests: [commands and user journeys]. Proof: [screenshots, logs, URLs, database evidence]. Preserve existing behavior. Stop after the criteria are proven and list risks before suggesting the next slice.
Launch check
Test the deployed URL like a real first-time user on desktop and mobile. Include refresh, sign-out/sign-in, empty data, duplicate submissions, slow network, and permission boundaries. Do not mark complete without evidence.
What to remember
- The useful unit of work is a proven MVP slice.
- Connect version control and deployment early.
- Protect server secrets and enable database policies.
- Requirements, tests, and proof make AI work legible.
Then test it, document it, and make the next version easier for someone else to reproduce.