Set Up Your Basic Tools on Day 1

The fastest way to waste your first week with an AI coding agent is to let it write features on a laptop with nothing behind them.

No database. No repo. No deploy target. The agent will do what it always does: take the path of least resistance. It stubs persistence. It hardcodes config. It builds auth against nothing. All of it runs beautifully on your machine, and none of it survives contact with the real world. Then you pay the plumbing tax in the middle of the build, when every change is expensive, and you rewrite half of what the agent already wrote.

My rule: day 1 is plumbing, not features. Before the agent writes a single feature, you need five things standing: a database, a repo, CI, hosting, and email for sign-ups.

The stack I reach for

Keep it simple and pick tools your agent can drive directly. The ones I use all have MCP servers, which means Cursor, Claude Code or Codex can operate them instead of you copy-pasting credentials and SQL between windows:

• Database: Supabase. Postgres with an API on top, migrations built in, and an MCP server your agent can use to run schema changes itself.
• Repo + CI: GitHub. The repo is obvious. The CI is GitHub Actions, free with the repo, and the agent can read failed runs and fix them the same way you would.
• Hosting: Vercel. Push to main, it deploys. Preview URLs per branch. The agent can check a deployment's status instead of asking you to.
• Email for sign-ups: Resend. Everyone forgets this one until launch day. Auth emails, password resets, welcome flows: none of it exists until you wire a sender, and wiring it late means re-testing every auth path.

That is the whole foundation. Nothing exotic. The point is not which five tools, it is that all five exist before feature one.

The prompt to copy

Once the accounts and projects exist, hand the whole setup to the agent in one prompt:

We are starting a new project. Before any features, set up the foundation. Connect the Supabase project (URL and keys in .env), create the initial migration for a users table with id, email and created_at, initialize the GitHub repo with a main branch and a GitHub Actions workflow that runs lint and tests on every push, connect the Vercel project so main deploys automatically, and configure Resend with a from-address for auth emails. Verify each connection works: run a query against Supabase, push a test commit and confirm CI goes green, confirm the Vercel deployment URL loads, and send a test email. Report back what you connected and what is still missing. Do not build any features yet.

That last sentence is doing real work. Agents love to sneak a head start on features. Day 1 is plumbing. Hold the line.

Why this matters more with AI than without

A human developer grumbles about missing infra and works around it visibly. An agent works around it invisibly. It will not tell you it faked the database layer unless you ask. It will not flag that deploy does not exist. It produces working-looking code on top of nothing, and the nothing only shows up when you try to ship.

Plumbing first also compounds through the whole series. Tip #3 in this series is about keeping plans and docs in a separate folder that the agent maintains. Those docs describe the architecture. The architecture assumes the five tools above exist. If the foundation is real, the docs describe reality, and every future prompt builds on truth instead of stubs.

What breaks when you skip it

Skip the database and the agent invents persistence, usually as local files or in-memory stores, and you migrate real user data later under pressure. Skip the repo and CI and nothing is reviewable, nothing is reproducible, and "it works on my machine" becomes your deployment strategy. Skip hosting and the agent never thinks about production constraints: environment variables, build sizes, cold starts. Skip email and you discover on launch day that nobody can sign up, which is a special kind of launch.

All of these are cheap on day 1 and expensive on day 30. The agent does not know what day it is. You do.

Day 1: Supabase, GitHub, Vercel, Resend. Wire them up before the features, not after.

Dr Helmi posts a short AI coding tip every other day on Instagram at @doc.helmi. He is the founder of Orgenix, which builds AI that turns product decisions into software.