Hey,
Almost everyone who starts building with AI makes the same mistake. Not with the code. Before the code.
They open a new chat, type what they want to build, and let the AI start. It feels fast. Within a few minutes something is on the screen and working. They keep going, adding one thing at a time. An hour in, something breaks in a way that is hard to explain. A new section does not match the old one. Two parts of the app are doing the same job differently. The whole thing starts to feel unstable.
They blame the AI for going off track. But the mistake happened in the first ten messages, before any real code was written.
I know because I did this too.
An year ago I was building a simple client portal for one of my agency projects using Lovable. A login page, a dashboard, a place to upload files. Nothing complicated.
I opened the chat and typed: "Build me a client portal with a login and a dashboard."
The AI got to work immediately. Within a few minutes there was a working structure. I kept going. File upload section. Minimal sidebar. Recent activity on the dashboard. It handled all of it.
Forty minutes in, I asked for a simple settings page.
That is when everything fell apart.
The settings page came back in a completely different component structure than the rest of the app. The styling was inconsistent. The file naming did not match. When I tried to connect it to the login flow, there were two different authentication patterns in the same codebase.
The AI had been guessing the architecture from the start. Every message I sent, it filled in the gaps on its own. By message 10, the decisions were baked in. By message 30, undoing them would take longer than starting over.
I started over.
Why the First Messages Are Different From All the Others
Most people think of a build session like a conversation. You ask, AI answers, you ask again. Message by message. Each one roughly equal.
That is not how it works.
The first few messages do not just produce code. They set the rules for everything that follows. When you tell AI to build something without context, it makes decisions immediately. What framework to use. How to structure components. How to name files. How to handle state. Whether to build one big file or ten small ones.
It does not ask for permission. It just decides. And then every message after that builds on those early decisions.
By the time you notice something is wrong, the problem is not the last thing you asked for. The problem is buried in message three, where AI made a quiet choice you did not know about and everything since has been building on top of it.
What a Bad Session Opener Looks Like
The most common openers I see from people who are just getting into vibe coding all look like this.
"Build me a landing page."
"Create a dashboard with charts."
"Make a booking form for my business."
These are not bad intentions. They are bad instructions. There is no context, no constraint, no definition of done. The AI has almost nothing to work with, so it invents everything. It picks a framework. It picks a structure. It picks a visual style. It picks an approach to state management.
All of those decisions will shape everything you do for the next two hours. And you had no say in any of them.
What a Good Session Opener Looks Like
Before you type a single instruction in a new build session, give the AI four things.
What you are building.
One or two sentences. Not a vision document. Just enough that it knows the purpose. "A simple booking form for a photography studio. Clients fill in their name, date, and type of shoot. There is no payment. Form submits and they see a confirmation message."
What tech you want to use.
If you have a preference, say it. "Use plain HTML and Tailwind. No React." Or "Use Next.js with Tailwind. Keep it simple." If you do not have a preference, say that too: "Keep the tech as simple as possible. I do not want to learn a new framework."
What it should not include.
This is the one people skip. AI defaults to adding things. User authentication. A database. Admin panels. Analytics hooks. Responsive breakpoints for six screen sizes. Unless you tell it what to leave out, it will add everything it thinks a complete product needs. "No database. No auth. No admin panel. Just the form and the confirmation page."
What done looks like.
Not "a fully functional app." Something specific. "Done means I can open it in a browser, fill in the form, and see a thank you message. Nothing else needs to work yet."
That is it. Four things. It takes three minutes to think through and write out. It saves hours of undoing bad decisions later.
This Is Still the ROOM Method
If this sounds familiar, it should.
In Issue 01, we talked about the ROOM Method. Reduce to one feature. Outline the steps. One prompt per step. Map the whole build before you start.
The session opener is the Outline step in practice. You are not outlining the full build. You are outlining the context. You are giving AI everything it needs to make the same decisions you would make, instead of guessing.
And just like the 3-Part Prompt from Issue 03, the session opener works because it removes ambiguity before AI gets a chance to invent an answer to it.
The pattern is the same across all of these: think first, reduce ambiguity, then prompt. The session opener is just what that looks like at the start of an entire build.
A Note on Existing Builds
If you are continuing a build from a previous session, the same principle applies.
AI does not remember your last session. When you open a new chat, it is starting completely from scratch. If you just paste in the code or share the link and ask it to continue, it will re-read the structure and start guessing again. And its guesses may not match the decisions from last time.
Start every session, even a continuation, with a brief recap. What was built. What you were in the middle of. What the next step is. What must not change.
Three sentences is usually enough. The point is to hand AI an accurate starting position instead of letting it infer one.
Your Action for This Week
Start a new build. It does not have to be big. A landing page, a simple form, a personal dashboard, anything.
Before you type the first instruction, write out the four things. What you are building. What tech you want. What it should not include. What done looks like.
Paste that as your very first message. Nothing else. Then wait and see what the AI sets up before you give it any more instructions.
Compare that first session to the last time you opened a build without doing this. The difference will be obvious.
Reply and tell me what you built and what changed when you opened deliberately. I read every reply.
If you want a ready-made session opener template you can fill in and reuse, the VibeShip Vault has it along with the rest of the prompt library. And if someone you know keeps running into messy codebases they do not know how to untangle, send them here to join the community free.
See you next week.
Prathamesh
Founder, VibeShip