How Vibe Coding Works Here

You describe what you want in ordinary sentences. The agent writes the code, runs the tools and reports back. You never type a command or read code — every page gives you something to say, then tells you what you should see.

You do not need to understand the output

Almost none of the text agents print is for you. Look only for what this guide points at — usually a web address, a file path, or a window opening.

The one habit worth learning

Ask for a result you can see, not for an action.

Instead of saying	Say
"Set up the project"	"Set up the project, then tell me the web address I can open"
"Add a products page"	"Add a products page, then list every file you changed"
"Build the app"	"Build the app, then tell me exactly where the installer file is"
"Fix it"	"It shows an error. Read the error, tell me the cause, then fix it"

The address, path or file list is what lets you check the work. An agent that says "done" and nothing else has given you nothing to check.

Platform-specific examples

Instead of	Say
"Set up a project"	"Create a project containing only the website-based management system"
"Fix the build"	"The docs build fails. Read the error, tell me the cause, then fix it"
"Add a table"	"Add an article list page, following the conventions in apps/web-admin/AGENTS.md"
"Make it work"	"Start everything the management system needs, the standard way, and tell me when the page is ready to open"
"Set up a desktop project"	"Create a project with only the desktop app — answer the selection question and pick Desktop"
"Start the app"	"Start the desktop app and tell me when the window is open"
"Add a products page"	"Add a product screen, following the step-by-step workflow in apps/desktop/AGENTS.md"
"Build it"	"Build the release version — I need an installer file, tell me its exact path"
"Set up an Android project"	"Create a project with only the Android app — answer the selection question and pick Android"
"Put it on my phone"	"Install the app on the phone I have plugged in, and tell me when it is on there"

Same pattern every time: say what you want, name the constraint, point at the rules file, and ask for something you can look at.

Your first project, start to finish

Say this

I want to create a new project called admin-platform. I only need the website-based management system — no phone app, no desktop app. Set it up and tell me what you created.

What you should see: a short confirmation of the two pieces it is building.

✅ Selected 2 modules:
   - Web Admin
   - API Service

Those two are the default combination. Follow up once — "which pieces did you actually create?" — and the answer should be exactly those two. Anything extra means it built the wrong thing; restarting now is much cheaper than later.

Say this

Now install everything this project needs and start it running. Tell me the web address to open in my browser.

What you should see: a web address, usually starting http://localhost:. Open it — that is your project running on your own machine.

Your project's history is saved once automatically before you change anything, so if a later experiment goes wrong you can always ask the agent to restore it.

Asking for a different combination

The default pair is the backend plus the admin website. Anything else makes the tool stop and ask which pieces you want — the agent must answer that question, not skip it. You only have to name the pieces:

Say this

I only want the desktop app — nothing else. The tool will ask you which pieces to include. Read its list and pick the line labelled Desktop by its label, not by guessing a number.

What you should see: a confirmation naming only the piece you asked for.

✅ Selected 1 modules:
   - Desktop

Say the name, never a number

The numbers in the printed list move between versions, so "pick 6" eventually picks the wrong thing — always name the label. Entries marked [developing] are unfinished and get quietly ignored if chosen.

Things you can ask for

Say this	What you get
"Start just the admin website"	the website opens, but every page errors until the whole system is running
"Start the whole system"	everything the website needs is up, and the login page opens
"Build everything for release"	finished files you could hand to someone else to install
"Update my project to the latest template"	the shared template files refresh; your own work is left alone
"Check what my computer is missing, then install it"	a plain list of what is missing, then the installs
"Start everything the system needs"	its data store and cache are started the standard way, with nothing for you to choose

Two rules to repeat in your prompts

Agents get these two wrong more than anything else. Paste this whenever the agent is about to write documents or save its work:

Before you start: read the AGENTS.md file nearest to whatever you are changing
and follow it. Notes and plans for me go in .docs/. Pages for the public website
go in docs/content/ and must also be added to docs/sidebars.ts. Commit messages
are in English.

When something looks wrong

You do not need to diagnose it. Hand the problem back:

What you notice	Say this
It says "done" but you have nothing to open	"Tell me the exact web address, and the file paths you created"
It seems frozen with no new output	"Are you waiting on something that never finishes? Start it the way that lets you keep talking to me, and tell me when it is up"
An error you cannot read	"Read that error, explain the cause in plain language, then fix it"
It changed more than you expected	"List every file you changed and why. Undo anything I did not ask for"
You are not sure it did the right thing	"Show me your plan before you change anything else"

Ask for the plan first

A wrong plan takes one sentence to correct. A wrong pile of code takes an afternoon. For anything bigger than a small tweak, say "show me your plan first" and read it before agreeing.

Next

Start a Project — pick what you want to build and get set up in one go.