Vibe Coding

Vibe Coding

Overview

Vibe coding is building software primarily through natural language prompts to AI coding agents. Done well, it dramatically accelerates development — the AI writes boilerplate, implements features, fixes bugs, and handles routine tasks while you focus on decisions, architecture, and quality. Done poorly, it produces a mess of half-working code that's hard to reason about.

The difference between productive vibe coding and chaos comes down to structure, context, and iteration.

Instructions

The core loop

Spec → Generate → Review → Test → Iterate

1. Spec: Write a clear description of what you want
2. Generate: Let the AI produce the code
3. Review: Read and understand what was generated
4. Test: Run it and verify it works
5. Iterate: Refine with follow-up prompts

Never skip Review. You own the code. The AI is the keyboard, you're the brain.

Step 1: Structure your project for AI

AI agents work best when they can understand your project at a glance. Key files:

CLAUDE.md / AGENTS.md — Project context for the AI agent:

# Project: MyApp

## What this is
A SaaS app for freelancers to track invoices and clients.

## Stack
- Next.js 15, TypeScript, Tailwind CSS, shadcn/ui
- Prisma + PostgreSQL (Supabase)
- Auth.js with GitHub OAuth
- Stripe for billing

## Architecture
- `app/` — Next.js App Router pages
- `components/` — Reusable UI components
- `lib/` — Business logic and DB queries
- `lib/db/` — All Prisma queries (never query DB from components)

## Key conventions
- Named exports everywhere
- Zod validation on all API inputs
- React Query for client-side data fetching
- Error handling: throw in lib/, catch and return proper HTTP responses in API routes

## Commands
- `npm run dev` — Start development server
- `npm run test` — Run Vitest tests
- `npm run db:migrate` — Run Prisma migrations
- `npm run typecheck` — TypeScript check without building

Why this matters: Without CLAUDE.md, the AI guesses your stack and conventions. With it, every prompt inherits your context — no need to repeat "we use TypeScript with strict mode" every time.

Step 2: Write effective prompts

The anatomy of a good prompt:

[Context] + [What you want] + [Constraints] + [Examples if needed]

Weak prompt:

Add a user settings page

Strong prompt:

Create a user settings page at app/settings/page.tsx.
It should let users update: display name, email, avatar URL.
Use the existing UserForm pattern from app/profile/page.tsx.
Fetch current user with getServerSession, update via PATCH /api/user.
Add a Zod schema for the update payload.
Show a toast on success using the sonner setup in lib/toast.ts.

The difference: the strong prompt gives the AI enough to produce code that fits your project without guessing.

Step 3: Context management

AI coding tools have context windows. Use them wisely:

Keep files small and focused — A 50-line file that does one thing beats a 500-line file. The AI can read the whole thing, understand it, and modify it correctly.

Name things descriptively — getUserInvoicesWithLineItems.ts tells the AI more than queries.ts.

Use barrel files — lib/db/index.ts that re-exports everything lets the AI import without hunting:

// lib/db/index.ts
export * from "./users";
export * from "./invoices";
export * from "./clients";

Reference specific files in prompts:

Looking at lib/db/users.ts, add a similar pattern for clients.
Use the same error handling and Zod validation approach.

Step 4: The spec-first workflow

For features bigger than a few lines, write a spec before coding:

## Feature: Invoice PDF Export

### What
User clicks "Export PDF" on an invoice → receives a download of a formatted PDF

### How
1. Frontend: Add "Export PDF" button to InvoiceDetail component
2. API: POST /api/invoices/:id/export → returns a PDF blob
3. Generation: Use @react-pdf/renderer to build the PDF template
4. The PDF should include: company logo, invoice items, totals, payment instructions

### Edge cases
- Draft invoices can be exported (show DRAFT watermark)
- Long item descriptions should wrap, not overflow
- If company logo URL is broken, show placeholder

### Template
Mirror the layout of the existing invoice preview at components/InvoicePreview.tsx

Hand this spec to the AI: "Implement this feature. Start with the API route, then the PDF template, then the frontend button."

Step 5: Iterative refinement

Generate in chunks, not all at once:

Iteration 1: "Create the database schema and Prisma model for invoices"
Iteration 2: "Add the CRUD API routes for invoices using the schema we just created"
Iteration 3: "Build the invoice list page that fetches from the API"
Iteration 4: "Add the invoice detail page with edit functionality"
Iteration 5: "Write tests for the API routes"

Each iteration is reviewable, testable, and committable. This beats "build me a full invoice system" which produces 2000 lines you need to review all at once.

Step 6: Review patterns

When the AI produces code, check:

1. Does it compile? Run tsc --noEmit immediately
2. Does it follow your patterns? Compare to existing similar code
3. Are there obvious security issues? Input validation, auth checks
4. Does it handle errors? What happens when the database is down?
5. Does it do what you asked? AI sometimes solves a slightly different problem

If something looks wrong, tell the AI specifically:

The getUserById function doesn't handle the case where the user doesn't exist.
It should return null, not throw. Fix that.

Step 7: When to code manually

Vibe coding isn't always the right tool:

Use AI for:

• Boilerplate (CRUD routes, forms, migration files)
• Implementing well-understood patterns
• Refactoring and renaming
• Writing tests for existing logic
• Debugging with error messages

Write manually when:

• The algorithm requires careful thought (complex business logic, performance-critical code)
• Security is critical (auth flows, cryptography)
• You're prototyping and exploring the problem space
• The AI keeps getting it wrong (you understand it better than it does)

Step 8: Git hygiene with AI coding

Commit small, commit often. When the AI generates a bunch of code:

# Review changes before committing
git diff

# Commit in logical chunks
git add lib/db/invoices.ts
git commit -m "feat: add invoice DB queries"

git add app/api/invoices/route.ts
git commit -m "feat: add invoice CRUD API"

This keeps your history readable and makes it easy to revert if the AI went off track.

Examples

Example 1: Starting a new project with Claude Code

# 1. Create project
npx create-next-app@latest myapp --typescript --tailwind --app

# 2. Write CLAUDE.md
cat > CLAUDE.md << 'EOF'
# MyApp
Next.js 15 + TypeScript + Prisma + Supabase + Auth.js + Stripe
[... stack and conventions ...]
EOF

# 3. Start Claude Code
claude

# 4. Prompt:
"Read CLAUDE.md. Set up the project:
1. Install Prisma and configure with Supabase connection string from .env
2. Create the initial User and Account models for Auth.js
3. Add the Auth.js configuration with GitHub OAuth
4. Create a basic layout with header and sidebar using shadcn/ui"

Example 2: Debugging with context

I'm getting this error:
PrismaClientKnownRequestError: Foreign key constraint failed on field: invoice_id

It happens when I call deleteClient() in lib/db/clients.ts.
The client has existing invoices. I need to either:
a) Block deletion if invoices exist (return an error)
b) Cascade delete the invoices too

Looking at the business rules in CLAUDE.md, what approach makes sense,
and how should I implement it?

Example 3: Refactoring session

The components/forms/ directory has grown to 15 files and they're all doing
their own validation and submission logic differently.

Please:
1. Identify the common patterns across these form components
2. Extract a useFormSubmit hook that handles: loading state, error handling, success toast
3. Refactor UserForm.tsx and InvoiceForm.tsx to use it as a demonstration
4. I'll refactor the remaining forms myself following that pattern

Guidelines

• Read what the AI writes — you're responsible for the code, not the AI
• Commit before major changes — easy to revert if the AI goes sideways
• One thing at a time — smaller tasks produce better results than "build the whole feature"
• Invest in CLAUDE.md/AGENTS.md — 30 minutes writing context saves hours of prompt repetition
• Small files beat large files — better AI comprehension, easier diffs, more modular
• Test after every generation — catch problems early before they compound
• Tell the AI your constraints — token budgets, existing patterns, what NOT to change
• When stuck, be more specific — vague prompts produce vague code