A Confession

- Emre Varol, 8+ years at Google, Palantir, Liftoff

What Does That Mean?

It does not mean I stopped thinking

It does not mean I blindly accept output

It does not mean engineering skills don’t matter

It means I describe behavior instead of typing syntax

It means I review outcomes instead of reviewing every line

It means what used to take a dev team 3 weeks, AI now does in 15 minutes

The Old Way vs The New Way

The World’s #2 Coder Works for You

In December 2025, AI models placed 2nd in the world’s top competitive programming competitions.

Better than 99.99% of human programmers

Solves complex algorithmic problems in seconds

Understands every language, framework, and pattern

Has perfect memory - never forgets a single detail

Available 24/7, infinitely patient, never gets tired

“Vibe Coding”

Coined by Andrej Karpathy (OpenAI co-founder, ex-Tesla AI lead) in February 2025. Named Collins Dictionary Word of the Year 2025.

In 2026, it has evolved into Agentic Engineering:

Write natural-language specs

AI generates code under structured human oversight

Multi-model orchestration with persistent project context

The developer specifies, reviews, and steers - AI implements

Meet the Tools

These Are Not Autocomplete

The 2000x Claim

Let’s do the math with real numbers:

A Real Example

Without AI: 2–3 weeks. With AI: 4 hours.

What Changed?

LLMs have perfect memory and infinite knowledge.

They were trained on virtually all public code ever written - and they memorized it

Context windows are now 1M+ tokens - they remember your entire conversation

They read your entire project before writing a single line

They understand your naming conventions, architecture, and style

They know every framework, library, and best practice ever published

This is not a search engine. This is a world-class engineer with perfect memory who has read every codebase on GitHub and never forgets a single detail.

Your Role Has Changed

Key Insight

Prerequisites

Node.js 18+ - required for Claude Code and most modern tools

Git - version control is non-negotiable

A terminal - macOS Terminal, iTerm2, Windows Terminal, or any Linux terminal

A text editor - optional, for when you want to read code (VS Code, Vim, etc.)

# Check your versions
$ node --version    # v20.19.0 or higher
$ git --version     # any recent version
$ python3 --version # 3.10+ (optional, for Codex)

Installing Claude Code

# Install globally
$ npm install -g @anthropic-ai/claude-code

# Navigate to your project
$ cd my-project

# Start Claude Code
$ claude

# That's it. Claude reads your entire codebase and is ready.

Reads your entire project automatically

Can edit files, run commands, create new files

Remembers context across the conversation

Works with any language, any framework

Installing Codex CLI

# Install globally
$ npm install -g @openai/codex

# Set your API key
$ export OPENAI_API_KEY="your-key-here"

# Start Codex in your project
$ codex

# Open-source, runs locally, sandboxed execution

Open-source (MIT license)

Sandboxed - runs in a safe environment

Multiple modes: suggest, auto-edit, full-auto

Installing Gemini CLI

# Install globally
$ npm install -g @google/gemini-cli

# Authenticate with Google
$ gemini auth login

# Start Gemini in your project
$ gemini

# Multimodal: can analyze images, screenshots, and diagrams too

Your First Command: Build Your Website

$ mkdir my-portfolio
$ cd my-portfolio
$ git init
$ claude

# Now type (paste your CV text first, then):
> Here is my CV: [paste your CV]. Build me a beautiful personal
  portfolio website like emrevarol.com. Include a hero section with
  my name and title, an about section, my experience timeline,
  skills, education, and a contact section. Use a dark theme with
  modern animations. Make it responsive.

# Watch it create your personal website in under a minute.

The Terminal Is Your IDE

You do not need VS Code, IntelliJ, or any editor.

Claude Code reads, writes, edits, and creates files

Claude Code runs your dev server, tests, and builds

Claude Code commits, pushes, and creates PRs

Claude Code deploys to Vercel, Netlify, or any platform

Of course, you can use an editor alongside it. But you don’t have to.

CLAUDE.md: Your Project’s Memory

Create a CLAUDE.md file in your project root. This is your AI’s persistent memory.

# CLAUDE.md

## Project: Bean There Coffee Shop
- Tech stack: Next.js 16, Prisma, PostgreSQL
- Style: Tailwind CSS, warm earth tones
- Language: TypeScript
- Database: Neon PostgreSQL (connection string in .env)

## Rules
- Always use server components by default
- Use Turkish lira (₺) for prices
- Mobile-first responsive design
- Never hardcode secrets

Your Stack Beyond Code

Building a product is not just code. You need infrastructure. Here’s what we use - all set up from the terminal:

Behavior-Driven, Not Code-Driven

The Iteration Cycle

This is your new workflow. Memorize it.

Step 1: Start Small

Begin with the simplest possible version.

> Build a complete English
learning platform with AI tutor,
assessments, writing/speaking/
listening modules, vocabulary,
grammar, leaderboards, friends,
18 languages, and auth.

> Create a Next.js app with a
landing page for an English
learning platform. Show 4
learning modules (Writing,
Speaking, Listening, Vocabulary)
as cards. Hardcoded for now.

Step 2: Add One Feature

Once the base works, add exactly one feature.

# Iteration 2: Add a database
> Replace the hardcoded data with a PostgreSQL database using Prisma.
  Create Lesson, Module, and Progress models. Seed sample lessons
  for each module.

# Iteration 3: Add user auth
> Add NextAuth with email/password. Users should be able to sign up
  and log in. Show their name in the navbar when logged in.

# Iteration 4: Add AI tutor
> Add a /tutor page with a conversation interface. Users type a
  message, Claude API responds with English corrections and tips.
  Track conversation history.

Step 3: Test the Behavior

After each feature, test the behavior, not the code.

Open the app in your browser

Click through the feature like a user would

Try edge cases: empty inputs, wrong data, double clicks

Check mobile view

You do NOT need to:

Read every line of generated code

Understand every helper function

Review the CSS class names

Step 4: Modify and Iterate

Something not right? Tell the AI what to change - in behavior terms.

# Not: "change the onClick handler to use a different state variable"
# Instead:
> When I finish a lesson, the progress bar doesn't update until I
  refresh the page. Fix this so the progress updates immediately.

# Not: "add a try-catch around the API call"
# Instead:
> When the AI tutor API call fails, the chat just freezes. Show a
  red error message that says "Connection lost, please try again"
  and re-enable the send button.

Step 5: Bake It → Move On

A feature is “baked” when:

It works as expected in all cases you can think of

It handles errors gracefully

It looks good on desktop and mobile

You’ve tried to break it and couldn’t

Once baked: commit, and move to the next feature.

> Commit all changes with message "Add AI tutor with conversation history"

Don’t Review Code. Review Behavior.

This is the hardest mindset shift for experienced engineers.

Trust the Machine

A hard truth from my experience:

It never forgets to handle edge cases

It follows conventions perfectly

It writes consistent patterns across files

It never gets lazy or takes shortcuts

It knows every best practice from every framework

This doesn’t mean it’s always right. It means your default should be trust, then verify - not distrust, then rewrite.

Your New Role: Architect + PM + QA

You are now three roles in one:

The Prompt Is Your Specification

A good prompt is a good spec. A bad prompt is a bad spec.

> Add a dashboard

> Add a dashboard page at /dashboard.
  Show the user's learning streak,
  completed lessons, vocabulary
  mastery %, and recent activity
  (last 10 sessions). Include a
  chart showing progress over time.

Incremental Over Monolithic

Philosophy Summary

Case Study 1: emrevarol.com

Built in: ~4 hours   Traditional: 2–3 weeks with a web developer   Note: I’m a backend & ML engineer - I don’t know web development!

Case Study 2: evimibul.org

Stack: Next.js 16 + Prisma + Neon PostgreSQL + Claude API + Cloudinary + Bright Data

evimibul.org: Conversation-First UX

Instead of filters and forms, users just talk:

User: I'm looking for a 3-bedroom apartment in Çankaya,
      max 15,000 TL/month, with a balcony

AI:   I found 23 listings matching your criteria. Here are
      the top 5 based on value score:

      1. 3+1, Kızılay, 12,500 TL - Score: 92/100
         ✓ Balcony  ✓ Central heating  ✓ Near metro
      ...

Case Study 3: english.cash

Stack: Next.js 16 + Prisma + Neon PostgreSQL + next-intl

Case Study 4: pusulam.ai

Complexity: Financial math, real-time pricing, payment integration - all built with AI

Case Study 5: deneme.live

Stack: Next.js 16 + Prisma + SQLite   Domain-specific: Turkish education system rules, net calculation

Case Study 6: neyesek.biz

Stack: Next.js 16 + Prisma + PostgreSQL   Domain-specific: Food/restaurant data, review aggregation, Ankara-focused

Live Demo: a2sv.org

How Many Prompts? How Much Time?

I built all these projects in parallel over ~3 weeks. Here are the estimated prompt counts:

Common Patterns Across All 7 Projects

Every project started with a single claude command

Every project used incremental, behavior-driven development

Every project achieved multi-language support via AI translation

Every project was deployed to production by AI (Vercel, GitHub)

Every project uses the same modern stack (Next.js, Prisma, PostgreSQL/SQLite)

No project required a team - one person built each one

Time Comparison

The Scope Multiplier

Things I would never have built manually:

10–18 language support with full UI translation across all projects

AI-powered conversational search with voice input

CPMM financial math engine for prediction markets

150,000+ listing web scraper with deduplication

Interactive slide presentations with 90+ slides

AI exam generation with Turkish net scoring

Real-time AI tutoring with adaptive difficulty

Case Studies Summary

Parallel Agents with GenAI

Need to translate to 9 languages? Don’t do it one by one.

# Claude Code can spawn sub-agents that work in parallel
> Translate the lecture page to all 9 supported languages
  (TR, FR, DE, ES, AR, PT, RW, SW, AM). Use parallel agents
  for each language.

# Result: 9 agents work simultaneously
# 36 files translated in ~5 minutes instead of ~45 minutes

Context Is Everything

The AI is only as good as the context you give it.

CLAUDE.md - project-level instructions, tech stack, rules

Memory files - persistent knowledge across conversations

Clear prompts - specific, behavior-focused descriptions

Existing code - the AI reads your codebase and follows your patterns

# Bad context: starting from scratch every time
> make a page

# Good context: AI knows your project inside and out
> Add a /friends page similar to /leaderboard. Show mutual friends
  at the top, then suggested friends based on similar learning paths.
  Use the existing FriendRequest model.

When AI Gets Stuck

It happens. Here’s what to do:

Simplify the prompt - break it into smaller pieces

Give examples - show what you want, not just describe it

Share error messages - paste the exact error, AI is great at debugging

Try a different approach - “instead of X, try Y”

Start fresh - sometimes a new conversation with clear context works better

Debugging with AI

AI is an incredible debugger.

# Just paste the error and describe the context
> I'm getting this error when I click the Buy button:
  "TypeError: Cannot read properties of undefined (reading 'price')"
  This happens only when the market has no trades yet.

# The AI will:
# 1. Find the relevant file
# 2. Identify the bug
# 3. Fix it
# 4. Explain what went wrong

Multi-File Changes

One of AI’s biggest advantages: it edits many files atomically.

> Add a "favorites" feature. Users can bookmark listings.
  Show a heart icon on each listing card, a /favorites page
  with saved listings, and a count badge in the navbar.

# AI will modify in one go:
# - schema.prisma (new Favorite model)
# - API route (CRUD for favorites)
# - ListingCard component (heart button)
# - Navbar (favorites count)
# - New /favorites page
# - Database migration

What NOT to Do with AI

Don’t paste AI-generated code into Stack Overflow answers (it’s not yours)

Don’t skip testing because “the AI wrote it”

Don’t use AI for security-critical code without expert review

Don’t stop learning fundamentals - you need to understand what you’re building

Don’t give up after one failed attempt - rephrase and try again

Don’t try to build everything in one giant prompt

Production Safety (1/2)

Real incidents from building these projects with AI.

1. Always Have Backups

git push --force destroyed a live site (Impact page, Press page, GenAI lecture — gone)

Recovery was only possible because Vercel kept old deployments and a second clone existed

Rule: Never force push. Never rebase shared branches. Always have a rollback strategy.

2. Guard Your API Keys

AI models can autonomously use API keys and rack up thousands of dollars without you realizing

Rule: Set spending limits. Use separate keys per project. Monitor usage daily.

Production Safety (2/2)

3. Log Everything & Generate Reports

You need daily activity reports and finance reports — know what your AI agents did and what they spent

Use services like Resend to email yourself automated daily summaries

Rule: If you can’t see what happened yesterday, you’re flying blind.

The Human’s Irreplaceable Role

AI cannot replace you in these areas:

Vision - What should we build? Why? (for now)

Empathy - What do users actually need? (for now)

Judgment - Is this the right approach? (for now)

Taste - Does this feel right? Is it good enough? (for now)

The Future Is Already Here

What’s coming next:

AI agents that deploy, monitor, and fix production issues autonomously

Multi-agent systems where AIs collaborate on large projects

AI that designs UX, writes copy, and creates assets

Voice-first development: “Hey Claude, add a search bar to the header”

AI that learns your preferences and anticipates your needs

Key Takeaways

1. The manual coding era is over - embrace AI-first development

2. Think in behaviors, not code - describe what, not how

3. Start small, iterate incrementally - one feature at a time

4. Test behavior, not code - if it works correctly, it is correct

5. Trust the machine - verify outcomes, don’t rewrite implementations

6. Your value is in vision, judgment, and taste - not syntax

7. CLAUDE.md is your project’s brain - invest in good context

8. Never stop learning fundamentals - AI amplifies knowledge, not ignorance

9. Guard your API keys - AI models don’t care about your bill, set limits and monitor daily

10. Always have backups - never force push, log everything, generate daily reports

Your Homework

Build something real this week.

Resources

Claude Code: npm install -g @anthropic-ai/claude-code

Codex CLI: npm install -g @openai/codex

This lecture: emrevarol.com/lecture/coding-with-genai.html

All lectures: emrevarol.com/lecture/

Live Projects (See Them in Action)

emrevarol.com - Portfolio & lectures

evimibul.org - AI real estate

english.cash - AI English learning

pusulam.ai - Prediction markets

deneme.live - AI exam platform

neyesek.biz - AI food reviews

a2sv.org - Africa to Silicon Valley

Final Words