Week 2: How AI Coding Actually Works

"The hottest new programming language is English." — Andrej Karpathy

---

Quick Pulse Check (Start)

Before we begin, rate your confidence (1-5):

1. How well do you understand what happens when an AI generates code?
2. Have you experimented with different AI coding tools?

---

Learning Objectives

By the end of this week, you will:

• Understand how LLMs generate code (tokenization, context windows, next-token prediction)
• Recognize why AI hallucinations happen and how to prevent them
• Apply prompt engineering techniques to maximize AI effectiveness
• Compare different AI coding tools and select the right one for your task

---

LLM Fundamentals for Engineers

Tokenization: The Foundation

A token is not a word—it's a subword unit. Models process and charge by tokens.

"function" → 1 token
"XMLHttpRequest" → 4 tokens (XML / Http / Request)
"const myVariable = 42;" → 6 tokens

Why this matters:

• Cost scaling: Verbose prompts = higher cost, fewer tokens left for generation
• Context efficiency: Be precise to maximize useful output
• Model differences: Different models tokenize differently

Context Windows: AI's Memory

Model	Context Window
Claude 3.5 Sonnet	200K tokens
Gemini 2.5 Pro	1M tokens
GPT-4	128K tokens

Practical insight: By message 50 in a conversation, the AI begins "forgetting" earlier exchanges. Start fresh sessions for unrelated tasks.

Temperature: Controlling Randomness

Temperature 0.0  → Deterministic (same output every time)
Temperature 0.7  → Moderate randomness (default)
Temperature 1.5  → High creativity (unpredictable)

When to use:

• Code generation: Temperature 0.0–0.3 (reproducibility)
• Brainstorming: Temperature 0.7–1.0 (creativity)

Why Hallucinations Happen

The model has no "doubt" mechanism—it's always confident. It predicts the next token based on patterns, not understanding:

// AI might generate this with high confidence:
function queryDB(id) {
  return DATABASE.query(`SELECT * FROM users WHERE id = ${id}`)
    .then(users => users[0].encryptionKey) // ← Hallucination!
}

Always validate AI-generated code, especially for:

• API calls (made-up endpoints)
• Database queries (non-existent columns)
• Library function names (close but wrong)

---

The AI Coding Tool Landscape

Tool Categories

Category	Tools	Best For
IDE-Native	Cursor, Windsurf, GitHub Copilot	Integrated development, tab completion
Terminal-Native	Gemini CLI, Claude Code, aider	Server-side work, scripting, automation
Rapid Prototyping	Bolt.new, v0, Lovable, Replit	Full-stack prototypes in minutes

Free Options for Students

Tool	Free Tier
Gemini CLI	Free tier (check current quotas)
GitHub Copilot	Free with GitHub Student Pack
Cursor Pro	1 year free for students
Bolt.new	300K daily tokens free

The principle: Tools change every 3 months. Master the discipline, not the tool.

---

Prompt Engineering for Code

System Prompts vs. User Prompts

System Prompt (persistent context):

You are a Senior Software Engineer specializing in TypeScript.
You prioritize: Type safety, error handling, test coverage >85%.
Non-Goals: Don't optimize prematurely.

User Prompt (specific request):

Implement a function that fetches user data and caches the result in Redis for 10 minutes.

The Specificity Spectrum

Poor:

Build a todo app

Excellent:

Create a React 19 todo app with:

SCOPE:
- Task CRUD operations
- Persistence: localStorage only
- UI: Tailwind CSS, responsive to 320px

NON-GOALS:
- Server synchronization
- User authentication

CONSTRAINTS:
- React Hooks only
- Zero external dependencies except React & Tailwind

Few-Shot Examples

Show the AI examples of code patterns you want:

// Style I want:
const [formData, setFormData] = useState({ email: '', password: '' });

const validate = () => {
  const newErrors = {};
  if (!formData.email.includes('@')) newErrors.email = 'Invalid email';
  return Object.keys(newErrors).length === 0;
};

// Now generate a similar hook for...

Context Engineering

Prompt engineering = "How do I phrase my request?" Context engineering = "How do I organize ALL information so the AI can't misunderstand?"

Provide:

1. Project structure
2. Existing code patterns
3. Test examples
4. Then ask for the specific change

---

AI Tool Examples

Exploring a Codebase

Gemini CLI:

gemini "Read all files in src/ and explain the architecture"

Claude Code:

claude "What's the structure of this codebase? List main components."

Generating Code with Constraints

Gemini CLI:

gemini "Generate a login form in React with email validation. Use Tailwind CSS. No external libraries."

Claude Code:

claude "Create an Express endpoint for user registration with bcrypt password hashing and input validation."

---

Where AI Excels vs. Struggles

Strengths (Pattern-matching)	Weaknesses (Reasoning)
Boilerplate code	Novel algorithms
Refactoring existing code	Complex math proofs
Generating test cases	Long chains of logic
Documentation from code	Security vulnerability discovery
Fixing syntax errors	Performance optimization

Rule of thumb: AI is a search-by-pattern tool, not a reasoning engine. Give it patterns to follow.

---

Quick Pulse Check (End)

Reflect on today's session:

1. What's one thing about how AI generates code that surprised you?
2. Which AI coding tool do you want to try first?

---

Resources

Fundamental Papers

• "Attention Is All You Need" (Vaswani et al., 2017) — The transformer architecture
• "Language Models are Few-Shot Learners" (Brown et al., 2020) — Why prompting works

Prompt Engineering

• Anthropic's Prompt Engineering Guide
• OpenAI's Prompt Engineering Best Practices

Tool Documentation

• Gemini CLI Documentation
• Claude Code Documentation
• Cursor Documentation

---

Instructor: Goker Ezberci | gokerez@gmail.com