Step 1 — Why AI tools now

Early ChatGPT was a copy-paste chatbot. Late 2024 onward brought agents that read codebases, call tools, and edit dozens of files in one go.

1. 2023 vs 2025+

2023	2025+
Human asks → AI answers	AI edits and runs code directly
One file at a time	Dozens of files simultaneously
Terminal separate, AI separate	AI calls shell / DB / browser
Context = chat	Context = MCP / CLAUDE.md / memory
Static replies	Iteration, self-verification, tool chaining

2. Three IDE flavors

Agent CLIs — Claude Code, Aider, Codex CLI, Gemini CLI
Agent IDEs — Cursor, Windsurf, Zed
Browser / desktop agents — Comet, Atlas, Dia

This course centers on Claude Code (CLI); concepts transfer.

3. Model landscape (2026)

Vendor	Model	Strengths	Tier
Anthropic	Claude Opus 4 · Sonnet 4	code, long context, agents	Pro $20 · Max $200
OpenAI	GPT-5 · o3-pro	reasoning, multimodal	Plus $20 · Pro $200
Google	Gemini 2.5 · 3.0	multimodal, cheap, fast	free quota · Pro $20
Meta	Llama 4	open, local	free
xAI	Grok 4	realtime, X integration	paid

Claude is the coding leader; Gemini wins on volume; Llama wins on local autonomy.

4. First command

npm install -g @anthropic-ai/claude-code
cd my-project
claude
# or
claude -p "translate the README to English"

5. Learning curve — three shifts

CLAUDE.md

# Project rules

## Stack
- Next.js 16 · TypeScript strict · Tailwind v4

## Rules
- Only pnpm
- Korean commit messages
- Use logger.info, not console.log

Skills

Repeatable workflows invoked via slash commands.

MCP

Standard protocol to connect DBs / Slack / Figma / GitHub to the agent.

Later steps cover each in turn.

6. Real workflow

User: "add a comments feature"
↓
AI:
1. Read project structure
2. Analyze existing Post model / API
3. Write Comment schema + migration
4. Add service / repository
5. Add controller / DTO
6. Add frontend form + component
7. Write tests
8. Run, verify
↓
User reviews → requests tweaks → AI revises
↓
Commit (by user or AI)

7. Costs

Service	Monthly	Daily use
Claude Pro	$20	2–4 hours
Claude Max	$200	all-day
OpenAI Plus	$20	2–4 hours
Cursor Free	$0	limited
Cursor Pro	$20	all-day

8. Safety

Never paste API keys / secrets — system-prompt leaks possible
No production DB from AI — use read-only accounts
Don't blindly approve destructive commands (rm -rf, git push --force)
Verify package names (typosquatting)

9. Limits

Creative architecture — humans still
Deep domain knowledge — may miss post-training content
Rare libraries / cutting edge — hallucination risk
Security / performance — final judgment stays human

Treat AI as a fast junior; you stay senior.

10. Psychology

The hardest shift is ego. "I didn't write that."

Shift into code-review mindset
Reading, understanding, revising is the skill
Judgment, direction, verification are where value lives

11. Why now

2026 models are stable
Tool ecosystem (MCP, Skills) is mature
Productivity gap opens up at multi-x levels
Foundation + AI beats either alone

Deeper

Step 2 — spin up Claude Code and run your first command.