Module 0: The Mental Model Shift
Claude Code Is an Agent, Not a Chatbot
The single most important thing to understand: Claude Code operates as an autonomous agent in your terminal. It doesn't wait for you to paste code. It reads your files, navigates your project, runs commands, writes code, verifies its work, and iterates, all within a single conversation turn.
This distinction matters because it changes how you interact with it. You stop asking questions and start delegating tasks.
The #1 Mistake Experienced Devs Make
Treating Claude Code like a chatbot: pasting code snippets, asking "what does this do?", manually copying suggestions back into files. This works, but you're using maybe 10% of its capability.
Instead, give it a task with a clear outcome: "Refactor the authentication middleware to use JWT instead of session cookies. Update the tests." Then let it work. It will read the relevant files, make the changes, run the tests, and fix what breaks.
The Agentic Loop
Every Claude Code interaction follows the same cycle:
- Observe: Read files, search code, understand current state
- Plan: Determine what needs to change and in what order
- Act: Edit files, run commands, create new files
- Verify: Run tests, check builds, validate output
- Repeat: Fix issues found in verification, iterate until done
This loop can run multiple times within a single prompt. A refactoring task might involve reading 15 files, editing 8, running the test suite twice, and fixing two regressions before returning a summary to you.
How It Compares
| ChatGPT | Copilot / Cursor | Claude Code | |
|---|---|---|---|
| Interface | Browser chat | IDE panel + inline | Terminal CLI |
| Context | Paste what you want it to see | Open files + workspace indexing | Entire codebase (reads on demand) |
| Actions | Generates text | Chat, inline edits, agent modes | Reads, writes, runs shell commands |
| Workflow | Single Q&A turns | Chat or inline suggestions | Multi-step agentic plans |
| Tool use | Limited (web, code interpreter) | IDE-scoped (editor, terminal) | Full shell, git, MCP plugins |
Other tools have agent modes too. Cursor has its agent, Copilot has workspace agents. The difference is that Claude Code is CLI-native with full repo context by default. It doesn't need an IDE open. It doesn't index your workspace upfront. It reads what it needs, when it needs it, using the same tools you'd use: grep, find, git log, your test runner, your build system.
When Claude Code Shines
- Changes to existing projects: it reads surrounding code, matches conventions, updates related files
- Understanding unfamiliar codebases: "explain the request lifecycle in this app" while it actually reads the code
- Multi-file refactoring: rename a concept across models, controllers, tests, and docs
- Automating repetitive tasks: "add error handling to all API endpoints following the pattern in
users.ts" - CI/CD integration: headless mode (
-p) for code review, test generation, migration scripts in pipelines - Git workflows: branch creation, commit messages, PR descriptions with full diff awareness
When Other Tools Are Better
- Quick Q&A about concepts: use claude.ai or ChatGPT, no need for an agent loop
- Vibe-coding a new UI from scratch: tools like Lovable or v0 give you visual iteration speed
- Real-time inline suggestions while typing: Copilot's ghost text while you code is a different (complementary) workflow
- Debugging a running application: Claude Code can't attach to your dev server or inspect runtime state
Limitations to Know
- Hallucination: it will confidently generate code that uses APIs that don't exist, or reference files with wrong paths. Always verify.
- Context decay in long sessions: after many turns, earlier context fades. Start fresh sessions for new tasks.
- Scope creep: ask it to fix a bug, it might refactor the whole module. Be specific about boundaries.
- No built-in browser: it can start a server and make HTTP requests via
curl, but it can't visually inspect a running UI by default. With MCP tools like Playwright, it can interact with web pages including localhost. Without those tools, it reasons about code statically. - Cost per interaction: every turn consumes tokens. Long agent loops on large codebases add up, whether you're on a subscription or API billing.
AI Doesn't Write Worse Code, It Exposes Process Gaps
If the code quality from AI-assisted development seems lower than expected, the issue is usually not the AI but the lack of verification. Teams without proper testing, code review, and linting processes will notice this more. AI tools amplify what's already there: good processes get faster, weak processes get exposed. This course teaches the verification discipline that makes AI assistance reliable.
Language-Agnostic by Design
Everything in this course applies regardless of your stack. Claude Code works the same way whether your project is TypeScript, Python, Rust, Go, C#, or a monorepo mixing several of them. It reads files, runs your tools, and writes code. The agentic loop doesn't care about the language.