Vibe EngineeringSMIT · AI Agentic Engineering
Class 2 · 2.2

The Missing Manual

A powerful alien tool was handed to the whole profession — with no instructions. This course is the manual: the vocabulary, the mental models, and the map that make the chaos navigable.

🧭 Goal: build the vocabulary you’ll use for 8 weeksRead: 9 min

01The tweet that started it

A big part of the motivation for this course came from a post by Andrej Karpathy — a founding member of OpenAI, formerly of Tesla and Stanford, and the person who coined the term vibe coding.

A
Andrej Karpathy
@karpathy
𝕏
I’ve never felt this much behind as a programmer. The profession is being dramatically refactored… There’s this new layer of abstraction that involves agents, sub-agents, their prompts, context, memory, modes, permissions, tools, plugins, skills, hooks, MCP, LSP, slash commands, workflows, IDE integrations… Clearly some powerful alien tool was handed around, except it comes with no manual, and everyone has to figure out how to hold it and operate it while the resulting magnitude-9 earthquake is rocking the profession. Roll up your sleeves to not fall behind.

That tweet is this course. The promise: by the end, you can read that list back — agents, MCP, hooks, skills, Ralph Loops — and say “yes, I know what all of these mean, how they fit together, and where they work and where they don’t.” You’ll be able to separate the hype from the facts.

02Who this is for

Two archetypes — and honestly, everyone in between:

Aspiring engineers

New or fairly new to code

Maybe this is one of your first times writing code. You’ll learn to build complete, large-scale products with confidence, assisted by AI agents.

Senior / staff engineers

Experienced, but new to agentic coding

You can code, but the landscape feels overwhelming — too much going on. You’ll learn to connect the dots and accelerate without losing the joy of building.

Something for everyone

Some parts will be too advanced — get the gist and keep moving. Some parts (what’s a project directory?) will be obvious to experts — put it on 2× and skip ahead. Across the spectrum, there’s juicy stuff for you.

What you’ll be able to do

  • Build, debug, troubleshoot and enhance software at any scale — from a toy game to a large multi-repo team codebase — assisted by AI agents.
  • Take full advantage of today’s landscape, and “tweet back at Karpathy”: I get all this. I have the manual.
  • Meet the expectations now baked into job descriptions — MCP, skills, hooks, and bringing it together to ship quality software fast.
Enthusiast, not hype-merchant

These tools are incredibly powerful and they have real drawbacks. Success with agentic AI means knowing both the strengths and the weaknesses — maximizing one while managing the other. This course makes you aware of both.

03Vibe coding, vibe engineering, agentic coding

The terminology is genuinely muddled — the same words mean different things to different people. Here’s the map this course uses:

TermWho coined itWhat it points at
Vibe codingAndrej Karpathy“Give in to the vibes, embrace the exponentials, forget that code even exists.” Often means a quick, amateurish, throw-it-away-if-it-breaks approach — or just the whole field.
Vibe engineeringSimon WillisonThe professional side: building real software with the discipline of engineering, but with AI agents doing much of the typing.
Agentic coderemergingAn expert using agents (like the Cursor agent) as sidekicks — often several at once — to build a product together.
Careful: the words flip around

“Agentic coder” can mean someone who codes with agents (us) or someone who builds agents. Platforms like Cursor and Claude Code are themselves sometimes called “coding agents.” When in doubt, ask the person to clarify. On this course, an agentic coder is someone who codes with the help of AI agents.

The emotional arc of 2025 was a rollercoaster — surprise, denial, astonishment, frustration, even anger (people literally yelling at their agents) — then acceptance, and finally an inflection point in late 2025 (Claude Sonnet/Opus 4.5, then Gemini 3 and GPT-5.2) when agents got dramatically more reliable and a new wave of progress began.

04The three surfaces

There are three places you actually interact with coding agents — same brain, different cockpit:

Surface 1

IDE

A full app with windows, menus and panes. The agent lives beside your files and diffs. Best for staying in flow while you learn.

Cursor · Codex · Antigravity · Windsurf

Surface 2

Plugin / extension

Bolted onto an editor you already use — usually VS Code. Lower switching cost, narrower scope.

GitHub Copilot

Surface 3

CLI

The agent runs in your terminal — a retro “80s” black screen that turned out to be a wildly productive way to work. Scriptable, composable, closest to pro automation.

Claude Code · Cursor CLI · Codex CLI · Gemini CLI · OpenCode · AMP

Why the CLI matters

The command-line interface was almost an accident — a Claude Code side project that blew up. Stepping “back” to a retro terminal turned out to be one of the most powerful ways to work with agents. Weeks 2–3 of the course dig deep into it.

Signal vs. noise

This space moves absurdly fast — a tsunami of announcements every day. One of the hardest and most valuable skills is distinguishing true, durable value (like Claude Code’s release) from the talk-of-the-town that fizzles in a fortnight. Focus on the signal.

Key takeaways

  • The tools arrived without a manual — this course is the manual.
  • It’s for beginners and experts; take what’s useful, skim the rest.
  • Vibe coding → vibe engineering → agentic coding is a rough arc from casual to professional to expert.
  • Three surfaces: IDE, plugin, CLI. We start in the IDE and end in the CLI.
  • Be an informed enthusiast: hold the strengths and the weaknesses at once.