Module 0: The Map
By the end of this module you will know where the 21 skills come from, why AI software is built differently from every other kind of software, and what you will build across the rest of this course.
Where the 21 Skills Come From
In August 2026, Andrew Ng published the AI Engineering Skills Map. His team built it from more than 10,000 job postings, dozens of structured interviews with AI experts, hiring managers, and recruiters, and survey data on top of that. The method was closer to clustering a large dataset of real jobs than to a panel of opinions.
The top level has four areas. Building and deploying AI applications. Software engineering fundamentals. Using coding agents. Shaping the build. Ng expanded each of the four into four to seven sub-skills in the letters that followed, and this course covers all 21.
One point from the first letter changes how you should read everything after it. Ng writes about AI engineering skills rather than the AI engineer job title, and he does it on purpose. His comparison is the cloud. Every developer works with the cloud today and almost nobody carries a cloud engineer title. He expects the same here: full-stack, data, DevOps, and ML engineers will all need these skills. So this course is for you even if your title never changes.
Where a lesson needs something concrete, this course uses Claude Code, Python, and whichever model API you already have access to. Every principle transfers to Codex, Cursor, Gemini CLI, or anything else you prefer.
You have this skill when: you can name the four areas and explain to a teammate why AI engineering is a skill set rather than a role.
Why AI Software Is Built Differently, and What You Will Build
“When you prompt an LLM, you don’t know what you’ll get back.”
Traditional software behaves the same way every time you run it. AI software does not, and that single fact drives everything else in this course. Because you cannot predict the output, you cannot plan the build in advance the way you plan a normal feature.
Ng's second letter describes what skilled AI engineers do instead. They build a piece, examine it, decide what to try next, and take a sequence of steps shaped by the intermediate results. His phrase for the goal is building reliable systems out of unreliable components.
Most of the 21 skills exist to manage that unpredictability. Evals measure it. Grounding reduces it. Observability catches it in production. Software fundamentals keep the rest of the system predictable, so the AI core has something stable to sit inside.
The running project
Every hands-on step in this course applies to one build: Triage, an internal support agent. Triage reads incoming support tickets, classifies them by product area and urgency, pulls the relevant internal docs, drafts a reply for a human to approve, and escalates anything it is not confident about.
It is small enough to build in a week and it touches all 21 skills. You will make model, data, architecture, eval, security, and scaling decisions on it, and you will use a coding agent to build most of it.
- Create a repo called
triage. - Add a
SPEC.mdwith three lines: what Triage does, who uses it, and what done means. - Expect to rewrite it. You will change this file four times before the course ends. That is the point of it.
You have this skill when: you can explain to a non-engineer why an AI feature takes an iterative process while a normal feature takes a plan.
The Code: Your daily unfair advantage in software engineering.
Join 350,000+ software engineers, tech leads, and CTOs who start their morning with The Code.
By this point you should have:
- The origin of the map, and why it describes a skill set rather than a title.
- The one property, unpredictable output, that shapes every other skill on the map.
- A
triagerepo with a three-line spec you are going to rewrite four times.
Module 1: Building and Deploying AI Applications
The largest area on Ng's map, and the one job postings mention most. Seven lessons on turning an LLM demo into something people can rely on.