The AI Engineering Skills Map: Andrew Ng's Four Skills and What They Mean for Hiring

Andrew Ng's AI Engineering Skills Map is drawn from an analysis of over 10,000 job postings and dozens of structured interviews with AI experts, hiring managers, and recruiters. It identifies the four skills that matter most for developers working with AI today. This is the full guide: what each skill contains, why the order matters, and how hiring teams can actually assess them.

It is worth reading side by side with the AI-First Talent framework, which looks at the same shift from the hiring side rather than the individual skill side.

The four skills at a glance

Skill 1: Building and deploying AI applications

The key difference between AI applications and non-AI software is that the output is less predictable. Because of this, building AI systems is far more iterative — you build, examine, and decide what to try next, guided by intermediate results. Ng breaks this into six sub-skills: LLM foundations, grounding models with data, building agentic systems, evaluation-driven development, operating in production, and machine learning foundations.

The single trait Ng says distinguishes strong builders is a disciplined evals and error-analysis loop. That is a craft skill, not a tool.

Full breakdown: Building and deploying AI applications in detail.

Skill 2: Software engineering fundamentals

Coding agents write more code than ever, which raises rather than lowers the bar on judgment. Someone has to decide what the architecture should be, what trade-offs are acceptable between cost, latency, scalability, and reliability, and whether the code that came back is actually correct. "Vibe-coding" without understanding the consequences is how teams accumulate systems nobody can debug.

Full breakdown: Software engineering fundamentals in the age of coding agents.

Skill 3: Using coding agents

Using coding agents well is a distinct skill from writing code. It means knowing what context an agent needs, when to let it run and when to intervene, how to structure verification loops so mistakes surface early, and how to break work into units an agent can complete reliably. The output ceiling is set by how well the human steers, not by the model.

Skill 4: Shaping the build

This is where Ng says the real shift happens. As coding agents get better at executing a given spec, value moves upstream to deciding what belongs in that spec — which requires product sense and domain understanding, not tool fluency. It is also the skill hiring processes are worst at assessing, because it does not show up in a coding test.

Where the AI-First Talent framework starts

The framework asks a related but different question: not "what should a developer learn?" but "how does an organization recognize and hire the right people?" Survey data on AI-First hiring readiness shows:

• Only 12% of organizations have clearly defined AI skills for non-technical roles.
• Only 16% feel very confident they can assess AI-First skills reliably.
• 76% of organizations are still in an early or mid-stage of readiness.

This produced the AI-First Hiring Maturity Model — four archetypes (Observers, Experimenters, Builders, Leaders), each with a distinct risk and a priority action. You can read the full breakdown on the maturity model page.

The connection between the two maps

Ng's map is individual-centric: it tells developers what to learn. The maturity model is organization-centric: it tells companies how ready their hiring process is to spot those same skills. Laid side by side, both point at the same sequence — and the order matters.

Ng lists software engineering fundamentals and shaping the build — deep craft plus business and product context — as two of his four skills, alongside the more visibly "AI" skills. The AI-First framework's core hiring signal is the same pairing: an AI-native mindset combined with deep domain expertise. Neither framework treats AI fluency as the foundation. In both, it is the layer added on top of real expertise in a craft and an understanding of the business it serves.

That ordering is where leverage shows up. Someone with deep domain expertise and strong product sense who then adds coding agents and AI-native workflows does not just get faster — they make qualitatively better decisions about what to build. Someone with tool fluency but no craft can move quickly and still build the wrong thing. We unpack that multiplier effect in Hiring for leverage, not capacity.

How to screen for the four skills

1. Building and deploying AI applications — ask for a system they shipped where the model failed in production. What did the eval loop look like before and after?
2. Software engineering fundamentals — hand them agent-generated code with a real flaw and ask what they would change before merging.
3. Using coding agents — ask them to narrate a recent agent session: what context they gave, where they intervened, how they verified.
4. Shaping the build — ask what they argued against building in the last year, and why they were right or wrong.

Our Agent Skills library contains ready-to-use markdown prompts for CV screening, screening calls, and take-home design that operationalize this rubric.

Frequently asked questions

What is the AI Engineering Skills Map? It is Andrew Ng's framework, published via DeepLearning.AI's The Batch, identifying four skills that matter most for engineers building with AI: building and deploying AI applications, software engineering fundamentals, using coding agents, and shaping the build.

Is AI tool fluency enough to be hired as an AI engineer? No. Both Ng's map and the AI-First Talent framework place craft depth and business context ahead of tool fluency. Tools amplify existing judgment; they do not replace it.

How should hiring teams use the map? As an assessment rubric for technical roles — but the first question is where your organization sits on the hiring maturity model, because most companies are still Observers or Experimenters and lack role-specific standards to screen against.

Read the original from Andrew Ng at The Batch.