4 AI Skills That Actually Get Hired (Andrew Ng's Roadmap)

0:00 On the 14th of August, Andrew Ng published a diagram, four boxes and a claim about what the job market is buying. Those four boxes are what over 10,000 real job ads are hiring for in AI

0:11 engineering. Prompt engineering is not among them. Box one, building and deploying AI applications. Box two, software engineering fundamentals. Box three, using coding agents. Box four,

0:23 and this is the strange one, shaping the build. Shaping the build is the odd one out. The other three you can go and study tonight. This one Andrew Ng ranks with them anyway. On his post, bookmarks

0:34 ran ahead of likes, which is what people saving something to act on looks like. This was not a guess, and that is the part worth slowing down for. Andrew Ng and his team read over 10,000 job

0:44 postings first. They ran dozens of structured interviews with AI experts, hiring managers, and recruiters, then layered in survey data. His own description of the method is a line an

0:54 engineer will recognize, clustering on a massive data set of jobs and expert interviews. So, this is a labor market measurement published by the man whose course taught machine learning to

1:03 millions of people. And the measurement says the engineering job split in half. Agents took one half, the other half is the half that now pays. Start with box one. The difference between an AI

1:14 application and a normal one is that the output is unpredictable. Prompt a model and you do not know what comes back. Train one and you do not know what it will predict on the next example.

1:24 Andrew Ng lists the building blocks, models, context engineering, retrieval augmented generation, agentic workflows, machine learning. Then he puts the weight somewhere else entirely,

1:35 statistical technique to measure, steer, and govern the system so it behaves predictably. His words for the core of it are disciplined evals and error analysis loops. Anthropic published what

1:46 that looks like in January. Start with 20 to 50 tasks pulled from real failures, not hundreds. You measure pass at one, then pass to the K, because an agent that succeeds three times in four

1:57 on a single try clears three trials in a row only 42% of the time. The second box is software engineering fundamentals and on the map it sits above coding agents. The reason Andrew Ng gives is narrow and

2:09 useful. Fundamentals are what let you see that a trade-off exists at all. Cost against scalability, reliability against speed, security and privacy layered over both. His example is pointed. An

2:21 inexperienced developer vibe codes a solution without knowing which trade-offs the agent is making and the agent tends to make poor ones because that developer cannot tell it which

2:29 context matters. Fundamentals are how you steer in the precise language of software engineering. That ranking is Andrew Ng answering a question the internet keeps shouting. Is software

2:39 engineering dying? Two days after the map went up somebody put the blunt version of it to Hacker News. The thread, do you know of any company that went back to handwritten code? 101

2:49 points, 118 comments and a careful framing. Not whether AI is good, only whether anyone has walked it back. One reply named Ford, Commonwealth Bank of Australia and IBM underneath it. Three

3:01 separate people took that apart. Those firms rehired staff after AI layoffs and Ford's returning engineers were rebuilding the data pipelines that feed its own AI training. They cut people

3:11 early. Nothing in the thread showed they banned the tools. What the thread did produce was a picture of sculpt use. One founder allows AI for review and not for writing. The work is deep, full

3:22 understanding of the code matters more than speed and a client catching a whiff of generated code would end the deal. A second commenter who adopted every wave from Tab Nine to

3:33 still writes the interesting code by hand. His reason is the sharpest line in the thread. Deriving the code and building the mental model is about 90% of the work and the code itself is the

3:43 other 10. A reply from outside software entirely. A doctor moved back from AI scribes to typing her own notes and several colleagues did the same because checking the output and turning verbose

3:54 prose into a usable note ate the time saved. That is the verification tax, and it shows up in code review, too. The market numbers tell the same story. Indeed's Hiring Lab in July, software

4:05 development postings are still 27.5% below their pre-pandemic level. Call that the doomer half of the chart. It is real, and it is the half that gets quoted. The same report holds the other

4:16 half. Those postings are up almost 15% since Claude Code launched, while postings across the whole economy fell 7%, and of the rise over the last year, 71% is senior roles, 37% is jobs with AI

4:29 in the title. So, the hiring is real, and it is concentrated at the top, which puts the weight back on box three, because using coding agents appearing on a skills map from Andrew Ng is new.

4:39 An agent workflow used to be a productivity hack. Here, it is a hiring line. Andrew Ng spells out what the skill contains, a mental model of how agents work, their limits, and how much

4:49 to intervene versus leave them alone, managing the agent context, trading planning against execution, then the rest of it. Giving the agent verifiers so it can close its own loop,

4:59 orchestrating several at once, and keeping any of them away from your production database. One more clause makes it a skill rather than a tool, because agentic coding moves fast.

5:09 Andrew Ng says you need routines for trying new tools and evolving your workflow as best practice changes. The skill is the habit of relearning it. That is measurable. In the three days

5:18 before Andrew Ng hit publish, Claude Code turned sub-agent forking on by default. So, a forked agent inherits the whole conversation in the prompt cache. Codex shipped importing your setup,

5:29 instructions, settings, skills, projects straight out of Claude Code and Cursor. The harness you learned in July is not the harness you are running in August. All three boxes so far are about getting

5:40 code produced correctly. The fourth asks whether it should be produced at all. Andrew Ng sentence for it is short. Given a clear spec, coding agents are rapidly improving at

5:49 delivering to it, so our work as engineers is shifting toward deciding what should be in the spec. He pushes it further. Engineers should no longer expect a pixel-perfect design handed

5:59 over with an instruction to implement it. This work needs product sense, business context, customer goals, and the judgment to know when a rough MVP in front of users beats another week of

6:09 care. If that sounds soft, look at what the tooling did. Spec kit, GitHub's spec-driven development toolkit, has passed 129,000 stars since it went up

6:19 last August. Writing the spec has an ecosystem. The pay data points the same way. PwC read more than a billion job ads across 27 countries. Wages for AI skills carry

6:30 a 62% premium, up from 57 a year earlier. And jobs asking for AI skills grew 69% while the whole market grew nine. Then the number that explains box four. Entry-level roles most exposed to

6:43 AI are seven times more likely to require traditionally senior, human-intensive skills. Leadership, creativity, dealing with people face-to-face. The workforce lead at PwC,

6:55 Pete Brown, named the mechanism. AI is removing some of the routine work that once acted as an apprenticeship while increasing demand for judgment, leadership, and adaptability much

7:04 earlier in careers. Which is why that founder on Hacker News is the whole map in one anecdote. His problem was not the AI writing bad code. It was feature bloat and developers he could not slow

7:15 down. The coding went fine. The shaping did not. So, the verdict. The market did not stop hiring engineers. It changed which half of the job it pays for, and that half is judgment. Andrew Eng put

7:26 two judgment skills in the top four. Indeed says 71% of the software hiring rebound is senior. PwC says exposed entry-level work demands senior skills seven times as

7:37 often. That holds for most of you and not for all. For maybe one shop in 20, deep systems work. A domain where the mental model is the product, hand authoring still wins, and box two on

7:49 that same map is their argument. And the idea worth arguing with is not Andrew Ng, and not the people on that thread. It is the reflex that sold prompting as the skill of the decade while the market

7:59 was paying for something else entirely. Which leaves the question the map does not answer. If shaping the build decides your next five years, how do you practice something with no course to

8:09 take and no ticket to close? Andrew Ng says he has more to say and will keep updating the map. His next letters are where that answer has to live.