Module 4: Shaping the Build

Driving the Build Loop

Ng's fifth letter opens with a claim about roles. Before AI tools expanded what one developer could do, companies settled on a split: product managers and designers specify, developers build, and maybe a project manager drives the timeline. Those roles are blurring. A developer skilled at AI engineering builds the software and takes part in the other roles too, and product managers and designers are moving the other way at the same time.

The first sub-skill is the one that makes the rest possible. Most software gets built in a loop: write some code, get some feedback, decide what to do next. Ng's point is that a skilled AI engineer drives that loop, repeatedly choosing the next step, with a bias for action and at the speed AI has made possible.

The failure mode: Triage sits in a branch for three weeks while you wait for someone to tell you whether the confidence threshold should be 0.7 or 0.8. Nobody was ever going to tell you. The loop stalled because you stopped driving it.

Driving the loop is a set of concrete calls. Build a quick prototype to test a technical idea, or an MVP to put in front of users, or add features, or invest in an enterprise-grade version. Ship in small batches to keep velocity up. Know when to ask users for feedback and when to run a technical experiment instead, such as training a model to see whether the data supports the idea. Weigh each call against the product vision, the stage of the project, feasibility, risk, effort, and budget. On a mature project, define the key metrics and manage the work toward them.

You have this skill when: you can name the next step for your project right now, and it is a decision you made rather than one you are waiting on.

Making Product Decisions

But you will make the decisions the spec does not cover, and if you are asked to build without a spec, you know how to write one. Ng breaks the skill into four parts. Product sense, so you can pick a direction that meets real user needs without waiting for a PM on every call. Basic design sense, so what you build is pleasing to use and not merely functional. Basic business sense, so you can think through go-to-market, market size, unit economics, and profit and loss, and make tradeoffs that are economically sensible. And underneath all three, user empathy.

The failure mode: Triage is technically excellent and support staff do not use it, because the drafts sound nothing like how they talk to customers. Nobody on the build asked them.

Ng is specific about how empathy gets built, and it is not by intuition. Quick informal interviews with two or three users. Surveys of hundreds. Large-scale A/B tests. Analyzing the behavior of thousands or millions. Each method fits a different stage, and the skill is using the resulting input to keep correcting your picture of the user. A spec that starts with system behavior produces software that is correct and unwanted. A spec that starts with what the user is trying to do produces software people open twice.

You have this skill when: your spec starts with what the user is trying to do, and you can say what the build costs against what it saves.

Communicating and Leading

Ng's third sub-skill follows from the first two. AI engineering skills let you take part in a wider scope of work than traditional development allowed. He has written before about specialized developers moving to a full-stack role. This goes further: you might take part in marketing, finance, legal, and whichever other functions touch your project. That makes your ability to communicate with those functions matter more than it used to, because you can move the project forward by aligning and coordinating the people around it.

There is a second reason, and it is specific to this moment. AI is changing fast enough that many people outside engineering are trying to work out what it can do, what it means for their jobs, and what it makes possible. Your technical skill puts you ahead, and Ng's framing is that this lets you play a unique role in shaping those perspectives. You can say whether an initiative is technically feasible. That is leadership, whether or not your title says so.

The failure mode: the head of support announces a plan to auto-close tickets Triage marks as low urgency, because nobody told them that the urgency classifier is right 78 percent of the time. You knew. You did not think it was your job to say.

You have this skill when: the people who depend on your project know what it can and cannot do because you told them, before they found out.

High-Agency Ownership

Ng's last sub-skill is the one he sounds most excited about. AI engineering skills give you an unusual amount of room to make a difference, partly because many people, including some executives, do not yet understand what AI can do and so cannot name good project directions. That gap is an opening for someone with technical skill: spot problems, propose solutions, execute on them, respectful of the organization's priorities and constraints but without waiting for precise top-down direction.

He lists what high agency looks like in practice. Identify opportunities, prioritize what matters, act. Own an initiative end to end. Take accountability when things go wrong. Act in the face of ambiguity. Persist through setbacks. And measure your work by the value it creates rather than by tasks completed.

Then one more habit, which Ng puts last and which sits under all four areas of the map: invest in your own skills. Track the technology frontier, pick up new tools, tune your workflows, and keep learning, so you get better over time. His letter on coding agents said the same thing more sharply, that keeping up needs a routine of experimenting and building, not a one-time course.

The failure mode: Triage works, nobody asked for the next version, and you go back to the backlog. Six months later a vendor sells your company the thing you could have built in a month.

You have this skill when: you bring the next problem to your manager before they bring it to you, and your eval has run on at least three model versions.