Andrew Ng and Spec-Driven Development: What His 2023–2026 Course Catalog Actually Changed

Andrew Ng’s own 2023–2026 course catalog documents a shift from prompt engineering to spec-driven development. Here is the dated evidence, course by course.

Everyone still files Andrew Ng under “the guy who taught the world prompt engineering.” His own course catalog disagrees. That matters for any team that still treats a well-crafted prompt as the finish line, because the April 2026 additions to that catalog show the industry’s most visible AI educator has moved the finish line to a written specification — and did it without ever announcing a pivot.

I’ve been reading DeepLearning.AI’s catalog the way I read a client’s architecture decision records: not for what the marketing copy claims, but for what the artifacts admit. Course titles are cheap. Course syllabi are expensive because someone has to build the labs.

In the next ten minutes, you’ll get the dated timeline — course by course, artifact by artifact — of exactly when and how that shift happened. You’ll also get the part most takes skip: the evidence that cuts against the pivot thesis, and why it does not save the “prompting is all you need” position.

Quick Win — Core Insight

Ng did not repudiate prompt engineering. He demoted it.

In April 2023, his course with Isa Fulford taught one skill: writing precise, single-window instructions for a model with no memory of the project around it. By April 2026, the “Spec-Driven Development with Coding Agents” course, built with JetBrains, teaches something structurally different — writing a project constitution (mission.md, tech-stack.md, roadmap.md) that a coding agent reads before it writes a line of code, then validating each feature against its own spec before the next one starts.

The core insight: prompting never left the curriculum. It stopped being the whole job and became one building block underneath a larger discipline.

The proof is in the dates. 2023: prompting alone. 2024 through 2025: agentic workflows plus evals-first discipline. April 2026: an explicit constitution-and-loop architecture, launched in the same month as a brand-new prompting course — because those two skills now serve different audiences, not different eras.

Ng invented spec-driven development about as much as he invented prompt engineering, which is to say not at all. Both times, his real function was identical: to take an idea already circulating among practitioners and hand it to millions of learners at the moment it becomes worth their time.

With that established, let’s open up exactly what changed in Ng’s own teaching, and when.

History of the Prompt Engineering Phase

The 2023 course was correctly scoped for its moment, and that scoping is the only reason the later courses look like such a sharp departure.

“ChatGPT Prompt Engineering for Developers” launched in the last week of April 2023 — a DeepLearning.AI collaboration with OpenAI, taught by Isa Fulford alongside Ng. Roughly ninety minutes. Nine video lessons.

The entire curriculum rested on two principles: write clear and specific instructions and give the model time to think. Four tactics sat underneath the first one — delimiters, structured output formats, condition-checking, and few-shot prompting, meaning you paste two or three examples of the output shape you want directly into the prompt itself.

The course also taught an iteration loop, and it is worth naming because of what it omits: write a prompt, run it, analyze the result, refine. Every one of those four steps happens in the developer’s head or in a chat window. Nothing gets written down. Nothing survives the session.

The loop closes on human judgment and nowhere else. The tasks matched that shape exactly — summarizing, inferring, transforming text, expanding text, and wiring up a custom chatbot. All of them fit inside a single context window, and all of them were verifiable by eye in under a minute.

That was the whole discipline — no specs, no requirements documents, no multi-file project artifacts. Nothing in it assumed the model might need to know what the project was for.

The architectural assumption underneath it was statelessness — the model retains nothing about the last request when the next one arrives. Every task, therefore, needed an exhaustive, self-contained ticket. Constraints, output format, edge cases, tone: all of it re-declared, every single time.

None of that was a design flaw. It was an accurate reading of what the models of early 2023 could hold and reliably do, and pretending otherwise is hindsight cosplay.

The 2023 Baseline

Ng’s 2023 course, in one picture: one window, one request, no memory of the last one.

The corporate analogy that fits is a talented intern with no memory of yesterday. Brilliant at any task you can fully describe in one page. Useless for anything requiring continuity because you re-brief from zero every morning.

The market treated that narrow skill as a profession anyway. Anthropic’s 2023 “Prompt Engineer and Librarian” listing ran from 175,000 to 335,000 dollars — and the posting itself conceded “the entire field of prompt engineering has only existed for fewer than two years,” which made hiring for it awkward. That same spring, a Stanford economist told a Davos panel — as reported by the World Economic Forum — that it was becoming one of the hottest emerging technology jobs.

Andrej Karpathy had already supplied the era’s defining one-liner in January 2023: the hottest new programming language is English. That line aged into something its author almost certainly did not intend — a slogan for treating natural language as the whole interface rather than one layer of it.

Here is the practitioner consequence, and it is uncomfortable. If your team still evaluates AI competence by asking, “Can this person write a good prompt?” you are grading against a 2023 baseline that even its most famous popularizer has since layered structure on top of.

That baseline held for roughly a year. Then Ng’s own courses started teaching something the 2023 course never mentioned: process discipline.

Signals of Change, 2025 to 2026

The shift did not arrive as an announcement. It arrived as three deliberate steps, and reading them as one continuous arc rather than three separate news items is what makes the quiet pivot visible at all.

Step one was evals-first thinking. An October 2025 agent-design course from Ng reframed the central skill entirely: the strongest predictor of whether someone builds well was their ability to drive a disciplined process for evals — automated tests that score an AI system’s outputs against defined criteria, replacing subjective vibe checks — and error analysis. Not prompt wording. Process. The pre-signal is dated even earlier: at an unscripted Buildathon Q&A in late August 2025, roughly six weeks before that course shipped, Ng was already discussing test-driven development in an agentic-coding context.

The Constitution, File by File

Step two is where the abstraction becomes an artifact. “Spec-Driven Development with Coding Agents” — DeepLearning.AI with JetBrains, taught by Paul Everitt — landed in mid-April 2026. Spec-driven development means writing a precise specification first and having the coding agent implement and verify against it, instead of iterating through freeform chat. It is, in the vocabulary I keep coming back to, a structural specification: a document the agent is forced to read before it is allowed to act.

Ng’s own announcement is blunt about the target: “Vibe coding is fast, but often produces code that doesn’t match what you asked for… Many of the best developers already build this way.”

One precision note for anyone citing this later. Ng’s personal post says developers “already build this way”; the course page says they “already work this way.” Small divergence, real one — the kind that turns into a misquote three reposts downstream.

The concrete architecture breaks down into three layers:

• Project level, written once — mission.md (why this exists and for whom), tech-stack.md (the approved stack, which stops the agent from hallucinating unapproved libraries), roadmap.md (what ships in what order).
• Feature level, repeating — plan, implement, validate. Each feature runs in a fresh agent context, on its own Git branch, with spec files committed alongside the code.
• Between features — a replanning checkpoint that catches drift before it compounds into a rewrite.

In other words: the agent stops being a colleague you ping on Slack and starts being an outsourced vendor you brief with a document. You do the hard cognitive work up front, once, instead of re-deriving it in every chat window.

The artifact Ng’s 2026 course is built around — three files an agent reads before writing any code.

An Hour, Unsupervised

Step three names what the structure buys you. In a June 2026 newsletter, Ng formalized loop engineering — designing the nested feedback loops around a coding agent rather than crafting one-off prompts — into three tiers: the agentic coding loop measured in minutes, the developer feedback loop in minutes to hours, the external feedback loop in hours to weeks. He credited Boris Cherny and Peter Steinberger with popularizing the term rather than claiming to have coined it himself — worth noting in a discourse that is largely a land grab over vocabulary.

The example he used is the most human thing in the entire evidence trail. Over a weekend, he had a coding agent build his daughter a typing-practice app with unlockable cat costumes. The agent worked on its own for roughly an hour, repeatedly opening the browser to check what it had built before coming back to him.

That hour is the payoff. An agent can only run unattended for an hour if something other than a human is telling it whether it succeeded — and that something is the spec, plus evals when a failure mode starts repeating.

None of this explains why Ng, of all people, is teaching a discipline he did not invent. That answer changes what “Ng says it” should mean to you.

What This Means for the Industry

“Diffusion is the process by which an innovation is communicated through certain channels over time among members of a social system.” — Everett Rogers, Sociologist

Who Was Already Doing This

The premise I began this investigation with was that education lags practice by twelve to eighteen months. The dates came in slightly worse than that.

Simon Willison was already describing a detailed-specification-driven workflow in a January 2025 episode of The Real Python Podcast — the earliest documented case I found of an independent practitioner using the concept without the label. An AWS spec-first coding IDE entered preview in mid-2025. A GitHub open-source spec toolkit shipped that September. Ng’s course landed in April 2026, six to fourteen months behind all three.

Karpathy’s “vibe coding” — the practice Ng’s course positions itself against — was coined on February 2, 2025, more than a year before the course that rebuts it.

The honest reading is not that Ng leads this movement. He confirms it — which is a more accurate model of a role he has held all along, given that he did not coin “prompt engineering” either. What he reliably does is tell several million learners when an idea has cleared the bar for their work, and that makes his catalog a timing signal, not a leading indicator.

I’ll correct one of my own assumptions here. I initially read the April 2026 course as opportunism — a big name arriving late to collect an audience. The August 2025 Buildathon material changed that: the process-discipline argument was already a talking point for him eight months earlier. Late to the label, not late to the substance.

Ng’s course wasn’t first. It was the moment the idea went mass-market.

The Measured Case for Structure

Structure got adopted because unstructured AI-assisted work produced numbers nobody wanted to defend in a review.

The sharpest of those is METR’s randomized trial published on July 10, 2025: allowing AI tools increased task completion time by 19% for sixteen experienced developers working on real issues in repositories they already knew, with a confidence interval running from +2% to +39%. The same group revisited a later, different subset in February 2026 under a changed design and came back with a roughly -18% speedup estimate for returning developers (confidence interval -38% to +9%) and a separate -4% estimate for newly recruited developers (confidence interval -15% to +9%).

Two periods, two methods, and a second result that is genuinely hard to read. Neither figure is “the” METR finding, and anyone quoting one without the other is selling something. What survives both is narrower and more useful: AI assistance on a codebase the human knows well is not a free speedup.

The enterprise-scale version of the same problem is MIT’s NANDA report from mid-2025, which put roughly 95% of enterprise generative-AI pilots in the “no measurable return” column across 300-plus deployments. The cause it identified was not model quality. It was the integration and learning gap — systems that cannot retain context and repeat the same mistakes across sessions.

Read that failure description next to the constitution artifact, and the shapes line up. Persistent context committed to version control answers exactly the complaint that a system cannot retain context — whether or not one directly caused the other.

The Two Ngs

Now for the evidence that cuts the other way, because the pivot thesis does not survive without it.

In March 2025, Ng published a widely shared career letter, calling the advice to skip learning code because AI will automate it some of the worst career advice ever given. Six months later, he was still listing prompting among an AI engineer’s core building blocks. Then on April 30, 2026 — two weeks after the spec-driven course — he shipped a seven-hour prompting course of his own.

A pivot that simultaneously ships a bigger prompting course than the original is not a pivot in the “abandonment” sense. It is stratification. Two audiences, two skills, one month.

For the knowledge worker, prompting remains the entire interface, and a 2026 model rewards it very differently than a 2022 one did. For the engineer, direct prompting has become an implementation detail generated by the orchestration layer, from a specification the human wrote. Ng now teaches non-engineers to talk to the machine, and engineers to write the paperwork that supervises it.

This lands where the evidence pointed all along: the best educators were teaching spec-driven development before anyone called it that. The label arrived last.

Summary and Takeaways

The promotion ladder holds all the way up. The company never stopped needing people who write clear tickets. It stopped trusting tickets alone to run production.

What a practitioner should take away from this:

• The shift is documented, not vibes. Ng’s own 2023–2026 catalog moves from evaluating prompt wording to evaluating a project’s constitution and its validated feature loop.
• It is a demotion, not a repudiation. Prompting still gets its own 2026 course — for a different audience than the spec-writing skill now taught to engineers.
• Treat “Ng says it” as a timing signal. His course arrives six to fourteen months after independent practitioners and two major vendor toolchains converged on the same practice. That convergence is the stronger signal.
• The credibility pressure behind the shift is measured. A 19% slowdown in one controlled trial and a 95% enterprise-pilot failure rate are why structure got imposed, though both figures carry scope caveats worth reading before quoting.
• The adoption path needs no tool purchase. Write a one-page mission and tech-stack file. Add a validation step per feature. Add evals only when a failure mode repeats.

That last item is the whole architecture review in three lines. You can propose it on Monday, and the citable case is not “Andrew Ng says so” — it is that independent practitioners, two vendor toolchains, and the field’s highest-reach educator converged on the same workflow from four directions.

The whole pivot, on one ladder.

Closing

“We shape our tools and thereafter our tools shape us.” — John M. Culkin, Communications Scholar, 1967 (describing Marshall McLuhan’s media theory)

Thank you for staying through a timeline this dense — the dates are the argument here, and skimming them would have left you with the same headline everyone else is repeating.

One thing I’d genuinely like to hear. If your team has already written a constitution file for a coding agent, tell me in the comments what went into tech-stack.md and what you deliberately left out — that omission list is where the real engineering judgment lives. If this changed how you read a course catalog, pass it to the architect on your team who is still grading candidates on prompt craft.

If you want to go deeper on the evals side — the point where a repeated failure mode earns its own automated test — I’ve written about that separately.