Why Universities Should Be the Ones Learning AI-Driven Development: How Spec-Driven Development Pairs With AI for Science

Hello, this is Ryuta Hamamoto from TIMEWELL.

We have designed a number of courses over the years that teach university students and young professionals the craft of building things with AI. One thing I have felt throughout is this: the number of people who can touch AI has grown, but the number who can hand AI a job properly is still small. Type a prompt and something comes out. Whether that something is of a quality you can use in production, though, is another matter entirely. I think the key to closing that gap lies in the approach now called spec-driven development.

And the place that should learn this craft most thoroughly is the university. That is what I want to get across in this article. As it happens, the government has just placed AI for Science, the effort to advance scientific research with AI, at the centre of national strategy. The craft of development and the advance of research are, in fact, connected at the root. If you are curious where you, your child, or your organisation currently stand in terms of AI fluency, checking your position first with our free AI literacy self-check will make the second half of this piece feel more concrete as you read on.

From the reckoning with "building on vibes" to spec-driven development

Let me first lay out what is actually happening on the front lines of AI development right now. Once this is clear, it connects to the university story.

Around 2025, an approach spread rapidly in which you build software just by giving AI instructions in plain, natural language. This way of letting AI build on momentum, without worrying about detailed design, picked up the name "vibe coding." It is reported to have been coined by the AI researcher Andrej Karpathy in February 2025¹. It is easy, and because something that runs appears right away, it is appealing as a first step.

There is a trap in it, though. When you let AI build loosely from vague instructions, a gap gradually accumulates between what you really wanted to make and what the AI actually made. With a small program you can push through on momentum, but the more features you add, the more people join, and the longer the software lives, the more this accumulated gap gets out of hand. It runs, but you no longer know what to fix or how. That is the state you fall into easily.

Out of that reckoning, spec-driven development has become the mainstream approach. In English it is Spec-Driven Development, or SDD for short. A specification is a design document that clearly states what you want to make. In spec-driven development, rather than making the AI write code straight away, you first fix in words what you are going to make, and then have the AI proceed in the order of planning, breaking down the work, and implementing, all based on that specification.

The tool that turned this way of thinking into something concrete is "Spec Kit," the open-source toolkit GitHub released on 2 September 2025². It splits development into four stages, defining the specification, drawing up a technical plan, breaking the work into small tasks, and implementing, and it proceeds while leaving each of them as a written document. In effect, it places a shared blueprint between the human giving instructions and the AI doing the hands-on work. Just as construction without drawings is precarious, development without a specification looks fast but is sure to slow down later.

What building courses taught me about the strength of people who can write a spec

From here, let me talk about what we have actually seen through running courses.

At TIMEWELL, we have been involved in developing the AI skills of more than 500 people. Including initiatives such as WARP ENTRE, which is based on an agreement with the Tokyo Metropolitan Government, we have repeatedly designed settings where university students and young professionals build something with AI as their weapon. What we grasped there, as a real sense rather than a theory, is that what separates the quality of the output is not skill in operating AI tools, but the ability to put into words what you want to make.

When we design a course, one of our aims is to reach a quality that holds up in production within a short period. My honest sense is that with roughly two months of concentrated learning, we can bring people to where a product that works in real practice is within reach. Please take this not as an external statistic but as our own read from designing and running courses. How fast someone reaches that point varies with the learner and the subject, and it is not a promise that anyone will always get there in two months.

Why does fixing the specification first let even beginners get close to production quality? The reason is simple. What beginners stumble over is usually not the way to write code itself, but the fact that "in the end, it has not been decided what to build." When you fix the specification in words first, the range you leave to the AI and the range you have to judge yourself split apart cleanly. AI speeds up the hands that implement, but it does not take over the judgment of what to make. The more a person can hold that part, the less rework there is, and the more directly they head toward completion.

Turn that around and the ability to write a specification becomes the foundation of every kind of knowledge work from here on. To hand AI a job properly is, in short, to put into language, without ambiguity, what you want done. As a place to train that ability, I do not think there is an environment better suited than the university. If your organisation is unsure how to root this craft in your younger people, our AI consulting service WARP has a part it can help with. Specialists who led DX and data strategy at major companies walk alongside you month by month and help with the design.

What the national strategy signalled about the changing role of universities

Let me take the university story one level deeper. The government too has officially set out that the very role of the university is changing.

On 31 March 2026, Japan's Ministry of Education, Culture, Sports, Science and Technology issued its "Basic Strategic Policy for Promoting AI for Science"³. AI for Science is a term for the effort to use AI to advance scientific research itself. The strategy positions AI as being deeply involved in every stage of research, forming hypotheses, designing experiments, analysing data and organising knowledge, and as changing the very shape of science. On that basis, it set out the aim of becoming not a country that merely uses AI, but a country that creates new science together with AI. The stated purpose is "the revival of science."

What I felt was most important in this strategy is that the way it draws the role of the university has clearly changed. Until now, the one-way flow was taken for granted: universities generate the seeds of knowledge, and companies put them to practical use. The strategy goes beyond this linear model and states plainly that it will shift to a circulating model in which scientific insight and AI technology complement each other and create together. And it writes that universities need to become not simply the side that provides insight, but co-creators of high-precision experiment and theory-building, who verify the hypotheses AI presents and assemble them into theory.

To support this shift on the talent side, the strategy sets out concrete numbers too. Train more than 3,000 people who can make advanced use of AI in scientific research over five years. Generate 3,000 or more cases of AI for Science use over three years. Double the speed of the academic information network by FY2028, and increase shared computing resources more than tenfold by FY2030. Goals like these are lined up. Alongside them, it sets out mechanisms for university faculty to join company front lines, recurrent education for working adults to relearn at universities, and efforts to raise the AI literacy of both faculty and staff.

One note, since the figure 3,000 appears twice and I do not want them confused. One is the talent-development goal of more than 3,000 people over five years; the other is the goal of 3,000 use cases over three years. People and cases are separate matters. Either way, the country has clearly repositioned the university as a place where people who advance science together with AI gather and grow. If you want to know more about the whole picture of this strategy, I cover it in detail in an article that explains the mechanics and national strategy of AI for Science in plain language.

Why spec-driven development and AI for Science pair so well

This is the heart of the article. Let me explain why spec-driven development, the craft of building, and AI for Science, the way of advancing research, are of a piece.

Break down the flow of scientific research and it actually looks a lot like development. First you raise a question, form a hypothesis, design the experiment or analysis that confirms it, verify the results, and organise them into an output called a paper. Across this sequence, the range AI can now handle has expanded astonishingly. AI proposes a hypothesis, assembles an experimental procedure, runs an analysis, and organises it into text. We are standing at the entrance to an era like that.

Several developments have been reported as symbolic examples. A paper summarising the methodology of an AI that advances research autonomously, developed by Japan's Sakana AI, was published in the journal Nature in March 2026⁴. It is described as a system that generates ideas, surveys the literature, runs experiments and even writes a draft of the paper. Overseas too, Google is reported to have announced, in February 2025, a mechanism in which multiple AI agents divide roles to generate hypotheses and experimental procedures⁵. Going further back, the developers of AlphaFold, the AI that predicts the structure of proteins, receiving the 2024 Nobel Prize in Chemistry⁶ was an easy-to-read milestone marking that AI had begun to move the frontier of science. I leave the detailed content of these to an article that surveys the front lines of AI-driven research.

What matters here is whether researchers and students can move to the side that uses these AIs with their own hands, rather than merely watching them. And the very craft needed for that is spec-driven development. To have AI help with research is nothing other than to put into language, without ambiguity, "what you want to confirm," "under what conditions," and "what result would count as success," and to hand that to the AI. This has exactly the same structure as the act of writing a specification in development.

Instruct on vibes and AI will return a plausible-looking hypothesis or piece of code. But whether it truly holds up to verification, or whether it drifted off course along the way, cannot be checked without a specification. In fact, autonomous research AI has also been noted to have limits, such as errors mixed into what it generates, or proposed experiments that do not run. That is precisely why the human role of verifying AI's output and building it into theory becomes the linchpin. The "co-creator of verification and theory-building" the national strategy asks of universities is, put another way, a person who can hand AI a proper specification and scrutinise what comes back. Learning the craft of development becomes, as it is, the ability to advance research. This continuity is what I think is the single greatest reason universities should learn AI-driven development.

Universities become the place that trains people who can write a spec

Finally, let me write about how universities should concretely move, and what we can do.

One realistic entry point is to build AI-driven development into research and problem-solving. Rather than learning operations the way you memorise grammar in a classroom, take your own research theme or a familiar problem as the subject, write a specification, and finish an output with your own hands alongside AI. This way of learning while tackling real problems is called PBL. It stands for project-based learning, a way of acquiring the abilities you need as you build, rather than cramming knowledge in first. I feel spec-driven development pairs very well with PBL. It is because you have something concrete you want to make that writing a specification finally makes sense in your gut.

Another is to cross the walls between fields. The national strategy tries to generate cases of AI for Science use across every field, from the humanities and social sciences to the natural sciences. Public calls that back individual challenges on the scale of several million yen each have also begun to move⁷. The less a researcher specialises in programming, the more the craft of writing a specification and leaving the implementation or analysis to AI comes into its own. Even when the specialty differs, the ability to put into words what you want to make becomes a shared foundation. On the continuity of culture and education, and how the younger generation should engage with AI, I also touch on this in our piece on AI education from the high-school stage.

To be honest, this kind of shift does not advance on orders alone. Hand out the tools and, unless the craft of writing a specification reaches the front line in a usable form, they become treasure left to rot. What we have felt we are getting somewhere on, in our courses and consulting, is exactly this "translation" part. Take the most advanced craft and land it as the next step for the specific lab or department in front of you. When there is a certain pattern to that, the speed of development changes.

If you are in a university, a lab, or a corporate R&D department and are thinking about building AI-driven development or AI for Science in-house, we can help from the stage of sorting out where to begin. With the use of public funding also in view, we design together the path by which young people come to write specifications. If you are interested, please talk to our WARP team. Specialists who led DX and data strategy at major companies walk alongside you month by month.

To sum up

This ran long, so let me organise the key points.

• Out of a reckoning with the limits of letting AI build on vibes, spec-driven development, which fixes what you want to make as a specification first, is becoming the mainstream. GitHub's Spec Kit is a leading example
• What separates the quality of the output is not skill in operating tools but the ability to put into words what you want to make. Fixing the specification first reduces rework even for beginners and makes it easier to get close to production quality
• The Ministry of Education set out a national strategy in March 2026 and repositioned the university as a co-creator that verifies the hypotheses AI puts forward and builds them into theory. It sets goals such as training more than 3,000 people over five years
• The flow of research closely resembles development, and the craft of handing AI a proper specification becomes, as it is, the ability to advance research. Learning development and advancing research are of a piece
• Within PBL and across fields, universities can become the place that trains people who can write a specification. TIMEWELL continues to walk alongside that in-house build

The number of people who can use AI will keep growing from here. What will make the difference is whether you can put into words, without ambiguity, what you ask AI to do. As a place to sharpen that ability, no environment suits better than the university. Set the craft of writing a specification as the foundation, and development and research both move forward with the same single step. Start from putting into words what it is you want to make right now.

References and primary sources

Footnotes

1. The coining of "vibe coding" is attributed to a post said to have been made by Andrej Karpathy on 2 February 2025 (reported by multiple technical media; the original post not directly confirmed). As an example of a term explanation, see @IT, "What Is Spec-Driven Development (SDD)?" https://atmarkit.itmedia.co.jp/ait/articles/2510/07/news022.html ↩

2. GitHub, "Spec-driven development with AI: Get started with a new open source toolkit," 2 September 2025 (author Den Delimarsky) https://github.blog/ai-and-ml/generative-ai/spec-driven-development-with-ai-get-started-with-a-new-open-source-toolkit/ . See also the Spec Kit repository https://github.com/github/spec-kit ↩

3. Ministry of Education, Culture, Sports, Science and Technology, "Basic Strategic Policy for Promoting AI for Science," issued 31 March 2026 (Reiwa 8) https://www.mext.go.jp/content/20260403-mxt_jyohoka01-000048752_1.pdf ↩

4. A paper reporting the methodology of Sakana AI's "The AI Scientist" was published in the journal Nature (26 March 2026). This includes the point that it was voluntarily withdrawn after passing peer review at an ICLR 2025 workshop (per Sakana AI's official account) https://sakana.ai/ai-scientist-nature-jp/ ↩

5. Google, "Accelerating scientific breakthroughs with an AI co-scientist," 19 February 2025 (a Gemini-based multi-agent system; specific research cases per reporting) https://research.google/blog/accelerating-scientific-breakthroughs-with-an-ai-co-scientist/ ↩

6. The 2024 Nobel Prize in Chemistry was awarded to Demis Hassabis and John Jumper, who developed AlphaFold, and to David Baker for protein design (per multiple reports). ↩

7. Ministry of Education, Culture, Sports, Science and Technology, "AI for Science Emerging Challenge Research Creation Project (SPReAD 1000)," public call information https://www.mext.go.jp/aifors_spread/ ↩