AI Tools and Building New Products

The ground has shifted

Between 2025 and 2026, the assumptions underlying software development changed fundamentally.

AI coding tools — Claude Code, Cursor, GitHub Copilot — reached a level of practical maturity where “writing code” itself became dramatically cheaper. What once took weeks now takes hours. The scale of projects a single person can handle went up by an order of magnitude.

This isn’t just efficiency. It means the criteria for what to build have changed.

From “can we build it?” to “should we build it?”

Previously, technical feasibility was the bottleneck. Constraints like “this requires skill X” or “implementation will take three months” narrowed the choices of what to build.

Not anymore. Most things can be built. The question has shifted to whether they’re worth building.

Here’s a trap many developers fall into: just because AI lets you build things quickly doesn’t mean building what AI can build has any value. Users have the same tools. Products that “anyone could ask an AI to make” are losing their worth.

So what should we build?

I believe lasting value comes from things like these:

1. Products rooted in unique data and context

AI has general knowledge but no specific context. Products born from data, experience, and perspectives that only you have are hard to replace.

For example, this site (fragments) is auto-generated from personal records I accumulate daily. Anyone could build the mechanism, but the content can only come from the accumulation of my own activity.

2. Tools with embedded judgment

AI can generate code, but “what should be done” is a human decision. Tools that formalize chains of judgment in a specific domain can’t be replaced by code generation alone.

ILP was born from this idea — making the “why” behind code traceable, giving intent and context even to AI-generated code.

3. Systems that assume continuous operation

One-shot creations are AI’s strong suit. Systems where daily operation, improvement, and accumulated judgment create value — those differentiate not at the moment of creation, but through the process of continued use.

How AI changes the development process

Working with AI tools every day, I notice the process itself changing.

Design carries more weight

With implementation costs down, more time goes to design. Thinking about “what to build” and “why to build it” takes longer than the building itself.

Refine the design through dialogue with AI, then implement in one burst once the direction is clear. This cycle spins remarkably fast.

Experimentation gets cheaper

“Just build it and see” became realistic. Where we used to deliberate carefully before implementing, now we can build a prototype, use it, and then decide.

This site — fragments — went from framework selection and database design to a working product in half a day, through conversation with AI.

Individual scope expands

Frontend, backend, infrastructure, design. Venturing outside your specialty used to carry a steep cost. AI bridges that gap, making it possible for one person to build full-stack products.

The question that remains

The evolution of AI tools is a tailwind for individual developers. But the fundamental question hasn’t changed.

What do you want to build?

No matter how powerful the tools become, only you can answer that. In fact, the “you can build anything” state makes this question weigh more, not less.

With technical constraints removed, the only constraints left are time and your own judgment about what matters.