Гео-аналитическая платформа за 2,5 месяца: как двое передали весь код AI

Two developers built a geo-analytics platform in 2.5 months using an approach that would have taken years in the pre-AI era — leveraging spec-driven development, where developers write only specifications and AI generates the code entirely.

Why a typical AI assistant in the IDE isn't the same thing

Most developers use AI as advanced autocomplete: Copilot suggests a line, Cursor completes a function, ChatGPT explains an error. Spec-driven development is a fundamentally different mode that changes the developer's role entirely.

Developers stop writing code altogether. Their job is to describe the desired system behavior in specifications: what should happen, under what conditions, what result is expected. AI implements the logic based on these descriptions, rather than complementing what a human has written.

This transforms the nature of work: instead of coding — design and task definition. The developer becomes an architect and editor, not an executor of details. The tool itself is not an IDE plugin, but a systemic process in which AI is embedded in every step.

What two people built in 2.5 months

A team of two created a geo-analytics platform from scratch — with no legacy code, no constraints from existing architecture.

• Team: 2 developers
• Timeline: 2.5 months
• Approach: spec-driven, AI generated all production code
• Starting point: blank slate, complete freedom in architectural choices
• Context: an analogous product in the pre-AI era would have taken years and required a larger team

The absence of legacy code is a critical success factor. Spec-driven processes develop organically from the start, whereas adapting AI to an inherited codebase is fundamentally harder: existing code resists — it has its own logic, context, and technical debt that AI handles poorly.

The platform's genre adds complexity: geo-analytics involves working with spatial data, maps, layers — not a typical CRUD project. The result is all the more instructive.

Where failure was expected — and what actually happened

The authors' initial skepticism was understandable: AI handles small, isolated tasks well, but on a large project it loses context, hits token limits, and starts proposing solutions that contradict what's already written.

"At first, I didn't believe it would hold up at real scale.

Experience suggested: the bigger the project, the faster AI gets confused and hits its limits," the author writes.

After 2.5 months, the conclusion changed. With a properly structured process, AI maintains context throughout a large project. The spec-driven approach structures interaction with the model such that each request is self-contained and doesn't require AI to keep the entire development history in memory — this eliminates the main risk of context loss.

A new critical skill for developers is not the ability to code, but the ability to formulate specifications precisely: without ambiguity, with the right context and clear success criteria. Those who write better specifications get better code.

What this means

The geo-analytics platform case is working proof that spec-driven development scales beyond textbook examples. Two people with the right process launched a product that previously required a larger team and years of work. The link between "task complexity — team size" weakens: AI handles implementation if a person can describe tasks accurately.

The question is no longer "should we trust AI with a big project," but "how do we structure the specification process so it works consistently."

Frequently asked questions

Do we have to start a project from scratch?

The authors are direct: starting without legacy code is the key condition. Spec-driven development emerges organically precisely on greenfield; adapting it to an inherited codebase is fundamentally harder — existing code resists.

What skill becomes primary in the spec-driven approach?

Based on the authors' experience — the ability to formulate specifications clearly: describing desired system behavior without ambiguity, with necessary context and criteria for a finished result. The ability to code takes a back seat.