Minesweeper III: "A Quality Leap"

Since January 2023, we have been using the same challenge — "generate a functional Minesweeper" — as a thermometer to measure the real capability of generative AI models. Why Minesweeper? Because it seems trivial, but it is not: it requires managing two-dimensional arrays, controlled recursion (flood-fill), event management, multiple states per cell, and complex conditional logic. It is the type of problem that exposes an LLM that only knows how to copy patterns.

Today, three years and three generations of models later, the results speak for themselves.

👉 Read the V1 article (2023)  ·  👉 Read the V2 article (2024)

2023 — OpenAI Playground: Asking AI for a Minesweeper was like supervising a junior's practices. It understood the general concept, but the grid broke, recursion entered an infinite loop, and mines were visible from the start. The impressive part was that it generated something, but the result needed hours of human debugging to be usable.

2024 — ChatGPT 4o with Canvas: The leap was real. There were no structural bugs; the AI understood flood-fill, generated clean CSS, and even implemented the safe first click without being asked. But you still needed to guide the process: initial prompt, correction, refinement. Three rounds to get to something publishable.

2026 — Claude Opus 4.6: This is where things change category. A single detailed prompt and the result is a complete game with 3 difficulty levels, elaborate 3D aesthetics with an industrial panel, animations (the mine cascade when losing is spectacular), chord-click, a record system with localStorage, touch support for mobile, and keyboard shortcuts. No bugs detected in the first execution.

What impresses the most is not the amount of code — it's ~1000 lines, something a senior could write in a couple of days. What impresses is the architectural coherence: the model separates state and presentation, uses consistent CSS variables, implements an IIFE to avoid polluting the global scope, correctly manages event listeners, and has solid stopping conditions in recursion. It does not copy fragments — it reasons about the complete problem.

Does it replace the programmer? No. But it changes the role: from writing code to designing prompts and validating results. The well-structured prompt is the new source code.

Result of a single prompt. Complete Minesweeper with industrial control panel aesthetics: 3D buttons with physical depth, metal frame with screws, seven-segment displays, status LEDs, and impact animations.

Controls: Left click = uncover · Right click (or long press on mobile) = mark flag ⚡ · Click on revealed number = chord (opens adjacent if flags match) · Keys: R restart, 1/2/3 difficulty.

Minesweeper has not changed in three years. The challenge is the same. What has changed radically is the ability of AI to solve it.

From a model that generated broken code and needed 8 iterations for something functional, we have moved to another that produces a complete application — with clean architecture, professional aesthetics, responsive, bug-free — in a single conversation turn. That is not an incremental improvement; it is a category change.

For us, who have been documenting this evolution since 2023, the conclusion is clear: generative AI is no longer a curious assistant — it is a serious production tool. And the most interesting thing is that this is just the current state. Within a year, this same article will likely seem modest.