Recreating Classic Games with AI Coding: Practical Inspiration from Codex and Claude Code

AI coding tools are changing the barrier to entry for indie game development. A Bilibili creator shared a series of game projects built with Codex and Claude Code, ranging from classic board game recreations to original rhythm game experiments, demonstrating the real possibilities of AI-assisted coding in game creation. Most of these projects aren't particularly complex, yet they cover a remarkably diverse range of gameplay types — offering useful reference for developers who want to build something themselves.

Why Starting with "Recreations" Is a Smart Choice

According to the creator's sharing, a considerable portion of his games are clones of existing works, including Tetris, Puyo Puyo, Liar's Dice, Reversi, and Connect 4.

The advantage of recreating classic games is that the rules are clear and the boundaries well-defined. AI coding tools can quickly generate a runnable code skeleton, and developers only need to adjust gameplay details. The creator specifically mentioned Liar's Dice — he felt there wasn't much content online about a "complete solution" for this game, so he implemented a solving approach with some originality of his own. Projects like this are well-suited as AI coding practice exercises: the rules are known, but the implementation path needs to be designed by yourself.

Codebreaker, a color-guessing game similar to Mastermind, also falls into the logic puzzle category. The creator described the implementation as follows: players think of different color combinations, and the system provides feedback responses. The logic layer of such games isn't complex, but there are plenty of details to polish in UI feedback and state management — exactly the areas where human intervention is needed after AI generates the code.

Tetris: From Playable to Practicable

Tetris is the most detailed project mentioned in this batch. The creator stated that his clone not only allows free play but also supports practice modes targeting specific techniques — he specifically mentioned "DT Cannon," a popular opening setup in the player community, where the game guides players to place blocks at specific positions.

This point is worth noting: tutorials for simply cloning Tetris are everywhere online, but adding a "setup practice" feature means the developer needs to understand the game's advanced techniques and translate that understanding into interactive training guidance. This is exactly where AI coding tools can help — generating basic line-clear logic is fast, but state judgment and hint logic for specific techniques require the developer to repeatedly converse with the AI and iterate on the code.

The creator also mentioned technique terms like P-Spin, indicating that this project targets players with some foundation rather than being a purely casual recreation.

DT Cannon is an advanced opening setup in the Tetris player community, short for Double Triple Cannon. Through a specific sequence of block placements, it constructs slots that can trigger both a Double (two-line clear) and a Triple (three-line clear) during the opening phase, sending a large number of garbage lines to the opponent in a very short time. Mastering DT Cannon requires precise block placement order and rotation operations, demanding high precision and planning ability from the player. The creator's addition of a setup practice mode in the game means the program needs to be able to recognize whether the player is placing blocks according to the standard sequence and provide hints when they deviate — this involves real-time tracking of game state and pattern matching, far more complex than basic line-clear logic.

P-Spin is another advanced technique that refers to rotating a T-piece (or other piece) into a narrow gap and completing a clear when it cannot be placed normally. P-Spin detection depends on the game's legality check for post-rotation positions, and rotation rules (such as SRS, Super Rotation System) vary greatly across different versions of Tetris. The creator's mention of these terms indicates that his project intentionally reproduces the core mechanics of modern competitive Tetris rather than being a simple casual version.

Original Experiments: Half-Finished Rhythm and Fighting Games

Beyond recreations, the creator also attempted more original projects. One of them, a rhythm game (which he calls Nyx), is described by him as "a bit buggy" but shows visible potential. The inspiration is similar to the "Twinkle Twinkle Little Star" melody paired with fast-moving wind/star elements, where players press keys (1, 2, 3, 4, 5, 7, 9) to trigger running actions.

He admits that this game "doesn't run very well," but a combo mechanism has already been added — ideally, the higher the combo and the more correct notes, the faster the wind in the scene should run. This unfinished state precisely illustrates the real workflow of AI coding: AI can quickly build a prototype, but gameplay feel and rhythm feedback still require repeated debugging by the developer.

Another project is a fighting game clone inspired by "Little Fighters," made by a Hong Kong developer, which the creator says was quite popular when he was in eighth grade. His version is called "Bro Fighters," with characters based on himself and classmates, and backgrounds based on his alma mater — a "half-finished attempt."

The value of such personalized projects isn't in their completeness but in validating an idea: with AI tools, turning friends into game characters and schools into scenes is no longer something that requires a professional team.

The combo mechanism in rhythm games is a common scoring system in rhythm games: the combo count accumulates when players consecutively hit correct notes, and resets to zero upon a miss or wrong hit. The combo count is usually tied to score multipliers and also affects visual feedback (such as the intensity of visual effects). The creator's ideal state — "the higher the combo and the more correct notes, the faster the wind in the scene should run" — essentially maps the combo count to the movement speed of dynamic elements in the game, a design technique for visualizing gameplay data. Implementation requires passing note judgment results to the rendering layer in real time and applying smooth interpolation; otherwise, sudden speed changes will appear jarring. In AI-generated prototypes, such mechanisms often only have the data flow framework set up, and feel and rhythm matching still require manual tuning.

Takeaways for AI-Assisted Developers

Several practical lessons can be distilled from this batch of projects.

Start with subjects that have clear rules. The rules of board games, dice games, and match-3 games can be clearly described in natural language, leading to higher accuracy in AI-generated code and lower rework costs.

Add a layer of differentiation on top of recreations. Pure clones tend to get lost among similar works. Features like Tetris setup practice mode and Liar's Dice's original solving method are ways to give a project distinctiveness.

Accept that prototypes aren't perfect. The creator's rhythm game and fighting game are both in half-finished states, with bugs present and feel not yet tuned. But getting a prototype running is itself a milestone achievement, and it can be iterated on afterward rather than waiting until you've "thought it through" before starting.

Draw subjects from personal experience. Popular games from middle school days, friends around you, your own school — these materials give projects a unique emotional connection and make them more relatable when shared.

The creator's call to action at the end of the video is quite direct: if you have time and ideas, you can absolutely develop games — there are plenty of opportunities, don't miss out. Considering that tools like Codex and Claude Code have lowered the coding barrier to a considerable extent, this advice is indeed worth taking seriously for those with some programming foundation who lack projects to practice on.

However, limitations should also be recognized: these projects are mostly small prototypes with limited complexity, and AI-generated code still requires substantial human processing in performance optimization, cross-platform adaptation, and release workflows. Treating AI coding as an accelerator rather than a replacement is the more pragmatic mindset.

Codex and Claude Code are two representative types of AI coding assistants. Codex was originally a code generation model developed by OpenAI and later evolved into one of the underlying technologies of GitHub Copilot, excelling at completing code based on natural language comments or function signatures. Claude Code is a command-line tool launched by Anthropic that allows developers to interact directly with the Claude model in the terminal to complete coding, debugging, and refactoring tasks. What both share is: they accept natural language instructions and generate runnable code snippets or complete files, but output quality depends heavily on the clarity of prompts and the developer's ability to review results. The creator using them to develop games essentially means using "describing gameplay rules" as input, letting the model generate basic code such as game loops and collision detection, and then manually adjusting details. This workflow is very efficient for small prototypes, but for parts requiring fine feel or complex state management, the developer still needs to intervene repeatedly.

One-line Summary

A Bilibili creator built multiple game prototypes with Codex and Claude Code, showcasing the practice and limitations of AI coding in indie games.

Paragraph Summary

A Bilibili creator shared a series of game projects built with Codex and Claude Code, including classic recreations like Tetris, Liar's Dice, and Connect 4, as well as original experiments like a rhythm game and a fighting game. The article argues that starting with classic games with clear rules is an efficient path for AI coding, and adding differentiated features on top of recreations (such as Tetris setup practice mode) can enhance a project's distinctiveness. The prototype state of the rhythm game and fighting game reveals the real workflow of AI coding: quickly building a skeleton, but gameplay feel and debugging still require manual iteration. The article concludes with practical advice on subject selection, differentiation, accepting imperfection, and drawing from personal experience, and notes that AI coding should serve as an accelerator rather than a replacement, with performance optimization and release workflows still depending on the developer's own abilities.