Vibe Coding a Login Page and an AI Tetris Player

I explore what Andrej Karpathy called vibe coding: describing what I want and letting an AI write the code, often without inspecting its changes. His example concerned a throwaway project; this experiment does not establish that unreviewed code is appropriate for production.

I use Windsurf, Superwhisper and initially o3-mini to build a login page. The assistant creates a Flask project, and I follow its terminal instructions to set up the environment and run it.

When the environment produces errors, I paste the messages back and accept the proposed changes. A basic username and password screen appears, but seeing a form does not verify that authentication is secure or complete.

I request Google authentication, encounter missing packages and environment issues, and eventually get the app running again. The Google button still does not appear, showing that generating code is not the same as completing the requested feature.

I ask the assistant to fix the missing button and switch to Claude 3.5 Sonnet when I want to share a screenshot. A button appears after further changes, although the page still has a broken image; the video does not validate the complete login flow.

Next I ask for a Python Tetris game. After installing requirements and starting the program, I verify that a piece can complete a line, then request a preview of the next piece.

The preview is missing on the first run, so I describe the problem and request another fix. The next version shows the preview and also adds a score, which I check by clearing a line.

I add pause and end-game buttons, but their positions overlap the board and the pause cannot initially be reversed. I report the issues and eventually copy and adjust suggested code manually, despite the goal of minimizing direct coding.

I ask for an AI that learns to play. The first result is a random-move placeholder rather than a learning system, so I switch models and ask again. The assistant proposes a deep Q-network using Stable Baselines3.

Training initially fails because of an import issue and then an observation-shape mismatch. I paste the errors back for fixes; after another run, training messages and rewards begin to appear.

After training, the game still has integration problems, including a circular import and a failure to locate the saved model. Further prompting gets the assistant to find the saved file and add an AI play mode.

The agent finally plays, although it performs poorly and loses. I am impressed by the prototype, but the recording also shows troubleshooting, manual edits and incomplete verification rather than a flawless one-prompt build.