Generate playable HTML5 games with Mistral Medium and GLM

Summary

1. 1.Initialize the client. The project environment requires the mistralai and python-dotenv packages, configured with an API key stored in a dot-env file. The Mistral client and its underlying HTTPX client are initialized with a timeout of 600 seconds, or 10 minutes, because GLM 5.2 generates large code outputs exceeding a thousand lines and requires more time than typical chat completions.
2. 2.Craft the game prompt. A prompt function structures system instructions and detailed game specifications. The system prompt instructs the model to build a self-contained HTML5 game using Canvas rendering without external CDNs or imports. The user prompt provides a high-level game description, such as a top-down procedural dungeon crawler, alongside explicit engineering requirements. These rules mandate spatial properties like x, y, width, and height for collision detection, numeric health properties for enemies, and dedicated key-state tracking objects so held inputs are not dropped.
3. 3.Generate the game. The script defines an extraction helper using regular expressions to retrieve clean HTML whether the model returns markdown code fences or raw doctype declarations. The generation function calls chat completion using the GLM 5.2 model identifier through the Mistral API and extracts the raw HTML string from the response.
4. 4.Review and fix the game. A dedicated review function passes the HTML code to Mistral Medium, prompting it to act as a senior QA engineer searching for runtime errors, undefined variables, and broken game logic. If issues are found, the fix function sends both the issues list and the original HTML back to GLM 5.2, requesting a corrected file. This review-and-fix cycle iterates up to five times until Mistral Medium replies that no issues remain.
5. 5.Edit the game. The script includes an edit function that reads an existing HTML file from disk and accepts a targeted user instruction, such as adjusting enemy speed or weapon damage. It sends the current game and instructions to GLM 5.2 to apply in-place modifications, bypassing the initial generation step entirely.
6. 6.Save and serve the game. Because modern browsers restrict certain features when opening HTML files directly from the filesystem, a helper function initializes Python's built-in HTTP server on port 8000. It writes the final HTML file to disk, launches the server, and automatically opens the game in the user's default web browser.
7. 7.Tie it all together in main. The main execution block parses command-line arguments to determine whether to run in generate mode or edit mode. In generate mode, the script executes prompt construction, code generation, up to five cycles of the QA review loop, file saving, and local server startup. In a live terminal demonstration, the script produces a functioning dungeon crawler featuring health pickups, enemy combat, and procedural room navigation.