Play-test your own game
I have a solid workflow for making my game with AI. Task: The author developed a Node.js wrapper for the browser game 'Crossword Dungeon' to enable AI playtesting without modifying game files. This wrapper exposes game state via text-based rendering and synthetic HTTP events, allowing AI agents to navigate and interact with the game logic directly. The workflow involves the AI executing milestones, creating custom test fixtures for bugs, and validating fixes through terminal-based playtests, while the human reviews outcomes and identifies 'feel' issues.; tools: custom-built or unnamed agent. Human role: The original post does not say how a person stayed involved.; limitations: This approach relies on the game being text-heavy and turn-based; graphical or real-time games may not be suitable for this specific wrapper method. The AI cannot fully replace human judgment for subjective 'feel' or complex visual states, requiring manual verification. Token costs and execution time vary based on the complexity of the milestone and the AI model used..
Tools
custom-built or unnamed agent
Published on Agent Work Atlas:
Overview
I have a solid workflow for making my game with AI.
Workflow
The author developed a Node.js wrapper for the browser game 'Crossword Dungeon' to enable AI playtesting without modifying game files. This wrapper exposes game state via text-based rendering and synthetic HTTP events, allowing AI agents to navigate and interact with the game logic directly. The workflow involves the AI executing milestones, creating custom test fixtures for bugs, and validating fixes through terminal-based playtests, while the human reviews outcomes and identifies 'feel' issues.
Reported result
The AI completed Milestone 13, including five new special features, in approximately 12 minutes using ~120k tokens with Claude Code and Sonnet 4.6. While this automated the bulk of functional testing, the author still spent less than 30 minutes on manual playtesting to identify bugs and gameplay 'feel' issues, which the AI then fixed in a single batch.
Human role
The original post does not say how a person stayed involved.
Limitations
This approach relies on the game being text-heavy and turn-based; graphical or real-time games may not be suitable for this specific wrapper method. The AI cannot fully replace human judgment for subjective 'feel' or complex visual states, requiring manual verification. Token costs and execution time vary based on the complexity of the milestone and the AI model used.
Try it
I am developing a turn-based, text-heavy browser game called [Game Name]. I need you to create a Node.js wrapper that exposes the game state via a text-based renderer and accepts synthetic HTTP events for interaction, without modifying the core game logic files. First, implement the wrapper and verify it works by navigating to a specific room configuration. Then, take the following milestone: [Milestone Description]. For each feature in this milestone, write a minimal test fixture to reproduce potential edge cases, playtest the feature via the wrapper to confirm functionality, and fix any bugs found. Do not proceed to the next feature until the current one is validated. Finally, provide a summary of the tests run, bugs fixed, and any remaining issues that require human review for 'feel' or subjective quality. Ensure you use [AI Tool/Model] for this task.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.