Engineering Turing's Dawn: Integrating AI Hints and Deterministic Engines in Web Games

Turing’s Dawn — race the dark before the longest day runs out

Turing’s Dawn is a browser puzzle game submitted to the June Solstice Game Jam. It implements six distinct codebreaking disciplines, including a functional Turing machine simulation.

Why This Matters

Integrating LLMs into puzzle games often creates a conflict between helpfulness and spoiler leakage. While standard LLM implementations tend to blurt out answers, this project demonstrates a constrained architecture where the model acts as an adaptive coach limited by authored hint ceilings, ensuring the technical challenge remains intact while leveraging generative AI for user experience.

Key Insights

• The cipher engine is built using pure TypeScript to ensure it remains free of React or DOM side effects, allowing for full unit testing via Vitest.
• A ‘traceTuring()’ engine manages machine configurations (tape, head, state, and rules) to simulate real machine operations rather than static descriptions.
• The hint system utilizes gemini-2.5-flash with a spoiler-safety architecture that excludes canonical solutions from the LLM prompt to prevent answer leaks.

Practical Applications

• Adaptive Hint Systems: Using LLMs for tiered guidance (reframe goal → name technique → point to step) while maintaining spoiler safety by withholding keys from the model.
• Data-Driven Game Design: Implementing puzzles as JSON entries rather than hardcoded logic to allow for scalable content addition without modifying core engine code.

References:

• https://youtu.be/kwdbez-77X8?si=WnSFTmtgGdSi5NRe
• https://github.com/pooja-bhavani/Turing-s-Dawn