Talk me to the moon

Inspiration

“Fly Me to the Moon” inspired two journeys: exploring space and learning to work with AI. To us, AI’s progress feels like a new moonshot—but we don’t want anyone left behind.

A 2026 Pew Research Center survey found that 51% of U.S. adults do not use AI chatbots. People have different reasons for not using AI; we want to help those who are interested but unsure where to begin.

What if their first step felt like playing a game?

What it does

Talk Me to the Moon is an English/Chinese browser game where players guide BOLT, an AI copilot, through six connected lunar missions:

Land: communicate a goal and fuel requirements.
Deliver equipment: provide missing context.
Navigate: inspect conditions before approving an action.
Sort samples: teach a classification standard through examples.
Audit reports: distinguish supported, contradicted, and uncertain claims.
Return home: test instructions across different situations.

Visible consequences, including crashes, encourage players to revise and retry. After each mission, a popup explains the skill they practiced and provides a complete everyday example.

How we built it

We built the game with React, Vite, a Node.js backend, and DeepSeek, and deployed it on Render. DeepSeek interprets conversational instructions, while programmatic rules determine fuel use, collisions, report approval, and mission completion.

Research informed our use of examples in prompts, systematic prompt testing, and evaluating and adjusting AI interactions. These studies guided our design; measuring the game’s learning outcomes remains future work.

Challenges we ran into

Our biggest challenge was realizing that completing a mission did not necessarily mean learning its intended skill. Development became a repeated cycle of designing, playing, questioning, and revising.

Our first mission gave away the answer. We originally labeled the fuel-efficient route. Players could simply tell BOLT to take it without explaining the actual requirement: arrive with at least 30 fuel remaining. We redesigned the information flow so players needed to connect their landing requirement with BOLT’s route information.

Players finished without knowing what they had learned. Our own playthroughs exposed this gap. We added post-mission explanations connecting each action to a prompting skill and an everyday application.

Correct answers were sometimes rejected. We improved validation to recognize natural wording and spelling mistakes, particularly when players evaluated reports. We also revised example prompts that referred to missing information, making them complete enough to try independently.

Accomplishments and what we learned

We built a playable bilingual expedition with six missions, sound effects, responsive layouts, and 61 automated regression tests.

More importantly, we learned to ask two questions: Can the player finish? And what does finishing actually teach? Working effectively with AI involves communicating requirements, checking evidence, and deciding when to approve an action.