An educational game & research project
What happens when learning becomes part of the game?
A Python game inspired by an English Bulldog, developed across three studies to explore math, player experience, and AI-assisted learning.
Independent research by Marcus Withers
Meet the project
Meet Coco
A real dog.
The start of an experiment.
Named after an English Bulldog and inspired by Number Munchers, Coco Crunchers began with a question: could a game help people practice math?
The project grew from a desktop prototype into a platformer, then an experiment with an AI hintbot. Each version asked something different about how people learn, navigate, and ask for help.
01 / The game
Original development recordings, from the first math game to a pixel-art platformer.
A screen recording of the pygame build: Coco moves through a pixel-art landscape, collects a bone, and encounters the math-problem interface.
The original desktop math game, with numbered spaces along a map and a moving Coco character.
These are development demos, not a playable browser release. The original recordings include the development environment.
02 / The development
The game changed as the research did. Each version carried forward a lesson from the last.
Math & learning
A desktop math game paired with pre- and post-tests for eight ninth-grade participants.
The first study compared scores before and after gameplay. Its small sample makes it an exploratory result rather than evidence that the same effect would hold across learners. Data was collected manually using Google Sheets.
Play & observation
A move to pygame and browser deployment, with attention to player accuracy and the friction of collecting data.
This iteration explored gameplay across ages. Collection remained manual, highlighting the need for in-game data capture. Getting a desktop Python game into the browser also required a different approach: replacing tkinter with a pygame-based build packaged by Pygbag.
AI & assistance
A comparison of players with and without an LLM hintbot. The observed completion-time difference was not statistically significant.
Twelve participants were split between a control group (five) and a hintbot group (seven). The assistant ran separately from the game and knew the level objectives, not the live game state. Small groups and widely varying completion times limit the conclusions.
03 / The research
Three studies exploring math through play: learning outcomes, performance across ages, and the role of an AI hint. These are exploratory findings, not proof of effectiveness.
Year 01 / Learning
A pre/post math assessment with ninth-grade students.
In this small pre/post study, the reported post-test mean was higher. The reported test gives p = 0.02459 with 8 participants; it does not establish that gameplay alone caused the change.
Source limitation. The available summaries disagree on the pre-test mean and size of the change. These bars reproduce the recorded aggregate scores; no percentage lift is claimed.
Year 02 / Age & accuracy
An exploration of how different age groups approached the game.
The exploratory study examined performance across age groups. This overall mean is reproduced as reported, not recalculated or independently validated.
Age comparisons withheld. The available summaries and underlying records have inconsistent participant counts and age-group totals. They need reconciliation before age-group comparisons or a definitive sample size can be presented.
Year 03 / AI assistance
Comparing completion times with and without Coco’s AI-powered hints.
No statistically significant difference (p > 0.05).
The Hintbot group had a lower observed mean, but variability was high in both groups. With 12 participants and p > 0.05, the study did not establish a reliable difference in completion time. Bars show means, not the range of individual outcomes.
Source note: figures reproduce the project’s recorded research summaries. They are not a reanalysis of participant-level data. Differences between source records remain unresolved; no cross-study effect or total sample is claimed.
04 / Under the hood
The hardest part was not choosing a stack. It was finding what worked for the game, the browser, and the study.
Flask and Django did not solve the desktop GUI constraint. Moving to pygame, then packaging with Pygbag, provided a workable path to browser deployment.
Three skill-focused levels introduced wall jumping, enemy combat and timing, and platform navigation. Tiled provided the visual level editor.
The hintbot ran separately from the game. It could explain level objectives but could not see the player’s current position or game state.
05 / Looking back
Coco Crunchers became a study of more than math. It exposed the practical challenges of browser deployment, unobtrusive data collection, and AI assistance that understands the game. The results point toward further testing, not a finished answer.
Explore the project source