Bounce Tales: One Game, Four AI Builds
SIN SCHOOL OF AI / AI AGENT XCELERATOR
During the AI Agent Xcelerator program's Prompt Engineering Master Class at SIN School of AI, Er Harshvardhan Purohit conducted a practical experiment proposed by Shefali Agarwal: recreate the childhood game Bounce Tales as a single HTML file containing HTML, CSS and JavaScript, with at least five levels.
18-page official report • Original prompt, findings and evidence
The classroom outcome
The instructor's corrected order is Gemini Argon after a corrective second attempt, ChatGPT Sol 6.1 High, Claude, and Qwen last among the four supplied builds. DeepSeek did not produce a successful result in the reported session and is listed separately. This is an account of a classroom exercise, not a universal ranking of AI models.
Class participation
Priyanshu, Vanshika, Bhanu Pratap, Kanika Gehlot and Shefali Agarwal were present during the creation of the classroom experiment. Shefali Agarwal proposed the challenge. These attendees are credited here in the blog; attendance does not imply a separately verified research role for each person.
The classroom brief
The class supplied reference images and requested a playable 2D browser game. The following prompt is reproduced verbatim, including its spelling and punctuation:
You are expert in the 2d web based game development and now you have to develop one of the old childhood game named as the bounce tales and make sure to create atleast 5 levels in the same and create an 2d game with all the dynamics of the same I want the final output in the form of the html css js combined to one file of html For the reference I am prioviding you the images of the game first take a look at the game first and then start creation of the same
Classroom results and what they mean
Gemini Argon: preferred after correction
The instructor described the first result as visually very close to the reference, with the requested features present. However, movement was too fast. A second attempt following corrective feedback produced the classroom's preferred result. The supplied filename identifies this as the best Gemini Argon version. The exact correction prompt and the first version were not supplied, so the before-and-after speed change cannot be independently reconstructed from the available files.
ChatGPT Sol 6.1 High: second in the classroom assessment
The instructor placed this build second. Its supplied HTML file is included in the combined game selector. The ranking alone does not establish which individual mechanics were stronger or weaker, so no unmeasured feature scores or completion claims are assigned.
Claude: iteration constrained by free-plan limits
Claude placed third in the corrected order. The instructor used the free version and encountered usage limits during the first attempt. Its supplied screenshots show a green character, diamond collectibles, a ground platform and a pause menu. These differ visibly from the red-ball and egg reference. The exercise compares results obtained under these conditions and does not isolate model capability from iteration time or access limits.
Qwen: last among the supplied builds
The instructor corrected Qwen's position to last. The supplied screenshot shows a pale blue empty playfield with a Level 1 - Green Hills heading, 0/10 counter and 0:00 timer. No character or platforms appear in this capture. It documents a visibly incomplete result at that moment, but does not establish the technical cause. No numerical similarity score was recorded.
DeepSeek: unsuccessful attempt reported
The instructor reported that DeepSeek did not work for this exercise. No DeepSeek output file or error transcript was supplied. The available evidence therefore supports an unsuccessful session outcome, not a diagnosis of its cause.
How CRAFT shaped the task
The supplied Day 3 Prompt Basics presentation defines CRAFT as Context, Role, Action, Format and Tone. In this experiment, the childhood-game reference supplied context; the opening instruction assigned the role of a 2D web game developer; the action was to develop Bounce Tales; and the format required one HTML file containing the CSS and JavaScript. Tone was not explicitly specified in the quoted prompt. That omission does not invalidate a coding task, but it would be inaccurate to claim every CRAFT element was explicitly present.
The four S's in practice
The presentation defines the four S's as Show examples, Split into steps, Structure the input and Specify limits. The reference images gave the model a visual target. The requirement for at least five levels and a single HTML file specified concrete limits. The broader lesson is to separate the task, reference material and output requirements, then evaluate the first implementation before requesting targeted changes. The supplied prompt is one paragraph; it does not itself document a fully separated four-block workflow.
Why the second attempt matters
The deck's TEST framework means Try, Evaluate, Sharpen and Track. The Argon speed correction illustrates the importance of evaluating behavior after generation. A game can look convincing while moving at an uncomfortable speed. Recording the original prompt, observed issue, exact correction and resulting version would make this exercise easier to reproduce. Only the original prompt and the supplied final game files are available here.
Similarity to the supplied game reference
The supplied reference shows a red ball, green platforms and a blue-sky environment. Meaningful comparison should examine silhouette, colors, platform shapes, camera behavior, controls, movement speed, collision response and level progression separately. The instructor considered Argon's visual result especially close. No percentage similarity was measured, and a static image cannot establish that the physics or level design matches the original game.
Explore all four builds
The accompanying index.html embeds the four supplied HTML files in an isolated game selector. Their embedded bytes were checked against the originals. Selecting another model restarts that build. The wrapper adds navigation and classroom findings without rewriting the embedded game code. The public research lab is available at https://sinbouncytails.netlify.app/. Deployment availability was checked on 17 October 2026; full gameplay verification remains incomplete.
Questions readers may ask
Which model won this classroom exercise? Gemini Argon after corrective feedback, according to Er Harshvardhan Purohit's classroom assessment.
Was this a fair controlled benchmark? No. Access limits and iteration counts differed, and complete model identifiers, timings and transcripts were not supplied.
Does one HTML file prove five levels work? No. The prompt requires five levels, but runtime testing must verify progression, completion, restart and collision behavior.
Can the results be reproduced? The original prompt, reference images and four supplied artifacts provide a starting point. Exact model versions, settings and correction prompts would also be needed for a closer replication.
Sources and editorial status
Primary sources: the instructor's classroom account; four supplied HTML files; supplied game-reference images; five user-supplied output screenshots (Argon, ChatGPT, Qwen, Claude gameplay and Claude pause menu); and Agent_Xcelerator_Day3_Prompt_Basics.pptx, especially slides 5, 7, 9, 10 and 11. Model names follow the instructor's labels and have not been independently verified as vendor release identifiers. Promotional compositions use the screenshots as inputs; the original captures remain the evidence. No independent gameplay session was completed by the report preparer.
Original classroom output screenshots
Original evidence gallery: Argon, ChatGPT, Claude gameplay, Claude pause menu and Qwen. The landscape banners above are designed compositions based on these captures.
argon original classroom output screenshot chatgpt original classroom output screenshot claude original classroom output screenshot claude-paused original classroom output screenshot qwen original classroom output screenshot
About the institution and research lead
This experiment forms part of the AI Agent Xcelerator classroom research at SIN School of AI. Er Harshvardhan Purohit is identified in the supplied institutional report as Founder and CEO, SIN Technologies, and Director, SIN School of AI Innovation Centers India.
Comments