A practical exploration of prompt engineering, visual fidelity and iteration—through the recreation of Bounce Tales.
Conducted by Er Harshvardhan Purohit during the Prompt Engineering Master Class at SIN School of AI.
Explore the four builds Read the findings
The educational setting for this project and the AI Agent Xcelerator program. The learning approach demonstrated here connects prompt design with practical creation, inspection of outputs, corrective feedback and a documented research record.
The supplied institutional report identifies SIN School of AI, Gayatri Innovation Center, as a Center of Excellence in AI and Management Studies.
Explore SIN School of AI ↗The technology organisation identified in the supplied institutional report and the website hosting this research blog. This page presents the educational experiment under SIN School of AI branding, with SIN Technologies providing the public website context.
Visit SIN Technologies ↗Founder and CEO, SIN Technologies
Director, SIN School of AI Innovation Centers India
Instructor and research lead for this classroom experiment. Leadership designations follow the supplied MODELTRUST institutional report.
The task was to recreate Bounce Tales with at least five levels in one HTML file containing HTML, CSS and JavaScript. The exercise forms part of the program's classroom research into how prompt structure, reference images, access limits and corrective feedback affect the outputs obtained.
Use the supplied game images and a fixed output requirement: a single browser-game file.
Compare visible design with the reference and assess how the game behaves. Appearance alone cannot verify playability.
Argon's initial movement was too fast according to the instructor. A corrective second attempt produced the preferred result.
Gemini Argon → ChatGPT Sol 6.1 High → Claude → Qwen
Argon was preferred after feedback. ChatGPT placed second. Claude produced visible gameplay but free-plan limits constrained iteration. Qwen's supplied capture shows a blank playfield with a visible HUD. DeepSeek is recorded separately as an unsuccessful attempt; no build was supplied.
This is exploratory classroom research, not a controlled benchmark or a universal model ranking. Labels follow the classroom account. No numerical similarity scores were measured.
Select a build, then click inside it to use its own controls. Switching builds restarts the game. Qwen is retained as supplied, including its observed blank-output issue.
Keyboard controls belong to each original game. Use its on-screen instructions. The wrapper does not alter the model code.
Preserved verbatim, including spelling and wording.
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
Context, Role, Action, Format, Tone. The reference game provides context; the prompt assigns a developer role, specifies creation and requires one HTML file. Tone is not explicitly stated.
Show examples, Split into steps, Structure the input, Specify limits. Images show the target, while five levels and one file set concrete limits. The original prompt itself remains a single paragraph.
Try, Evaluate, Sharpen, Track. Run the prompt, identify the issue, make a targeted correction and retain the evidence. The exact Argon correction prompt was not supplied.
These are original user-supplied screenshots. They document visible output at capture time; they do not establish completed levels or tested physics.
The instructor preferred its visual resemblance and the second version after movement-speed feedback. This is the recorded classroom assessment.
No. Each file is embedded as its original bytes and decoded into an isolated frame. Only this surrounding landing page has been redesigned.
No. Five levels were requested, but an independent full playthrough was not completed. The original artifacts are available here for inspection.