WorldSmith
Our user is a beginner aged roughly 14 to 22 who wants to make games but has only a school laptop or an old PC. Today they hit two walls. If they try Unity or Unreal, they must install multi-gigabyte editors that need a strong GPU, then face a steep learning curve, and many quit on the first evening. If they use AI instead, they type one prompt, get a finished game, and learn nothing, so the next game is just as impossible. Worldsmith fixes both. It runs in any browser: a live 3D scene on top, a code editor below. The heavy AI work, modelling a 3D object from a description, runs on our servers, so the learner's device only draws a light low-poly scene and needs no strong hardware. A Socratic AI coach asks questions instead of giving answers, and the page itself, not the AI, decides whether a goal is met.
Preliminary score
- Value for the user18.5 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI21.0 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing15.0 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility9.5 / 15
Data requirements, running costs and a clear next step.
- Originality6.5 / 10
A meaningful difference from common solutions.
What the judges wrote
- Claude Opus 5.5Preliminary69/100
- Strengths
- Your user is sharply defined: a 14–22-year-old beginner on a school laptop who quits Unity on the first evening, or who uses one-prompt AI and learns nothing. The architecture fits that user well. Gemini does the heavy work (AI Forge text-to-object, Prompt mode, the coach) on your server, while the browser only renders a light three.js scene. Your testing is unusually honest and specific: 30 scripted cases, 0/24 answer leaks versus 5/30 for your rule-based coach, a 1.35 s median latency and Forge at 12/12 valid objects. You also report what did not work, including the AI coach scoring lower on concept-naming and crude organic models. Having the page itself, not the model, verify goal completion is a smart design choice that stops the AI from hallucinating success.
- To improve
- Your evidence of learning value is still internal: run even a small session with five to ten real beginners. Compare time-to-first-working-scene and concept retention against a video tutorial, so the outcome is measured rather than asserted. Add a per-learner running cost estimate for Gemini calls beyond the free quota of about 15 requests per minute, and explain how you would keep the coach usable at scale. Only Unit 1 (4 missions) is playable, so finishing one more unit would show that the coach and goal-checker generalise beyond the first lessons.
- GPT-6 AstraPreliminary72/100
- Strengths
- Your browser-based scene and editor address a specific barrier for aspiring game creators with modest hardware. Your separation of Socratic AI coaching from deterministic goal checks gives the learning experience a meaningful distinction from prompt-to-game generators. Your reported evaluation includes concrete metrics, a rule-based comparison, and candid limitations, including quota fallbacks and the absence of beginner testing.
- To improve
- Your next demonstration should show a complete lesson and Forge generation, including actual model responses, rendered results, and lesson completion checks. Publish representative evaluation inputs and outputs, and explain the different denominators in the AI-versus-rules comparisons. Test with genuine beginners against a tutorial baseline, measuring independent task completion and retained understanding rather than only valid objects or answer leakage. Add measured performance on a school laptop and an estimated paid inference cost per learner session.
The team
Nərmin AbdullayevaLead
developer
Senan Azemmedov
developer
Isa Jafarov
developer
Toprak Sümer
developer
Sama Valiyeva
developer