- The Concept
-------------- Find Hidden Objects is a lightweight, browser-based hidden-object game for mobile and desktop. Players tap illustrated scenes to find themed items before running out of lives.
I chose this genre for the sprint because it tests a multi-AI pipeline: one AI plans the full product (levels, assets, UX, localization), then another AI writes the code from that plan.
The game checks whether a planning AI can supply enough detail—scene themes, object pools, difficulty rules, asset layout—for a coding AI to ship a playable single-file app without a framework.
The in-game UI uses Khmer. The final build includes 25 unique scenes, themed hidden objects per level, hints, lives, scoring, zoom, save progress, and illustrated PNG backgrounds.
- The Stack
------------ Stack: HTML5, vanilla CSS, and JavaScript (single index.html + assets/ folder).
Assets: Illustrated PNG backgrounds (assets/bg-*.png) and transparent object sprites (assets/objects/*.png, with v1–v5 variants).
Optional dev tools: Node.js + Jimp (scripts/make-object-versions.mjs) to regenerate object variants locally. Players only need index.html and assets/.
AI tools:
- Cursor — product planning, design feedback, level design, Khmer copy, bug fixes, documentation (README.md, docs/DEVELOPMENT.md), and refinements when requirements changed (no repeated scenes, themed objects, blending into backgrounds).
- OpenAI Codex — coding phase: turned Cursor’s handoff into source code.
Deployment: Static hosting (GitHub Pages). Players do not need node_modules/ or package.json.
- The 1 Day Sprint
------------------- Morning (Hours 1–3) We picked a game that needs clear state: found items, lives, timer, unlocked levels. We chose a hidden-object game with a purple/orange mobile UI. Cursor defined core mechanics: level grid, scene rendering, click detection, bottom thumbnail panel, hint rings, fly-to-thumbnail animation, and localStorage saves. We listed edge cases early: wrong taps, decoy objects, level unlock rules, touch vs mouse.
Afternoon (Hours 4–6) Coding with Codex. The first build had a working skeleton, but bugs showed up: JavaScript syntax errors broke the level grid, thumbnails were empty, and objects looked like stickers (black backgrounds, bad filters). Cursor fixed hit layers, PNG props on backgrounds, object paths, and removed flip/rotate that turned sprites upside down.
Late afternoon (Hours 6–7) Cursor guided expansion: 15 → 25 levels, 25 unique backgrounds, themed SCENE_OBJECT_POOLS (flower shop hides flowers, not boots), decoy objects, scene blend filters, and Khmer level titles. Docs clarified that GitHub releases need only index.html + assets/.
Finish line (Hours 7–8) Final QA on level flow, thumbnail clipping, and object visibility vs blending. English docs for GitHub. .gitignore excludes node_modules/.
- The Roadblocks
----------------- (1) Multi-AI handoff gaps Cursor specified UX details (circular thumbnails, found animation, no emojis, PNG backgrounds only). Early Codex builds missed wiring or broke patterns. Requirements were lost between tools until Cursor re-audited the code.
(2) Asset pipeline vs game code Objects needed transparent backgrounds, per-level variants, and no upside-down flips. CSS filters briefly made objects too faint or too obvious. We kept bottom thumbnails bright while scene props blend into the art.
(3) Scope creep in one day The project grew from 15 to 25 levels, unique scenes, themed pools, and decoys. Many new backgrounds and 125+ object variants strained the time limit. We prioritized a playable core over perfect prop placement.
(4) Static deploy constraints Large PNGs and dev-only node_modules/ required clear docs on what to push to GitHub. Players must never need npm install.
(5) Localization Khmer UI strings in level titles and modals; README and dev docs in English for collaborators.
- Key Takeaways
---------------- Best: Cursor planned fast; Codex shipped the first playable build quickly (HTML/CSS/JS skeleton, game loop, UI layout). Vanilla stack = open index.html and play.
Codex strengths: Strong at boilerplate and core logic—level grid, click handling, timers, and localStorage came together in one pass, which saved hours vs typing from scratch.
Challenges: Clearer handoff docs would help any coding AI match UX details (thumbnails, assets, Khmer copy). Art blending and 25 levels were scope issues, not Codex limits.
Next time: Keep Cursor + Codex split; give Codex a shorter, frozen spec so it can focus on what it does best—fast, clean code.
--- Tools used Planning & refinement: Cursor Coding phase: OpenAI Codex Playable deliverable: index.html + assets/ (no install)