Khla Si Ko Project Report
1. The Concept
Khla Si Ko (ខ្លាស៊ីគោ) is a lightweight, web-based reconstruction of the traditional Khmer tiger-and-cow hunt: four tigers versus twelve cows on a 4×4 board. I chose this game for a one-day sprint because it is a much stricter baseline than a calculator—placement, jumping captures, two win conditions, a bilingual UI, and an AI opponent all have to stay in one coherent state machine. The point of the sprint was to test a multi-AI pipeline: could one AI plan a traditional game from design notes, then hand that plan to another AI to ship a playable mobile-web duel?
2. The Stack
The Stack: React 19, TypeScript, Vite, Vanilla CSS, custom PNG art, and the Web Audio API. The AI Assist: ChatGPT-4o handled the initial product planning. It read the traditional rules, the Sakurai-style design notes, and the Khmer UI constraints, then generated a technical handoff document covering screens, phases, win conditions, and edge cases. Cursor / OpenAI Codex was then used exclusively for the coding phase, taking that markdown document and translating it into the React source, the 4×4 rules engine, the minimax AI, and the Khmer/English copy.
3. The 1 Day Sprint:
Morning (Hours 1–3): We kicked off by brainstorming small games that require strict state management, not just a pretty board. We settled on Khla Si Ko because the rules are compact but unforgiving: cows start off the board and are placed one by one; tigers hunt by jumping; tigers win by eating eight cows; cows win by trapping every tiger. We spent the next two hours prompting ChatGPT to map those rules into screens (title, mode, level, side, battle, result, how-to), bilingual labels, a 15-second turn timer, three hints per match, and a VS-AI path with Easy / Medium / Hard. The output was a strict markdown handoff for the developer. Afternoon (Hours 4–6): The coding phase. We fed the handoff doc to Codex. Screens and boilerplate appeared instantly—title art, gold/cyan buttons, a 1080×1920 phone stage—but the actual game logic was messy. Testing revealed a major flaw: the cows were already sitting on the board at kickoff, like a Western checkers setup, instead of waiting in the tray under the board to be placed one by one. Tiger jumps sometimes failed to remove the eaten cow, and hitting a win dumped the player onto a result stamp with no sense of how the hunt had actually unfolded. We spent these hours debugging the place/move phase, capture state, and whose turn it really was. The Finish Line (Hour 7-8): With time running out, we had to abandon extra motion, richer sound, and a harder AI rewrite to ensure the core hunt functioned properly. We spent the last hour running final QA tests—placement from the tray, orthogonal-only moves, jump captures, trap wins, Khmer/English switching, local two-player vs AI—polishing the painted buttons as much as possible, and pushing the repository to GitHub.
4. The Roadblocks
Roadblocks: The biggest roadblock was the "lost in translation" effect between the two AIs. ChatGPT planned a traditional Khmer hunt with an off-board cow tray, a smartphone-game layout (one screen, one job), and Khmer-first copy. Codex completely ignored several of those UX requirements. It generated a generic web-app board with every piece already placed, English-leaning chrome, and no working history of how cows entered or left the hunt. Worse, it missed the logic that after each cow is placed, a tiger must get a turn—so the opening phase felt broken. Because of the strict 7-hour limit, we had to pivot away from making every screen feel like a finished commercial game and focus entirely on hacking the TypeScript state so the rules would not crash, skip a side, or declare the wrong winner.
5. Key Takeaways
What was the part that worked the best? The structural planning phase. ChatGPT turned oral tradition plus the Sakurai notes into an architectural outline, screen list, and edge-case checklist in under 10 minutes: place vs move, eat-8 vs trap-all, diagonal ban, timer, hints, and three AI depths. That mapping would normally take a human designer a long afternoon. What was the part that worked the worst? Multi-AI handoffs and rule fidelity. Codex struggled to interpret the handoff doc perfectly, resulting in a Westernized starting position, a broken placement state, and a UI that kept drifting toward a website instead of a phone game. Heavy manual prompt-tweaking was required before the board played like Khla Si Ko instead of checkers.