Overview
A Wordle-inspired daily alphabet game. The player has 5 minutes to get a word for each letter of the alphabet based on the day's category. Once completed, the user submits their answers to be marked. To speed up performance, common answers were hardcoded in JSON while the unique answers were sent with a carefully written prompt to an LLM to be marked.
Tech Stack
TypeScript
React
TailwindCSS
Node.js
PostgreSQL
Key Decisions
1.
Cheating Prevention
A few of my testers made it their goal to game the system. They succeeded, which was actually necessary for me to optimise the system prompt. Taking into account edge cases and certain rules, the marking felt much fairer and led to less cheating across the board.
2.
LLM Model Choice
The cheaper models (i.e. Claude Haiku) would consistently make mistakes, I didn't want users feeling frustrated that their answers were marked as false negatives. Claude Sonnet cost way too much but performed much stronger. Balancing the pros and cons of both, I chose Claude Haiku with a strong system prompt. This reduced most of the false negatives, while saving me money!