English: A vs. An

This explains how to pick the English indefinite article for procedural text generation: use a_or_an(word) to decide whether to output a or an. The key is that the choice depends on the spoken initial sound, not the written first letter. Common counterexamples illustrate this: unicorn starts with the letter u but a consonant Y sound (so a unicorn), while hour starts with h but an initial vowel sound (so an hour). A simple letter-based check produces mistakes, so the function must consult pronunciation information.

To see how often letter-vs-sound mismatches occur, the author parsed the CMU Pronouncing Dictionary, built tree-style visualizations (including checks of whether the first one or two letters suffice), and analyzed 32,455 words. Only 129 words required explicit exceptions, a surprisingly small fraction. The write-up documents how those exceptions group, shows visualizations of letter-prefix determinism, and discusses implementation notes like trie simplification. The author also admits not using LLMs for the code work and reflects that an LLM might have saved time on some parts such as the trie algorithm and parsing/d3 work.