Common enough words like `function` and `class` are generally encoded as a single token by the tokenizer and may provide a slightly better context to the LLM. For openai you can test this stuff at https://platform.openai.com/tokenizer