The author noticed that their LLM workflow's system prompt has expanded significantly over time, as they repeatedly added instructions and examples to address undesirable outputs and edge cases. Consequently, every request now includes a large prompt block, even though much of it may be unnecessary for most runs.