Every day I see someone bolt an LLM onto something a shell script did better. The best AI practitioners I know are the ones who draw the line early: - Deterministic task, fixed rules? Script it. - One-off analysis with judgment? Ask a human or a cheap model. - Open-ended, branching, context-heavy? Now AI earns its keep. The $0 automation stack I run uses AI for exactly one step (summarizing news) and plain code for everything else. That's the whole secret: AI where it compounds, code where it ...