Stop Using AI For This

The current trend of using non-deterministic AI agents to perform deterministic tasks represents a significant inefficiency in software development. Many developers are embedding complex logic and "English if statements" into AI skills—such as Gary Tan’s GStack—forcing the model to re-process instructions and burn tokens for every execution. This approach ignores the fundamental principle of caching; instead of forcing an AI to interpret a "wall of code" repeatedly, developers should use AI to write standalone, deterministic command-line tools or programs. A video editing workflow serves as a prime example: using AI solely for transcript generation while leaving video stitching to stable tools like FFmpeg ensures reliability. Isolating AI to specific, non-deterministic sub-tasks prevents entire process chains from breaking due to a single hallucination, ultimately saving money, energy, and time while producing more consistent results.