The LLM is basically a runtime that needs optimized input bc the output is compute bottlenecked. Input quality scales with domain knowledge, specificity and therefore human time input. You can absolutely navigate an LLMs attention piecemeal around a spec until you build an optimized input.