Practice — Advanced Prompting Techniques (5 questions)

Your team built a decomposed pipeline for turning a legal contract clause into a structured risk assessment: step 1 extracts the parties and obligations, step 2 classifies the clause type, step 3 assesses risk given steps 1-2's output. On your 60-clause golden set, the single-prompt baseline (one call, CoT, no decomposition) scores 81% "risk level matches human reviewer." The decomposed pipeline scores 74% — worse — while costing roughly 3x the tokens per clause.

1. Explain, mechanistically, how a decomposed pipeline can score worse than a single well-prompted CoT call on the same task.
2. Describe a concrete debugging process to find where the pipeline is actually losing accuracy, rather than guessing which step is at fault.
3. Given what you find, is the fix "add self-consistency to each step," "go back to the single-prompt baseline," or something else? Justify your answer.