Best-of-N generates multiple complete responses to the same prompt, then selects one against predefined criteria. Use it when candidate quality differences matter and one final response is needed—not an average or vote. section 8.5 section 8.7
Run the selection workflow in this order:
- Define the scoring criteria before reading any candidates to reduce anchoring on the first response. section 8.5
- Generate N complete candidates from the same prompt and requirements. section 8.5.1
- Score each candidate against the criteria, select one final response, and archive the rubric for auditability. section 8.5.1
- Choose another technique when the output is a label, a single numeric score, or a low-consequence draft that will be heavily edited. section 8.5
Executive deployment update
For a failed production checkout deployment, generate three complete updates under the same prompt. The prompt sets a 120–160-word limit, targets senior engineering leadership, calls for a calm and action-oriented tone, and requires impact, mitigation, and next-update timing. Score the candidates against predefined criteria before choosing the final update. section 8.5.1
