Multi-Pass And Multi-Instance Review

Definition

Two review strategies for improving reliability: multi-instance (run the same prompt N times, then take majority/union) and multi-pass (run sequential passes with increasing specificity), often as a draft → review → refine chain.

Key points

• Multi-instance: same prompt run N× in parallel; aggregate by majority vote or union of findings — good for coverage/consensus.
• Multi-pass: sequential passes, each more specific than the last (self-correction pattern: draft → review → refine, each a separate API call so you can inspect intermediate outputs).
• Prompt chaining / self-correction stays useful even with strong models — it enforces a pipeline and exposes intermediate state.
• Fix inconsistency with examples, not prose: "inconsistent output → add 3–4 few-shot examples, NOT more prose" — see Few-Shot Prompting.

Why it matters for the exam

• The multi-instance vs multi-pass distinction is a named D4 concept; know that instances are parallel same-prompt and passes are sequential increasing-specificity.
• Feeds directly into review/reliability scenarios and the Opus Severity-Filter Trap fix (report all → filter separately).

Common gotchas

• Conflating the two: multi-instance = same prompt repeated; multi-pass = a chain of different, sharpening prompts.
• Reaching for "more prose instructions" to fix inconsistency instead of few-shot examples.

See also

• Few-Shot Prompting
• Opus Severity-Filter Trap
• Validation-Retry Loops

Sources

• docs-prompting
• netlight-study-guide