# Multi-instance and multi-pass review

*Claude Certified Architect - Foundations Study Guide · Prompt Engineering & Structured Output · Last updated October 22, 2026 · https://www.certifyforce.com/study-guide/claude-certified-architect-foundations/prompt-engineering-and-structured-output/multi-instance-and-multi-pass-review*

Why a session reviewing its own work misses errors, how to give an independent reviewer only the artifact and the criteria, and how verification passes, rubric-based evaluators and multi-instance voting turn review findings into reliable, routable signals.

- Certification: Claude Certified Architect - Foundations
- Exam section: Prompt Engineering & Structured Output
- Study guide: https://www.certifyforce.com/study-guide/claude-certified-architect-foundations

**Official exam objective:**

> Design multi-instance and multi-pass review architectures

## What the exam expects

- Explain why a model reviewing its own output in the same session misses errors, and why self-review instructions or extended thinking do not fix it.
- Use a second, independent Claude instance (a separate session, reviewer subagent or separate API call) that receives only the artifact, the requirements and the criteria.
- Separate writers from testers, and scope a reviewer to correctness and requirement gaps.
- Add a verification pass that checks each finding and records confidence and evidence so findings can be routed.
- Use a separate evaluator instance with a rubric, and choose aggregation thresholds when several independent reviewers vote.

Key terms: **self-review** (the generating session checking its own work); **independent review instance** (a reviewer without the generator's reasoning context); **verification pass** (a step that confirms or drops candidate findings); **LLM-as-judge** (a separate instance grading output against a rubric); **voting** (running several reviewers and combining their flags).

## Key facts

| Fact | Detail | Source |
|---|---|---|
| Self-review limitation | The generator retains its reasoning and is less likely to question its own decisions | [Exam guide](https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542750%2FClaude+Certified+Architect+%E2%80%93+Foundations+Exam+Guide.pdf) |
| Independent beats extra thinking | Independent instances catch more than self-review instructions or extended thinking | [Exam guide](https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542750%2FClaude+Certified+Architect+%E2%80%93+Foundations+Exam+Guide.pdf) |
| What a fresh reviewer sees | Only the diff and the criteria you give it | [Best practices for Claude Code](https://code.claude.com/docs/en/best-practices) |
| Reviewer subagent hand-off | Gaps return to the implementing session for fix and re-review | [Best practices for Claude Code](https://code.claude.com/docs/en/best-practices) |
| Subagent context | Own context window, system prompt, tools and permissions | [Create custom subagents](https://code.claude.com/docs/en/sub-agents) |
| Writer/test split | One Claude writes tests, another writes code to pass them | [Best practices for Claude Code](https://code.claude.com/docs/en/best-practices) |
| Reviewer scope | Flag only gaps affecting correctness or stated requirements | [Best practices for Claude Code](https://code.claude.com/docs/en/best-practices) |
| Code Review stages | Find, Verify, Rank, Post | [Code Review](https://code.claude.com/docs/en/code-review) |
| Chained review | Each step a separate API call you can log, evaluate or branch on | [Prompting best practices](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices) |
| LLM judge format | Single call, 0.0-1.0 scores plus pass-fail, most consistent with humans | [How we built our multi-agent research system](https://www.anthropic.com/research/multi-agent-research-system) |
| Voting thresholds | Different vote thresholds balance false positives and negatives | [Building effective AI agents](https://www.anthropic.com/research/building-effective-agents) |

## Check yourself

- Why is a separate reviewer call that receives the generator's transcript not truly independent?
- When should a team choose a reviewer subagent over a second terminal session?
- What does Code Review's verification step do?
- Why might a team flag on any single reviewer for payment code but require a majority for style?

## Sources

- [Best practices for Claude Code](https://code.claude.com/docs/en/best-practices)
- [Claude Certified Architect – Foundations Exam Guide v1.0](https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542750%2FClaude+Certified+Architect+%E2%80%93+Foundations+Exam+Guide.pdf)
- [Code Review](https://code.claude.com/docs/en/code-review)
- [Create custom subagents](https://code.claude.com/docs/en/sub-agents)
- [Prompting best practices](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices)
- [How we built our multi-agent research system](https://www.anthropic.com/research/multi-agent-research-system)
- [Building effective AI agents](https://www.anthropic.com/research/building-effective-agents)
- [Reduce hallucinations](https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations)

## About the full unit

The "Core concepts", "Exam traps" and "Worked scenario" sections of this unit, its 11 linked practice questions, and unlimited timed mock exams are included with the Claude Certified Architect - Foundations pack (USD $19): https://www.certifyforce.com/practice-exams