Review work — checking a contract against a playbook, scoring an RFP response, comparing two vendor proposals — suits Claude unusually well, because the ground truth stays in human hands. Claude flags; a person decides. But evaluation prompts fail differently from generation prompts: the default failure mode is a review that is polite, generic, and agreeable. The patterns below are about forcing specificity.
Give the reviewer a rubric, not a vibe
"Review this document" produces the LLM equivalent of "looks good, a few small suggestions." The fix is to make criteria explicit and force a verdict per criterion. Anthropic's core guidance — be clear and direct, as if briefing a capable new employee — applies doubly to review, where the interesting output is disagreement:
- Enumerate criteria. "Evaluate against these five checks: liability cap present, payment terms ≤ 45 days, …" Each check gets its own verdict.
- Define the scale. If you ask for scores, anchor every point: what does 2/5 mean versus 4/5? Unanchored scales drift toward the middle.
- Demand evidence. Every finding must quote or point to the passage that triggered it. Findings without locations are unactionable.
- License candor. Anthropic's docs advise explicitly requesting "above and beyond" behavior rather than implying it. For review: "Be blunt. A missed problem is far worse than a harsh finding. If the document is weak, say so plainly."
Structure: document on top, rubric at the bottom
Review inputs are long, so the long-context rules apply. Per Anthropic's best practices, put the document at the top of the prompt and the instructions and rubric at the end — placing the query after long content can improve response quality by up to 30%. For comparisons, wrap each document in its own <document> tag with <document_content> and <source> subtags, then reference them by name in the rubric ("for each criterion, state which of proposal_a or proposal_b is stronger and why").
For evidence you can verify, enable citations on the document block ("citations": {"enabled": true}). Claude then returns findings with a structured cited_text field tied to a location — character ranges for plain text, page locations for PDFs — instead of paraphrased recollections. Per the docs, cited text doesn't count toward output tokens, and it removes the "did the contract really say that?" round-trip. One constraint to plan around: citations cannot be combined with structured outputs in the same request (it returns a 400 error), so pick per call — cited prose findings, or schema-shaped scores with quotes as ordinary string fields.
Make scores machine-readable when they feed a process
If review results land in a spreadsheet or trigger a workflow, don't parse prose — request the scorecard via structured outputs (output_config with a JSON schema listing each criterion, verdict, and evidence string). Schema-constrained scoring also suppresses the classic reviewer drift where criteria are skipped or renamed between runs.
Keep generation and review separate
The strongest review pattern is organizational: never let the same conversation both write and grade the document. A model reviewing its own fresh output inherits the context and assumptions that produced the mistakes. Run critique as a separate request with a reviewer role in the system prompt — per the docs, even a one-sentence role ("you are a skeptical procurement reviewer; your job is finding gaps, not fixing them") measurably focuses behavior and tone. Instruct the reviewer to evaluate, not rewrite: "Do not produce a corrected version; list findings only." Otherwise Claude's helpfulness pulls it back into generation.
<example> tags) including at least one that reaches a negative verdict. A reviewer that has never seen a "reject" example rarely issues one.Finally, treat the reviewer itself as a system under test. Keep a small set of documents with known, human-agreed defects and check that prompt changes still catch them — the same empirical-evaluation discipline Anthropic's prompt-engineering overview recommends before any prompt tuning.
Where to go next
Review prompts pair naturally with self-critique patterns and confidence and hedging. For grounding findings in sources, see citations and sourcing; for building the test set, the evaluation framework.