What You Need to Know
Reviewing an output means checking what Claude wrote before anyone relies on it. A review asks two questions. Accuracy asks whether the specifics are right: numbers, names, dates, titles, quoted figures, and references to documents or rules. Completeness asks whether anything is missing. Both are checked against the source, the material the output was built from.
A bakery owner attaches a flour supplier's new contract and asks Claude for a summary. Her business partner will decide from it whether to sign. The summary reads well but could carry a wrong price per sack or leave out a clause.
The Reviewer's Stance
Treat every Claude output as a strong first draft from a brilliant colleague who can be confidently wrong. The Help Center says why: Claude "can occasionally produce responses that are incorrect or misleading", and users "should not rely on Claude as a singular source of truth". Fluency and confidence are properties of the writing, not evidence about the facts. Claude writes fluently whether it is right or wrong, so an output that reads beautifully has passed no check. The bakery summary would read just as well with a wrong price.
Ownership settles who does the checking. The professional who sends Claude's work owns it, in front of their team, their members or their funder. Review is the step that turns a draft into work you can put your name to. It is no optional extra for the cautious.
Accuracy: Are the Specifics Right?
Check the specifics first, ordered by consequence: the details that would be most costly or embarrassing if wrong. That usually means the figure a decision rests on, the name at the top of the letter, and the date of the deadline. A draft with twenty accurate sentences and one wrong subtotal is a wrong draft if the meeting decides on that subtotal. At the bakery, the price per sack comes first, because the partner decides on it.
Completeness: Is Anything Missing?
Completeness is harder, because the evidence is absent rather than wrong. A fluent summary can drop a section, a theme or a caveat, and nothing in the prose shows the gap. So check it against the source itself, never against your memory of it. Walk the source's own structure: its sections, funds, themes or line items. Confirm each element is either in the output or knowingly excluded. Reading the output alone and nodding is how omissions ship. The bakery owner ticks each clause in the contract against the summary.
Run accuracy first every time: a wrong specific is the most likely failure and the quickest to catch. Completeness comes second, since it needs the source open anyway.
Verify Against the Input
When an output is built from attached material, the first check is the attachment. It's right there in the conversation, and comparing the two costs minutes. Anthropic's prompting guidance points the same way for long documents: ask Claude to quote the relevant parts of the source before carrying out the task. In the chat, ask Claude where in the attachment each figure or claim came from. Then spot-check the claims that matter against those locations. An output whose claims trace to the source is checkable in minutes. One that cannot be traced gets no benefit of the doubt. The bakery owner asks where the price came from, then reads that clause.
Asking Claude whether it is sure is no check. The same model that made an error will repeat it with the same confidence, because no new information enters the check. The check that counts compares the output with its source.
In the exam
This domain carries more of the exam than any other, and every objective in it rests on the reviewer's stance. The certification's central claim is that its holder can be trusted with what Claude produces. That trust is earned at the review stage, not the prompting stage. With the stance right, the rest of the domain follows: spotting failure types, validating claims, knowing when to escalate. The exam treats self-reported confidence as a distractor, never as a check.
The register exemplar
The guide's Domain 2 sample item tests this objective. Claude summarises a regulation and cites a specific subsection. The keyed action is verifying that citation against the official text before the compliance team sees it. The named wrong answers are trusting Claude's expressed confidence and rewording the summary to sound more formal. Diligence beats confidence, and polish is never a substitute for correctness.
Exam Traps
Sending the output because it reads well or because Claude expressed high confidence
Fluency is a property of the writing and confidence is not a reliable accuracy signal, so neither has touched the facts. The keyed answer always compares the output against the source material before anything is circulated.
Asking Claude to confirm its own work and treating a confident yes as review
No new information enters the check: the same model that produced an error will confidently restate it. Self-check dressed up as review is one of this domain's favourite distractors, and it appears in stems about exactly the outputs that most need a real check.
Verifying one named figure while the stem hints at a dropped section
Accuracy checking alone misses the second review question. When a stem signals that source material had more in it than the output shows, the keyed answer walks the source's own structure for completeness, not just the details that caught the eye.
Practice Scenario
A scout group treasurer has Claude draft the annual financial summary for the AGM from the group's accounts spreadsheet. The draft reads perfectly, but she must be sure it is right and leaves nothing out, without taking Claude's word for it. What should the treasurer do before circulating it?
Build Exercise
Run the Two-Pass Review
What you'll learn
- How to check every specific in an output against the source material
- How to order accuracy checking by consequence
- How to walk a source structure to catch omissions
- How to make an output traceable by asking where each claim came from
- Take a document Claude drafted for you from real source material, a summary, a report or a data write-up, or generate one now from a spreadsheet or document you actually use
Why: The two-pass habit only proves itself against real work, where you know what the source contains and what a miss would cost.
You should see: A Claude draft and its source material side by side in one conversation.
- Accuracy pass: list every specific in the output, figures, names, dates and references, and check each against the source, starting with the ones a decision or reader depends on
Why: Ordering by consequence puts the limited checking time where an error would cost most.
You should see: A marked-up list of specifics: confirmed, wrong, or unverifiable.
- Completeness pass: walk the source’s structure, its sections, funds, themes or line items, and confirm each element is represented in the output or deliberately excluded
Why: Omissions leave no trace in the prose; only the source’s structure shows what should be there.
You should see: A tick against every element of the source, or a note explaining its absence.
- Ask Claude to indicate where in the source each remaining claim came from, and spot-check the claims that carry weight
Why: A traceable output can be verified in minutes; an untraceable one gets no benefit of the doubt.
You should see: Each claim tied to a location in the source, with the load-bearing ones confirmed.
- Record what the two passes caught and roughly how long they took
Why: The habit sticks when you can see its price and its payoff on your own work.
You should see: A short note: errors caught, omissions caught, minutes spent.