How a team reviews AI drafts before they go out

The check

1. Against the request: every question answered, the format and length asked for, the right reader.
2. Every factual claim, one line each, with where it sits in the draft.
3. Sorted: supported (in the source given), checkable (needs an outside check, and where), suspect (specific but unsourced, or contradicts the source).
4. Sources: each cited study, article, URL or quote looked up; anything you can't confirm is "verify before use".
5. Promises, data and tone: dates, discounts or guarantees the company can't keep; personal or confidential data; claims like "best" or "100%".
6. Consistency: totals, names and dates that disagree.
7. Fix list and verdict: must fix, should fix, could fix.

A worked example

Illustration: a made-up business. An account manager asks AI to draft the results section of a renewal proposal from the account notes. Before it goes to the client, the shared review returns:

The account manager fixes the three must-fixes, checks the user count in the usage report, and sends it. Nothing in the draft was wrong in tone; everything that mattered was a fact.

Make it the team's habit

1. Write the rule down: anything that leaves the team is reviewed first.
2. Use one shared check, so every reviewer looks for the same things.
3. Give the AI the source with the draft; a claim can only be supported by something it can see.
4. A person checks every suspect claim and makes the call.
5. Keep the fix lists for a month and add a rule when the same mistake comes back.

Next step: the check as a skill

A check people paste gets skipped on a busy day. Making your first Claude skill shows how to turn it into a skill everyone's AI runs. For how to judge an answer in the first place, see LLM evaluation.

Questions

How should a team review AI-written work before sending it?

With the same check every time: does it answer what was asked, which factual claims are supported by the source, checkable elsewhere or suspect, whether any cited source might be invented, whether it promises something the company can't keep, and whether its numbers and names agree. End with a ranked fix list and a verdict: ready, ready after must-fixes, or redo.

Which AI drafts need a review?

Anything that leaves the team: client documents, proposals, published pages, anything with numbers, prices, dates, legal or policy statements. Internal notes for yourself can skip it. The rule is about where the text goes, not who wrote it.

Can AI review its own draft?

It can do the first pass with a fixed check, which catches most gaps and lists every claim to verify. A person still checks the suspect claims against the source and makes the call; the review shows them where to look.