Claim Safety Review Ad Creative

Claim Safety Review

Shared quality bar: ../references/output-standard.md. All numbers cited here live in ../references/thresholds.md. This is a pre-launch gate on wording. It flags exposure and routes restricted categories to the platform, and it is not legal advice or a policy ruling.

Use this skill when

Copy exists and has not launched. Run it on a batch handed over by an agency or freelancer, on copy generated by a model, on lines lifted from a sales deck or a case study, and on anything a client wants live today. Run it again whenever a line is edited after this review, because an edited line is an unreviewed line. It examines the ad text only and never opens the landing page; the destination belongs to ad-to-landing-page-match-ad-creative.

Required input

• Every line to be reviewed, exactly as it will run: headlines, descriptions, primary text, on-image or on-video text, and the call to action button.
• The platform each line runs on, since restricted-category rules differ.
• The country or countries targeted.
• Any source the user already holds for each factual line: a document, a dashboard, a dated study or a named customer.
• The advertiser's category and what the product actually does.

Analysis workflow

1. Split the copy into individual claim units. One line can carry two claims and only one of them may be a problem, so review at the claim, not at the asset.
2. Classify each unit: factual claim, comparison, superlative, testimonial, guarantee, or non-claim. Non-claims pass without a row and are counted so the review shows its own scope.
3. For every factual claim, apply claims.evidence_rule. Ask for the named source. A source is a document, a dashboard or a named customer, and an internal assertion that a colleague said it is not one. Unsourced units are marked [claim: needs source].
4. For every superlative, apply claims.superlative_rule. A qualified and cited superlative passes. A bare one is cut, and softening it into a vaguer boast does not fix it, since a vague unevidenced claim is the same defect with worse copy.
5. For every comparison against a competitor, check that the comparison is like for like, dated, and that the competitor's own current terms were checked on the date stated. An undated competitor comparison ages into a false claim without anybody editing it.
6. Screen every unit against claims.restricted_topics. Where a unit lands in one, do not issue a verdict. Route the reviewer to that platform's own policy page for the named category and country, and mark the unit as blocked pending that reading.
7. Check implied claims, not just stated ones: a testimonial that implies a typical result, an image of an outcome, an on-screen figure with no source, a guarantee implied by a phrase like risk free.
8. Return a per-line decision and a rewritten safe alternative wherever the idea survives without the unevidenced part. Never invent the missing evidence, and list every line you declined to rewrite.

Decision rules

• Factual claim with a named source supplied: ship.
• Factual claim with no source: rewrite to remove the number, or approval_needed if the user says the source exists but has not produced it.
• Bare superlative: cut under claims.superlative_rule. Offer one qualified alternative only if the qualification is something the user has evidenced.
• Comparison with no date or no like-for-like basis: rewrite.
• Testimonial with no permission on file: approval_needed under claims.testimonial_permission.
• Any unit in claims.restricted_topics: investigate, pointing at the platform's policy page. This pack issues no compliance verdict, ever.
• Guarantee, refund or outcome promise: approval_needed from whoever can honour it commercially.

Output format

Open with one line: is this batch launchable as written, yes or no.

Then the rewritten lines, then a short list of lines left unrewritten with the reason, then What this could not see, Missing data, Approval gates.

This skill reads wording, not a test and not a time series, so it issues no high confidence label and does not borrow the comparison family's conditions: test.min_runtime and test.volume_floor describe a test, and no amount of running time makes a claim sourced. What stands in its place is the evidence label on each line, which is the whole judgement here: a claim whose source was produced and read is ship, a claim whose source is asserted but not produced is approval_needed, and a claim with no source is needs_data until one arrives. A review of a batch is high in only one sense worth stating, and it is stated in the output: every line was reviewed at the claim rather than at the asset, and the count of non-claim units is given so the reader can see the review's own scope.

Practical example

Illustrative made-up copy, not a real advertiser. Six lines submitted. Line 2: "Cut your reporting time by 73%." Line 4: "The fastest onboarding in the category." Line 5: "Approved in 24 hours, whatever your credit score."

Reading: "No, not launchable as written." Line 2 carries a real number and no source, so it is [claim: needs source] and marked needs_data until the user names the dashboard or study behind the 73%. Line 4 is a bare superlative and is cut under claims.superlative_rule; a rewrite is offered only because the user supplied a dated onboarding time of 11 minutes against a documented alternative. Line 5 touches credit, which sits in claims.restricted_topics, so this review refuses to rule on it and routes the reviewer to the platform's own credit policy page for the target country, so it is neither passed nor rewritten and carries no verdict from this review.

All six lines are accounted for, one outcome each. Lines 1 and 3 are non-claims and pass unchanged. Line 4 is rewritten, on the dated onboarding time the user supplied. Line 2 is not rewritten and does not pass: it keeps [claim: needs source] and waits for the dashboard behind the 73%, because removing the number would leave a different claim rather than a safer version of this one. Line 5 is routed to the platform's policy page under claims.restricted_topics and is blocked pending that reading. Line 6 is left unrewritten because removing the unevidenced part leaves no idea behind. Two pass, one is rewritten, one waits on a source, one is routed out of this review's scope, one is refused.

Guardrails

• Never supply a source, a study, a statistic or a customer name the user did not provide.
• Never issue a compliance or legal verdict on a restricted category.
• Never soften an unevidenced claim into a vaguer one to get it past this gate.
• Do not upload or launch anything. This gate ends in a human approval.