A checker that does not fix: review AI ad outputs

Give the reviewer a different job

The author’s job is to produce a usable asset. The reviewer’s job is to determine whether that asset meets the brief and to show where it does not. Combining both jobs can make a defect disappear from the report before its cause is understood. Give the reviewer the required inputs and acceptance conditions, then ask for specific defects supported by the delivered files.

Anthropic documents separate subagent contexts and tool permissions. Use those mechanisms to give a reviewer readable outputs and a report destination, without authority to modify the submitted files. Independence is a working boundary, not merely a different role name in the prompt.

In a hypothetical blender demonstration, the brief requires the closed lid to remain visible. The author may know that the source footage contains it, but the final crop can still hide it. An independent reviewer checks the delivered interval and reports the actual obstruction with its output identity. The correction should address that defect without reopening unrelated accepted wording. Review independence matters because intended evidence and visible evidence are different things.

Start with the expected deliverable list. Count the actual files, confirm that each assigned asset has a result and identify missing or extra candidates. A summary that says “ten complete videos” should be checked against ten playable, correctly named outputs.

Then inspect what the files are. A renamed image is not a video; an audio file beside an MP4 does not establish that the MP4 contains audio. Record the inspection method so another operator can repeat it. This first pass is deliberately mechanical. It prevents a fluent explanation from becoming evidence of completion when the deliverables themselves are incomplete or misidentified.

Source: Turntake editorial checklist (proposed), October 2026.

Write findings with a location and consequence

A useful finding names the asset, version, location, observed defect and reason it matters. “Caption group at 12.4 seconds covers the product control” is actionable. “Captions look odd” requires the author to rediscover the problem and may lead to a different fix.

Attach a frame, media property or source excerpt where it supports the observation. Be explicit about uncertainty. If the reviewer cannot access the original voice track, it can report that sync was checked only against the rendered audio. Do not claim a stronger check than the available material permits. The evidence report should remain understandable without access to the reviewer’s private reasoning.

1. 01Read accepted briefEstablish testable requirements
2. 02Count delivered filesInspect actual media
3. 03Record located defectsKeep author files intact
4. 04Review revised versionRecheck neighboring content

Separate measurable gates from judgment

Some questions have exact answers: does the file have an audio stream, does its duration match the manifest, are all required labels declared? Other questions need judgment: does the metaphor explain the mechanism, is the transition distracting, does the voice suit the scene?

In our production in October 2026, measurable checks run as scripts, while ambiguous creative decisions receive a separate review. A machine pass does not settle editorial meaning. Conversely, an attractive cut does not excuse a missing layer. claude for ads describes the broader studio method. Keep the two result types visible so the author knows whether to repair a factual defect, a technical failure or a proposed style change.

Return defects instead of rewriting the submission

Leave the submitted files intact. The report goes back to the author with priorities and acceptance conditions for the repair. If the reviewer starts moving captions or replacing clips, the reviewed output no longer matches the original submission, and the revision history becomes harder to trust.

1. Flagged
2. Redone

An exception needs an explicit new assignment: a reviewer may become the repair author, but then another check should inspect that revision. The separation need not be bureaucratic. Even a small team can preserve an original, create a new candidate and record which role made the change. The essential point is that review evidence remains attached to a definite version.

Reviewer actionAuthor action

• Reviewer actionDescribe missing overlayAuthor actionRestore declared overlay
• Reviewer actionIdentify timing defectAuthor actionRevise cue timing
• Reviewer actionReport unsupported claimAuthor actionSupply evidence or revise wording

Check the repair on its own version

When the author returns a revision, compare the actual output with the finding. A note saying “fixed” is not the repair. Verify the original defect and inspect the neighboring interval for unintended changes. A caption moved upward may solve overlap while entering a placement overlay area.

Our October 2026 production practice includes viewing rendered frames after a mechanical change. Keep previous versions for comparison. Version control for video ads: know the approved cut explains approval history, and A video edit JSON recipe that survives revision describes the recipe that should travel with the new cut. Use those records to distinguish a corrected render from an unrelated replacement that happens to have a similar filename.

State the reviewer’s limits

A reviewer can inspect the inputs it received, not every fact about the product or account. If a claim has no supplied source, report missing evidence. Do not invent a supporting explanation because the surrounding creative looks credible. Likewise, a synthetic persona response cannot establish how real viewers will behave.

End the report with passed checks, defects, untested areas and the next owner. AI project memory: decisions in files, not chat explains where that report belongs in project memory so a later session can continue from the same evidence. Turntake’s model plans describes Turntake’s planned model connection; the checker method on this page is a proposed role design and a studio workflow, not a currently available automatic Turntake checker.

The deliverable should be small enough for the author to act on and precise enough for another reviewer to verify. That makes independence useful: it creates a second look at the actual asset without turning review into an undocumented second production pass.

Prioritize defects by their effect

Put an unsupported claim or missing required layer ahead of a proposed stylistic adjustment. The report can explain that a cut is technically complete while factual approval remains blocked. Do not use a single numeric score to conceal that distinction. A repair author needs to know which issues must be resolved and which are optional suggestions. When a suggestion is rejected, preserve the decision instead of relisting it as an unresolved defect in every subsequent review.

Who wrote this We made the ads first, for SmartSolar in Denmark, Sweden and Ireland. Then we built Turntake. Where Turntake comes from