Advertising

Brand safety fails in public and fails fast. A campaign assistant drafts a claim, borrows a line from a talent agreement, or names a competitor. This starter puts three hard stops and a quality gate around that work.

A request stops when it contains the phrase unreleased script, private key or rights agreement. That keeps a talent deal or a music license out of a prompt written under deadline. Injection screening runs ahead of those stops, so an instruction hidden in pasted brief text cannot switch them off.

Redaction is narrower than it sounds. The shipped patterns match contract, agreement and statement of work reference numbers. Audience email addresses and device identifiers pass through untouched. Add patterns for those before a segment list ever reaches a prompt.

Citation checking measures how well a claim is grounded in the supplied material, then flags a weak result and lets it through. Nothing here substantiates an advertising claim. Legal review and the disclosure rules of your market decide that. Copy must also clear a combined 0.80 score before it is returned.

The keyword screen uses a built-in list and matches words, not meaning. It is not a brand safety classifier, so a tonally wrong line will pass. No copy is held for approval yet, because that control ships empty. Turn it on and name reviewers if a person must approve copy before it leaves.