ad-pipeline-skeleton

Follow four acts: DECIDE (write a one-page brief, then multiple genuinely different concepts), MAKE (copy first, then stills, then video), DISTRIBUTE (plan placements, launch, operate daily), and LEARN (data plan before launch, test on small budget, measure and feed back). Four human gates sit before irreversible spend.

Settle each decision in the cheapest medium that can settle it: copy settles the message, stills settle the look, and video settles the timing. Generate video from the approved still and describe the movement rather than re-describing the picture, or the model reinvents the approved look.

Clear the concept before any production spend, since an unlicensed face, voice, song, or claim is a liability the moment it exists. Then clear the finished files again before publishing, because what got made is never exactly what was pitched.

Run an automated pass against criteria written before seeing the output, and do not show the judge model the prompt, since a judge holding the prompt grades prompt-matching rather than ad quality. Cap retries based on the requirement's success rate, then change the method rather than re-rolling the seed.

Rework usually means a gate sits after the commitment it was meant to protect. Audit where issues are caught, what was already paid for, whether the prior gate produced a filled artifact, and whether a later stage re-described an approved ad instead of using it.

Name a counterfactual: a holdout, geo split, or matched market tells you whether the campaign caused results, while a race between your own ads only tells you which ad won. Prefer lift measurement over last-click attribution, and measure the guardrail metric alongside the winning number.