Ad Creative Remake

Status

This skill is a planning contract, not a live performance-data connector. It becomes executable only when the user or an approved ads/analytics agent supplies campaign performance data with its time window, attribution method, spend, impressions, conversions, creative ID, and comparison baseline.

Mission

Use evidence about an underperforming video ad to design targeted creative variations. Change the smallest plausible cause—hook, opening frame, pacing, proof, presenter, offer framing, or CTA—while keeping a parent/variant graph. Do not call a video a failure from a single noisy metric and do not launch spend from this repository.

Required inputs

• the original ad and creative/version ID;
• performance data, source, date window, spend/impressions, audience, attribution, confidence or sample-size context, and comparison cohort;
• approved product facts, claims, offer, brand, talent, and source rights;
• target platform, ratio, duration, language, and variant count; and
• review owner, budget, experiment hypothesis, and stopping rule.

If performance data is absent, return a data-request checklist and do not invent a diagnosis. If the data is confounded by audience, spend, placement, or tracking changes, label the inference as a hypothesis.

Workflow

1. Lock the original as parent and summarize observed results separately from creative interpretation.
2. Identify the weakest supported stage: attention, comprehension, proof, trust, offer, or CTA. Record what evidence supports the diagnosis.
3. Build a small matrix of controlled variants. Keep product geometry, claims, permissions, and required disclosures fixed unless the hypothesis explicitly changes one of them.
4. Use MODELS.md and the live schema for any generation or edit. Use I2V/reference inputs when product or presenter fidelity matters.
5. Generate drafts, poll jobs, and run video QA for product, claims, captions, audio, framing, and synthetic-media risks.
6. Return a test plan with variant ID, changed variable, unchanged controls, prediction ID, output URL, and measurement hypothesis. A human or ads agent must approve any experiment or spend.

Output contract

Return the data provenance, observed facts, hypotheses, variant matrix, model/endpoint, parent/child links, outputs, failures, QA, and proposed measurement plan. Save only permitted receipts under .video/ following reports-and-receipts.md.

No metric prediction, legal claim, platform approval, or campaign launch is implied by a generated variation.