---
name: magic-hour-ad-variants
description: Make finished video-ad hook, angle or offer variants while reusing approved footage. Use when Magic Hour is requested or configured for ad variations or creative iterations, not for managing ad spend or publishing campaigns.
license: MIT
metadata:
  author: magichourhq
  version: "1.0.0"
---

# Turn an approved ad into useful variants

Keep the accepted parts. Change the intended creative variable, finish each requested file and retain what changed so the next revision does not start over. An ad is a reference, not a proven winner, unless the customer supplies performance evidence.

> Make three new openings for this approved product ad. Keep its body, voice and offer. Show a truthful product action, a problem/solution angle and a different buyer context. Choose the treatment, use my existing budget, and deliver clean and captioned versions with a record of each change.

## Read the source before planning

Inspect the actual video with sound, its transcript where available, clean master, product references, exact offer/CTA, intended audience and destination. Use supplied customer wording and product facts; do not invent quotations, results, prices or urgency. A competitor's ad can inform structure, pacing and an original angle, but its footage, person, distinctive script and claimed results are not reusable permission or proof of profitability.

Record the source hook, angle, proof, body transition and CTA. Find cuts from playback and actual words or actions, never from a filename or guessed transcript timing. Keep the original master unchanged. If only a captioned master exists, recover the clean source when possible; disclose remaining burned text instead of placing incompatible captions over it.

## Choose concepts or iterations

| Request | Change | Keep comparable |
| --- | --- | --- |
| Explore a new concept | Buyer context, message, demonstration or format | Grounded product facts and the intended offer unless changed explicitly |
| Test a new opening | First visual, spoken hook and its matching overlay | Accepted body, voice, offer and placement |
| Revise an offer or CTA | Authorized copy and its dependent speech/layout | Unaffected accepted footage |

Use the requested count and budget, not a default batch. For a single-variable iteration, state the variable and avoid incidental changes. A new concept may deliberately change several elements; label it accordingly. Do not claim that an ad platform will classify two files as different creatives.

Write a compact plan: variant name, concept or iteration, changed variable, opening visual, spoken line, exact overlay, source body interval, transition and required new media. The visual, speech and text should express one coherent promise that leads into the body. Useful openings include a supplied question, a visible product action, a truthful comparison or the payoff first. Do not stage fabricated product protection, medical results or personal experience as evidence.

## Generate only missing media

Use the existing Magic Hour creation connection at `https://mcp.magichour.ai/`, or the project's authorized API integration and [live schema](https://docs.magichour.ai/api-reference/openapi.json). Skill installation does not authenticate the account. Discover the supported model, references, duration, resolution, audio behavior and account eligibility before promising an output. Never redirect work to another provider implicitly.

Reuse approved clips when editing, copy or sound alone fulfills the change. For a genuinely new opening, use an accepted product/presenter image directly with image-to-video when it already meets the brief. Edit the image only when the requested scene or composition needs it. Use new text-to-video only when an exact reference identity is unnecessary. Keep source product details and the transition into the accepted body consistent.

New speech needs an accepted recording or an available authorized Voice Generator/Voice Cloner route. Talking Photo animates a still with audio; Lip Sync changes speech on existing speaker footage. Neither synthesizes speech from text. If speech is not exposed in this client's tools, use an existing authorized API integration or identify the missing recording. Do not invent a voice preset or promise native dialogue from a video prompt. Faceless voiceover needs composition, not a lip-sync job.

Check the complete requested batch against the existing spending limit, including speech, images, video, captions and permitted repair allowance. An estimate is not the eventual charge. Do not buy credits, force extra takes or require another approval for already authorized work. Inspect the first changed output before committing the remaining budget. Upload local media through the supported presigned upload flow, submit each paid job once and retain its project ID. Recover the same project after an ambiguous response or polling timeout; a new creation is not retrieval.

## Join, caption and deliver

Match the actual source dimensions, frame rate, time base, pixel aspect and audio format before joining. Choose crop or contain based on the full motion path and requested placement; never hard-crop every input automatically. Handle silent clips explicitly. Cut on a complete word/action, preserve the body's first word and check the join for clicks, repeated speech, pauses or visual jumps. A fade cannot repair an inappropriate cut.

Apply captions to each final timeline, using real timestamps where available. Rebuild exact overlays and CTA from editable copy, respecting the destination's actual overlay regions. Keep the clean master separate. Watch and listen through the full export, especially the new opening, seam, product label and final sentence. A successful encode, metadata or contact sheet does not prove acceptance.

Deliver one file per requested variant, its clean version when requested, and a small manifest containing the source master/interval, concept, changed variable, exact copy, generated project IDs, observed credits, final path and failed or unverified criteria. Keep the same identifiers on later revisions. Do not infer sales lift from a hook or viewing metric, prescribe campaign budgets/kill thresholds, publish or modify an ad account without separate authority.

**Evidence limit:** this is an adapted production recipe, not a validated Magic Hour ad-performance or end-to-end quality benchmark. Mark missing playback, listening or independent acceptance explicitly.

Adapted from Varg's [product-video-hooks](https://github.com/vargHQ/skills/blob/04cd89af3dfa6ba1c0a8dc6470107d02579b198b/product-video-hooks/SKILL.md), [ad variants](https://github.com/vargHQ/skills/blob/04cd89af3dfa6ba1c0a8dc6470107d02579b198b/varg-ad-variants/SKILL.md) and [creative testing](https://github.com/vargHQ/skills/blob/04cd89af3dfa6ba1c0a8dc6470107d02579b198b/ad-creative-testing/SKILL.md), under the included MIT license. Provider calls, fixed batch sizes and performance claims are not carried over.