Magic Hour Media

Magic Hour media

Deliver a usable image, video, or audio result. A queued project ID is an intermediate result unless the user asked only to submit a job.

Requires network access and either a connected Magic Hour MCP server or a Magic Hour API key with an HTTPS-capable runtime. Generation consumes Magic Hour credits.

Work from the user's goal

Accept ordinary requests such as “put me in this clip,” “make my character do this” or “turn this portrait into an impossible scene.” Infer the placement and creative direction from supplied context; recommend one direction when needed. Ask only for missing information that changes the result or spending authority, not model names, schema fields or a completed prompt template. When creative choices are delegated, inspect and accept suitable intermediate assets yourself and continue. Respect requested review points and the existing budget. Lead delivery with the actual media; keep IDs and technical records in the project unless they are needed to explain a failure. For transformations or recreating a visual format, read the remix workflow; for music, read music and performance.

Connect to the right service

Use an already-connected Magic Hour creation MCP if available. Its hosted endpoint is https://mcp.magichour.ai/. The separate https://docs.magichour.ai/mcp searches documentation; it does not generate media.

If creation tools are unavailable, read references/setup.md in this skill for setup and the API fallback. Do not silently replace an existing client configuration. Never request an API key in chat or put it in source code; use the client's credential flow or MAGIC_HOUR_API_KEY in the local environment.

Use live tool schemas, or the API documentation index, for current models, allowed durations, resolutions, and prices. Web-app features and plans do not imply identical API support. Do not copy a remembered model catalog or price table.

Choose the workflow

Tool names can have a client-specific prefix. Discover their current input schemas before calling them.

For full-person photo placement, distinguish Body Swap's person_file_path and scene_file_path from face-only replacement. For Video-to-Video, a custom prompt is ignored under the default prompt mode. Music Video Generator's web scene-assembly workflow, Video Extender and Video Upscaler have no equivalent dedicated endpoints in the public schema as checked September 11, 2026. Recheck support; when still absent, prepare the accepted inputs and direct web handoff instead of inventing tools. The complete use-case map links those workflows.

When installed, use ai-video-generation for a broad video request, creator-video-remix for a supplied reference format and ai-media-quality-review before final delivery. Use the other focused skill for image-to-video, face/body swap, text-to-video, image editing, talking video, character replace, music video, existing-video editing, product visuals, campaign kits, character consistency or thumbnails. Distinguish face-only identity changes from whole-character replacement, and a portrait's new speech from an existing video's performance. Do not require the user to memorize a skill name or install a companion just to finish a supported task. Reuse the current connection, project brief and accepted media; only ask for missing information that changes the output or authorization.

Set the quality target

Before generating, translate the request into a short quality brief:

• intended use, aspect ratio, resolution, and duration
• subject, composition, viewpoint or camera move, lighting, and visual style
• details that must remain accurate, such as identity, product shape, logo, colors, text, clothing, or background
• failure conditions that would make the result unusable

Infer these from the user's project when possible. For a website hero, leave deliberate negative space where the interface places text. When exact identity, product geometry, packaging, or branding matters, prefer editing an authorized reference image over recreating it from text alone. For image-to-video, use the source image to define appearance and prompt mainly for subject motion, camera motion, timing, and what must stay fixed. Keep short videos to one coherent action unless the chosen duration and model support a more complex sequence.

Choose the model from the live schema according to the requested result. Prefer the current recommended model for a general request. When the schema describes a model as better for the needed behavior—such as typography, identity preservation, reference control, motion, audio, or speed—use that evidence to select it. Do not select the cheapest or fastest model when the user prioritizes final quality. Match resolution to the placement and critical detail: a tiny preview cannot prove small text is accurate. Use a cheaper preview when it answers a specific appearance or motion question, and label it as a preview. Check the total cost against the existing authorized budget; ask only when that authorization does not cover the intended work. Preserve the user's model, resolution, and spending limit.

Write a concrete prompt in natural language. Describe the subject and action first, then composition or camera, lighting and materials, style, and required constraints. Avoid conflicting styles, long adjective lists, and unsupported negative-prompt syntax. For edits, state the requested change and the elements that must remain unchanged. Request only missing inputs that materially affect the result. Preserve the user's chosen model and budget.