Magic Hour explainer video

Deliver a watchable explanation with a beginning, a useful idea and an ending, plus its editable script. A folder of disconnected clips is not the finished video.

Use the connected Magic Hour MCP at https://mcp.magichour.ai/ or the project's existing API integration. Use the connection guide if needed. Verify account_retrieve and discover current image, audio and video tool schemas. Discover a local or connected compositor before promising an assembled video; Magic Hour generation tools do not imply a general-purpose assembly endpoint.

Convert the brief into a clear story

Read the actual source material. Decide what the audience should understand or do afterward, then write a short script grounded in that source. Preserve uncertainty and distinguish an illustration from evidence. Keep source links alongside the relevant script passage so factual claims remain traceable.

Use the requested duration and audience. When choices are delegated, choose a fitting visual style, pace and available voice and proceed within the authorized budget. Ask only for a blocking source, rights/identity requirement or unresolved meaningful choice. Do not force a two-turn setup questionnaire, a one-minute minimum or identical ten-second scenes.

Plan natural beats: one idea and one visual purpose per beat. Estimate length from a read-through, then fit the cut to measured audio. For longer videos, read timing and assembly. Avoid writing more narration than can fit and compensating with unreadably fast speech.

Make speech and visual identity first

Prefer an authorized supplied recording when it meets the brief. For TTS, call ai_voice_generator_create_audio with a voice actually listed in the live schema; do not invent a neutral narrator ID. If the available presets cannot meet a professional voice requirement, use an authorized reference with ai_voice_cloner_create_audio or explain the exact missing capability. Do not silently substitute a named person for a requested original narrator. Label a synthetic demonstration voice accurately.

Record each created audio ID and use wait_for_audio_project. Retrieve and measure the actual file. Listen when the environment supports it, checking pronunciation, missing words, pauses, noise and the final sentence. If listening was unavailable, state that limitation; duration and successful API status do not establish speech quality. Fix a mispronounced segment before paying for its visuals.