Turn mental scenes into 15-second video prompts
If you want to write instructions that specify camera position, character movement, and lighting rather than just listing keywords, this repository is a fit. It shows how to break a scenario into shots, organize each cut's composition in a table, and convert it into a prompt the model can understand. The actual video generation is handled by the provider you choose; this repository focuses on the planning and prompt-writing steps that come before.
v6.7.0 and multilingual guides
The current version is v6.7.0, released under the MIT license. It defaults to English but supports Chinese, Japanese, Korean, Spanish, and Russian, with a quick start guide for each language that takes five minutes. Recent changes are listed in the changelog.
Prompt rules to keep model context
Video generation models do not retain information from previous shots when a cut occurs. To compensate, this repository includes a 'lock line' at the end of each shot that restates the light, characters, and location. It also assigns only one action per shot, instructs that other characters' movements remain natural, and describes emotional expressions by combining a feeling with a single physical anchor. Prop manipulation is specified through hand movements, and all prompts undergo a pre-publication moderation dictionary check to prevent rejection based on specific words rather than intent.
Seven published scenarios and shot tables
The README includes seven completed 15-second story examples. Each example follows a short drama structure of four to five shots, designed so the situation is clear by the second shot. The stories cover various genres, including 'Reservation-only restaurant,' 'Pier in a storm,' 'Birthday party,' 'Delivery driver,' 'Resignation letter,' and 'Last helmet.' Under each clip, a simple keyword input typical of user behavior is shown side-by-side with the detailed prompt generated by this tool to illustrate the difference. Some clips are not yet rendered and are marked with a slate.
Pre-use checks
This repository is not a video generation engine but an agent skill for planning prompts and linking reference materials. Therefore, you need an account with a video provider that supports Seedance 2.0, and generation costs are borne by the user according to that provider's policy. As a Python-based tool, it requires a Python version compatible with your installation environment. You should also understand that the model does not retain information between cuts and that all visual elements must be repeatedly specified in the prompt. The license is MIT, allowing commercial use, but the copyright of generated videos follows the terms of your video provider.