TLDR;
This video explains a method for creating optimized prompts for AI video generation using custom GPTs trained on official model documentation. The key idea is to focus on the creative vision and let AI handle the translation of ideas into model-specific prompts.
- Train a GPT on the official documentation of the AI model you're using.
- Input your idea into the GPT and let it generate the optimized prompt.
- Focus on expressing your creative vision and iterate based on the results.
Intro: The Art is Your Idea [0:00]
The presenter introduces a technique for writing prompts for text-to-video AI models. The core idea is that the art lies in the idea, not in the prompt itself. Prompting is viewed as a translation process, converting an idea into a language that the video model understands. The presenter will demonstrate how to train a GPT on official model documentation to generate optimized prompts.
Getting prompt documentation from Hailuo, Kling, Luma, Runway, Veo, Vidu [1:11]
The initial step involves gathering documentation for the specific AI model being used. The presenter shows where to find the documentation for various models, including Hailuo, Kling, Luma Dream Machine, VU, Runway (in multiple places), and VO (via the DeepMind site). The documentation is typically saved as a PDF for easy access.
Creating a custom GPT for writing better AI video prompts [2:07]
After obtaining the documentation, it can be uploaded to a chat interface (if on a free plan) or used to create a custom GPT or GEM (if on a paid plan for ChatGPT or Gemini). In ChatGPT, this is done by going to GPTs, then create, and uploading the file in the knowledge section. In Gemini, the user needs to go to explore gems, then new gem, and add the knowledge. The instructions section should state that the GPT provides optimized prompts based on the documentation. For free versions of tools like DeepSeek, the documentation can be uploaded to the chat window, but it's recommended to start a new chat for each project to avoid the model forgetting the document.
Example of iterating with a Prompt Sidekick [2:41]
The presenter demonstrates how to iterate on prompts using the custom GPT. Instead of manually rewriting a prompt that produced an undesirable result (too dramatic), the presenter asks the GPT to make it more cinematic and emotionally subdued. The resulting prompt is closer to the desired outcome, and the GPT provides additional suggestions. The presenter also shows how to adjust camera direction by describing the shot as envisioned and how to refine the attire of a character in the scene.
How I use my Prompt Sidekick [4:49]
The presenter explains how they use their "Prompt Sidekick," highlighting the use of placeholders like "outline the best practices." Clicking this placeholder allows the GPT to access uploaded knowledge and provide a list of best practices. The presenter can input a movie idea and request model-optimized prompts for various models like Runway, Hyo, VU, VO, and Kling to see which model best captures the idea.
Is Prompt Engineering Still Even a Thing? [7:13]
The presenter questions the value of manually writing prompts, suggesting that the focus should be on expressing ideas effectively. The ability to articulate what you want to express is the key skill for the future. The presenter advocates for focusing on creative expression and allowing AI to translate for another AI.
Conclusion [7:49]
The presenter concludes by stating that computers should write prompts for computers, while humans should write for humans. The emphasis is on telling a story and leaving the prompting to the machines.