๐ ๐ ๐ฉ๐ถ๐ฑ๐ฒ๐ผ ๐๐ฒ๐ป๐ฒ๐ฟ๐ฎ๐๐ถ๐ผ๐ป ๐ฃ๐ถ๐ฝ๐ฒ๐น๐ถ๐ป๐ฒ ๐ง๐ต๐ฎ๐ ๐๐๐ถ๐น๐ ๐๐๐๐ฒ๐น๐ณ
I built a two-minute video using only one prompt and a conversation with Claude Code.
I did not use an image editor. I did not use a video timeline. I did not click through any video software. I only spoke to the AI.
Claude built the entire pipeline itself. It created the tools for image generation, voice synthesis, and video editing. It even wrote the glue code to tie everything together.
Here is how the system works:
โข Image Generation: It uses OpenAI and Google models to create every still frame. โข Image-to-Video: It uses models like Seedance 2.0 and Kling to turn stills into motion. โข Voice: It uses ElevenLabs for narration and Google Gemini for drafts. โข Editing: It uses ffmpeg to handle cuts, zooms, and audio mixing. โข Director: This is the meta-skill. It manages the entire pipeline from start to finish.
This system grew one skill at a time. When I needed a voice, I gave Claude the API docs. It made one successful call and then wrote a permanent "skill" so it never had to learn it again. This pattern applied to every tool in the kit.
The cost of this project was $45.26.
Most of that went to trial and error. Generating video is expensive. A five-second clip can cost up to three dollars. This changed my workflow. I no longer ask for "just one more variant" without thinking. The real-time cost makes me more decisive.
There were major challenges:
- Claude cannot see the video. It only reads transcripts and image data. I had to use precise, written instructions to keep timing in sync.
- A git error once wiped two months of work. I had to implement a hard wall to prevent the AI from running dangerous commands in the wrong folders.
- Transitions are hard. To make a smooth zoom into an AI video, I had to use a trick with ffmpeg to paste the original frame back into the center of an AI-painted background.
The video is just the receipt. The real value is the kit of reusable skills underneath it.
Stop reading about AI. Start building your own pipeline through conversation.
Source: https://dev.to/hiper2d/my-video-generation-pipeline-that-built-itself-459n
Optional learning community: https://t.me/GyaanSetuAi