Many people have asked me how I went from being a programmer to creating AI videos with millions of views.
The answer is four characters: aesthetic accumulation.
When you see a beautiful video clip, be willing to spend credits to recreate it. Do this enough times, and you'll learn shot composition thinking, understand which models are best for specific types of clips (faces, actions), what prompts ensure character consistency, and pick up color matching and art styles.
It sounds simple, but the first step stops over 99.99% of people. Especially with current AI video tools, almost no one thinks they're cheap. If you're lucky, you nail it on the first try; if not, you might pull ten times and get nothing usable.
So today, I'm open-sourcing a project that has been running for nearly a month: awesome-seedance.
This epic production is my first major project where I spent real money to re-run all 463 AI video prompts. It resulted in 14 prompt templates covering basically all types of AI videos, plus 25 AI video Skills!
All open source.
You don't have to gamble alone anymore.
I barely slept this month, spending 20,000 RMB just on generation attempts.
🔗 github.com/LearnPrompt/awesome-seedance
There are many AI prompt collection websites out there. Why did I want to make a new one?
The reason for wanting "GoodCase" was simple: previously collected prompts varied wildly in quality, lacked a unified scoring mechanism, and often had no categorization. I might as well randomly search X or use fixed prompts when testing new models, which was boring and didn't test model limits.
So I burned through three reset cards for GPT-6 Astra and two Max accounts for Claude Fable 5.1. My goal was to ensure every prompt was tested and reproduced, scored, and categorized. You can find a direction you like, take the template, change two words, and use it.
That's how GoodCase came to be.
It automatically collects cases with prompts from GitHub, Douyin, Xiaohongshu, and X daily. Done!
Using four multimodal models—GPT, Claude, Grok, DeepSeek—to score reproduced cases and prompts, cluster similar video types, and title collected cases. All done!!
I even compared whether renting machines to deploy Minimax H3 was cheaper than using fal's Minimax H3 Max...
Midway through testing, I considered switching to awesome-gpt-image-2.5 and skipping video reproduction because it was much harder than expected. Different models have different max durations; a 15-second prompt might fail completely on a 10-second model. Some prompts need reference images for stable style, so I also burned through my GPT Image 2 quota.
Video evaluation dimensions are also several times more complex than images. Images are easy to judge at a glance. For videos, you must watch the whole clip, checking action continuity, camera movement, audio sync, etc. Initially, my eyes were strained. Thanks to Grok Bot having a high-speed cloud PC, I could screenshot videos by the second, allowing multimodal models to judge results based on frames. (Not an ad)
After reproducing over 200 videos, categorization became natural. At this stage, I felt fulfilled organizing everything learned into one project.
Let's look at the final 14 prompt templates and 25 skills.
Timeline & Storyboards, Reference Images & Character Consistency, Handheld UGC Vlogs, First-Person One-Shot, UGC Review/Talking Head, Product Commercials, Process & Transformation Montages, Dialogue & Lip Sync, Cinematic Narrative Shorts, Animation & Stop-Motion Styles, Fight Scenes & Physics Spectacles, Beat-Synced Music Videos, etc.
These were naturally summarized from the 463 cases. Each template contains distilled conclusions on "how to write prompts for this video type for the best effect," derived from running dozens of similar prompts.
I'll highlight two.
The Handheld UGC Vlog template took the most time.
We can improve realism by applying era-specific camera flaws. A 2005 DV looks different from an iPhone X. DVs have a jelly-like vertical shake and laggy tracking. iPhones have smooth electronic stabilization. If setting 90s VHS, include tape artifacts and color bleed.
To maximize usability, I kept reference videos. You can ask Codex to generate prompts based on the structure and examples in the template, minimizing information loss. As new examples are added, the template updates.
Pitfalls are included too.
Asking for both handheld realism and 4K cinematic lighting in one prompt yields weird results because no single device does both. Also, frequent boyfriend-perspective close-ups increase AI face exposure risk due to detail scrutiny. Using digital zoom as transitions breaks one-shot continuity.
These pitfalls are avoided in the template.
How to use it? Copy the long prompt into your Agent and ask it to adapt it to your needs.
Then you get this result instantly.
Previously, when making AI videos in unfamiliar fields, the old method was following experts or viral posts to create representative prompt structures. But you'd realize some visuals don't fit that framework.
Most people write fight prompts like "Two warriors fighting fiercely, punches and kicks flying." Lower-tier models produce people punching air.
After studying good fight cases, I found the key: specify contact points. Abstract descriptions don't work. Write "straight punch hits chin," "block then counter body kick." Tell the model where the fist lands and how the opponent reacts.
But don't overdo contact points.
For 5-10 second fights, keep it under 4 action segments. More than 4 causes blurring or face distortion in some models.
My solution is writing moves as chains: straight punch -> hook -> low sweep; opponent blocks -> slips down -> parries -> counters with body kick. The model understands sequence.
The wildest part: specifying martial arts biomechanics works. "Muay Thai roundhouse" vs. "Taekwondo side kick" produces different amplitude and power.
Muay Thai shows large arcs from body rotation.
Taekwondo shows linear hip snap bursts. Key differences are visible.
At this point, I wondered: do we need Skills if we have templates?
Yes. In plain terms, a template fills parameters for one shot; change subject/scene and go. One template = one shot.
A Skill is a full methodology: how to break shots, choose models, handle transitions, maintain consistency. Say "Make a 30s product ad," and it breaks it into shots, defines structures, and links them.
Think of templates as recipes for single shots, and Skills as shooting manuals for entire shorts.
Here's a selection table.
Example:
Start with templates. Find a case in awesome-seedance, copy the prompt, ask AI to adapt it, and generate.
I used the UGC Review/Sales template.
Copied modified prompt to Seedance 2.5. Result: Hair washing process shown clearly from display angle.
If you frequently make similar content (e.g., weekly product demos), use Skills. Install the corresponding method from GoodCase Skills, state your need, and it handles the workflow without rewriting prompts from scratch.
Workflow: Find direction, download Skill from GoodCase,
Install in Agent, call Skill for video request.
Result: Slightly different from template-only. More review-focused, more spoken lines, more selling points.
Over these two weeks, one word kept popping up: Tracing.
Like learning to draw by copying masterpieces stroke by stroke, you only understand why white space or dark shading matters when you do it yourself.
AI video prompt templates work similarly. Find an effect you like, swap subjects, adjust actions/cameras, observe changes.
Eventually, you question why CAMERA modules come first, what happens with too many contact actions, or the difference between specifying a style vs. just saying "fight well."
You develop judgment through practice.
With experience, you know how to write prompts for desired visuals.
But someone needs to help take the first step.
Previously, this meant burning money testing blindly.
Awesome-seedance aims to make this easier.
Open project, find a case, copy prompt, run it. With references, modifications become directional.
If my pitfalls save others trouble, the effort was worth it.
The project has immature parts, but I've released everything openly.
Future plans include more model tests and a camera language series (dolly, crane, Steadicam, FPV).
Another big hole dug...
Welcome to fill it together.
Submit good cases via PR; I'll include them in tests.
Your success becomes another's reference.
Gambling alone is lonely.
Join us.
