Here is my advice for anyone writing prompts for image or video models: start from a prompt that already worked.
The reason is simple. A good prompt carries a structure, and the structure is the part that is hard to invent. Once you have it, you only swap in your own subject.
So where do you find prompts that worked? AgentCraft is a free library built for exactly this job.
A quick guess
Let me start with a question. A 15 second film clip, 1940s war documentary look, made with MiniMax H3. How long do you think the prompt was?
One sentence? A short paragraph?
The answer: the creator split the 15 seconds into timed beats. An establishing shot, rising tension, the firefight, a human moment, then a pull back. Each beat has its own camera move. Below that sit separate blocks for the environment, the characters, the visual style and the physics, plus a long negative prompt.
That structure is why the clip holds its 35mm look from start to end. You can copy it for any period piece you want to make.
This prompt came from @RuzainaMeer on X. AgentCraft keeps it word for word, next to the clip it produced, with a link back to the post.
What AgentCraft has
The whole library is open, and the numbers are large.
- Video prompts, 940. Seedance 2.5, MiniMax H3, Wan 3.0, Kling 4.0, Veo and Gemini Omni, grouped by type: cinematic, action, sci-fi, anime, ads and more.
- Image prompts, 1,307. GPT Image 2 and 2.5 and Nano Banana, grouped by use: posters, portraits, infographics, product shots, UI mockups, logos.
- Agent prompts, 353. System and role prompts for Claude Code, Codex and other coding agents, grouped by field.
- Claude Code skills, 4,636. Each one read and screened before it was listed, with its install command.
Every video and image prompt shows the output it made and the place it came from, either the model maker's guide or a creator's post. You see the result first, then decide whether the prompt is worth your time.
Links: video prompts, image prompts, agent prompts, skills.
The open data
This is the part I like most. Everything behind the site is public, so you can take the data and build on it.
- awesome-image-video-prompts on GitHub, 3.4k stars. The video and image prompts, with outputs and sources.
- awesome-prompts on GitHub, 9.2k stars. Prompts, frameworks and papers for LLMs and agents.
- awesome-prompts on Hugging Face, 3,149 prompts in two subsets. The engineering subset has 372 system prompts; 73 of them cite arXiv papers by id. The library subset has 2,777 everyday prompts for writing, marketing, study and design, each in English and Chinese.
Loading the dataset takes two lines:
from datasets import load_dataset
eng = load_dataset("Adrian2016/awesome-prompts", "engineering", split="train")
lib = load_dataset("Adrian2016/awesome-prompts", "library", split="train")
A prompt dataset with sources attached is handy for evals, for fine-tuning a prompt writer, or for studying which patterns models follow.
Three patterns worth stealing
Browse the library for a while and a few patterns keep showing up. Here are three that are easy to reuse!
1. Give video a timeline. Write the clip as timed beats with a camera move for each, as in the war clip above. The model then plans the shots in order.
2. Give images a count. One GPT Image 2 prompt asks for a town scene with exactly 100 elements, then asks the model to list every item. The model plans the picture as an inventory, and you get a dense illustrated town under a numbered list.
3. Name the layout. Another prompt asks for a catalog of 1990s anime character styles: 16 labeled archetypes in a grid, one per panel, at 16:9. One sentence, and the result is a clean reference sheet.
The common thread: tell the model how to organize the work, and it fills in the rest well.
Lesson
Good prompts are templates. Borrow the structure, keep your own subject.
Which model are you prompting most these days, and what structure works for you?
Top comments (0)