Why we open-sourced our AI prompt library (260 prompts, MIT)

The pitch

We just open-sourced 260 prompts from ZSky AI's production library at github.com/zsky-ai/zsky-prompt-library. MIT licensed. Use them with any AI tool — not just ours.

Why open-source prompts?

Most AI prompt collections you find online are either:

1. Aesthetically curated but not technically useful (Pinterest-mood-board style)
2. Hidden behind paywalls and "prompt courses"
3. Tied to a specific tool's syntax that breaks elsewhere

The ZSky prompt library is different in three ways:

Tested in production. Every prompt has been run through actual generation. We kept the ones that worked, dropped the ones that produced inconsistent output.

Tool-agnostic. Phrasing follows photo metadata conventions (camera, lens, light direction, color temperature) that any modern image model has been trained on. They work in Midjourney, Stable Diffusion, ZSky, etc.

Categorized for actual use cases. 11 categories: studio backgrounds, character portraits, cinematic lighting, product shots, anime styles, architectural rendering, food photography, fashion editorial, abstract textures, scientific illustration, narrative scenes.

What's in the library

zsky-prompt-library/
├── studio/                    # Backdrop, lighting, lens setups
├── portrait/                  # Character + face consistency patterns
├── cinematic/                 # Film stock, camera angles, focal lengths
├── product/                   # Commercial product photography
├── anime/                     # Style preservation across iterations
├── architecture/              # Renderings, scale references
├── food/                      # Plating, lighting, mood
├── fashion/                   # Editorial, runway, lifestyle
├── abstract/                  # Texture, pattern, mood
├── scientific/                # Diagrammatic, illustrative
└── narrative/                 # Scene composition, storytelling

Each category has 20-30 prompts with example outputs.

The pattern that makes prompts work

After running thousands of generations, the consistent finding: describe physical setup, not aesthetic mood.

❌ "professional studio photography, photorealistic, cinematic, high quality"
✓ "matte seamless paper backdrop, key light camera-left at 45° softbox, 5500K daylight, 85mm at f/1.8, subject 6 feet from backdrop"

The first reads like Pinterest. The second tells the model what to physically render. Models trained on photo metadata respond dramatically better to specifications than to vibe words.

Get it

git clone https://github.com/zsky-ai/zsky-prompt-library

PRs welcome — if you find a prompt that beats one in the library, send it. We'll attribute and merge.

The prompt research came out of building ZSky AI (free unlimited AI image + video generator). Open-sourcing the prompt library because the prompts shouldn't be the moat — the platform should be.

I'm Cemhan Biricik. I shoot photography (Sony WPA top-10, two Nat Geo awards) and run zsky.ai. The prompt library is the part of our stack we're happiest to share.