dharani2201-mini/nano-banana-prompting

Nano Banana Prompting is a collection of JSON prompt templates for Nano Banana, Google's Gemini based image generation model, and other structured image generators. It is not code you install and run, it is a set of prompt patterns and guides meant to solve two common problems people run into with AI generated images: the same character's face changing between different generations, and images having a plastic, overly smoothed, obviously AI-made look. The repository is organized into a few folders. A SKILL.md file explains the full method and is written so it also works as a Claude Code skill. An examples folder has several complete, ready to copy prompts, including a before and after comparison showing a normal prompt next to one built with this project's approach. A templates folder has starter JSON files with placeholders a user fills in for their own character or scene, and a guides folder explains the two main techniques in more detail. The first technique is an identity_lock block inside the JSON prompt, which describes a character's specific features, such as hair, skin, and any small distinguishing details, along with a list of things to preserve and a list of rules like no face swap or no beautify. The idea is that vague descriptions produce a different looking person every time, while specific, repeated details keep the same character recognizable across many generations. The second technique combines positive realism_rules describing imperfections a real photo would have, such as visible skin texture or harsh sunlight, with a negative_prompt list explicitly banning things like porcelain skin or a beauty filter look. The README notes that this format is for still images only, and suggests generating a still first, then feeding it into a video model like Veo 3, Sora, or Kling as a reference image, since those tools tend to work better with prose prompts rather than JSON. The project can also be installed directly as a Claude Code skill by cloning it into the user's skills folder. It is released under the MIT license.