Convert a folder of design screenshots into a SKILL.md that you attach to an AI coding session to keep generated UI consistent with your visual preferences.
Extract repeatable design rules from a reference set and use them to guide an AI agent building a new frontend from scratch.
Run the local pipeline privately with your own images without using the hosted web app or any cloud storage.
| jaytel0/taste | 5bv57zcm44-max/noxus-ai-open-whatsapp | javlonbek1233/amali-ish-9 | |
|---|---|---|---|
| Stars | 27 | 27 | 27 |
| Language | TypeScript | TypeScript | TypeScript |
| Setup difficulty | moderate | moderate | moderate |
| Complexity | 3/5 | 4/5 | 2/5 |
| Audience | vibe coder | developer | vibe coder |
Figures from each repo's GitHub metadata at analysis time.
Requires OPENAI_API_KEY and ANTHROPIC_API_KEY (or an OpenRouter key), run npm install then npm run taste with images in reference-images/.
Taste is a tool that looks at a collection of reference images and produces a written style guide, stored as a file called SKILL.md, that an AI coding assistant can follow when generating interfaces. Instead of describing your design preferences in words, you show the AI examples of what you like, and it extracts the patterns itself. The pipeline works in five steps. You drop JPG, PNG, or WebP images into a folder. The tool indexes them, removes duplicates, and assigns each one an identifier. Then two AI vision models analyze each image independently, looking only at visual details like layout, spacing, color use, typography, and density, deliberately ignoring what the app in each image actually does. The notes from both models are merged into one synthesized description per image. Those descriptions are grouped into chunks, and each chunk is distilled into a set of design rules. A final writing step turns those rules into concrete, specific instructions rather than vague aesthetic labels. The output might say things like "use neutral sans-serif typography" and "keep accent color localized" instead of "make it look premium." You can run the full pipeline locally with a single command after setting API keys for OpenAI and Anthropic. The result lands in a timestamped folder in your project. There is also a hosted web demo at taste.jaytel.com, which uses a database and cloud storage to run the same pipeline in a browser without any local setup. The repository is a TypeScript monorepo with a web frontend, a reusable AI package for prompting and provider handling, and a scripts folder for the local pipeline runner. It supports routing through AI gateway services like OpenRouter as an alternative to direct provider keys. The production web app adds Vercel hosting, Postgres for workflow state, and Vercel Blob for artifact storage. The repository includes comparison screenshots where agents with and without a Taste-generated skill produced dashboards and chat interfaces, showing the difference in visual consistency. No license is specified in the repository.
A pipeline that analyzes a set of reference design images using vision AI models and produces a SKILL.md instruction file that guides AI coding agents to match that visual style.
Mainly TypeScript. The stack also includes TypeScript, OpenAI, Anthropic.
No license specified, all rights are reserved by default, meaning you cannot reuse or redistribute the code without the author's permission.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly vibe coder.
This repo across BitVibe Labs
Verify against the repo before relying on details.