味觉将一组参考图像转换为可重复使用的SKILL.md。

• 托管演示：
• 尝试味觉技巧：pipeline/taste/taste-skill/SKILL.md

收车形状

网页应用是更广泛流程的演示。本地流程是最简单的 私密使用该仓库的方法：它将文物写入.taste/runs/...，并且 不需要Postgres、Vercel Blob、Cron或托管认证。

快速入门：本地管道

如果你想从自己的图片中创造味觉技能，可以用这条路径。

在 .env.local 中设置直接提供者密钥：

在 reference-images/ 中放置 JPG、PNG 或 WebP 参考图片，然后运行：

最后一个技能写为：

方便的替代方案：你可以设置一个，而不是两个直接键 网关密钥：OPENROUTER_API_KEY 或 AI_GATEWAY_API_KEY。Vercel 的 AI 网关 使用 AI_GATEWAY_API_KEY 并路由纯 AI SDK 模型字符串 默认是Gateway。

对应另一个文件夹：

实用选项：

比较

我们用三个代理做了一个小型前端设计对比：一个基础模型 没有技能，一个使用人类的frontend-design技能，另一个使用该技能 从pipeline/taste/taste-skill/SKILL.md生成味觉技能。每个特工是

Taste

Taste turns a set of reference images into a reusable SKILL.md.

• Hosted demo:
• Try a taste skill: pipeline/taste/taste-skill/SKILL.md

Repo Shape

apps/web/          hosted frontend demo and API worker implementation
packages/ai/       reusable prompt, provider, chunking, and generation package
scripts/           local pipeline runner for humans and agents
reference-images/  ignored local drop folder for your own JPG/PNG/WebP inputs
pipeline/taste/    Jaytel's example generated taste skill and pipeline notes
docs/              README images and examples

The web app is a demo of the broader pipeline. The local pipeline is the easiest way to use this repo privately: it writes artifacts to .taste/runs/... and does not need Postgres, Vercel Blob, Cron, or hosted auth.

Quick Start: Local Pipeline

Use this path if you want to create a taste skill from your own images.

npm install
cp .env.example .env.local

Set the direct provider keys in .env.local:

OPENAI_API_KEY=...
ANTHROPIC_API_KEY=...

Put JPG, PNG, or WebP reference images in reference-images/, then run:

npm run taste

The final skill is written to:

.taste/runs//SKILL.md
.taste/runs//04-skill/SKILL.md

Convenience alternatives: instead of the two direct keys, you can set one gateway key: OPENROUTER_API_KEY or AI_GATEWAY_API_KEY. Vercel’s AI Gateway uses AI_GATEWAY_API_KEY and routes plain AI SDK model strings through the Gateway by default.

Run against a different folder:

npm run taste -- ./path/to/images

Useful options:

npm run taste -- ./images --out ./my-skill-run
npm run taste -- ./images --skill-name my-taste
npm run taste -- ./images --provider openrouter
npm run taste -- ./images --model openai/gpt-5.5

Comparison

We ran a small frontend design comparison with three agents: one base model with no skill, one using Anthropic’s frontend-design skill, and one using the generated Taste skill from pipeline/taste/taste-skill/SKILL.md. Each agent was given the same prompt within each test and asked to produce a standalone HTML file.

Test 01: Opus 4.7

Prompt: Design a dashboard. Build in an HTML file. Ask no questions.

Test 02: Opus 4.7

Prompt: Design a chat interface. Build in an HTML file. Ask no questions.

Test 03: GPT-5.5

Prompt: Design a chat interface. Build in an HTML file. Ask no questions.