jaytel0/taste

What it solves

Taste turns a handful of reference images into a reusable skill file (SKILL.md) that captures the visual style of those images. Instead of describing your taste in vague words, the skill spells out concrete design rules an AI agent can follow when generating new interfaces — for example, which typography, spacing, colors, and shadows to use. This solves the problem of getting AI to reproduce a specific visual style consistently, without needing to hand-write long style guides.

How it works

The process has five stages:

1. Curate references – You provide a tight set of JPG, PNG, or WebP images. Close-up crops and high-resolution images work best because the visual details are easier to read. The idea is to show repeated examples of layout, spacing, color, type, texture, composition, density, and hierarchy.

2. Index the corpus – The pipeline scans every image, removes exact duplicates, and gives each image a stable ID (up to 20 images per run). The raw file list is saved as runs/{runId}/01-corpus/images.jsonl.

3. Analyze each image – Vision models read each image independently (default: OpenAI GPT‑5.5 and Anthropic Claude Sonnet 4‑6). The prompts tell the models to ignore what the app does and focus only on visual evidence. Raw notes are saved per model as runs/{runId}/02-image-notes/raw/{imageId}/{model}.md.

4. Fuse and chunk – A synthesis model merges each image's raw analyses into one canonical note (with the source anonymized). Then those notes are grouped into chunks, each chunk becomes a rule synthesis, and larger runs go through merge layers until a final rule set is produced. Intermediate files are stored under runs/{runId}/03-rule-set/.

5. Write the skill – The final rule set is turned into concrete instructions. The skill avoids vague labels like "premium and tasteful" and instead says things like "Use neutral sans-serif typography" or "Keep accent color localized." The final artifact is runs/{runId}/04-skill/SKILL.md.

The repo includes a hosted demo (taste.jaytel.com), but the local pipeline is the simplest way to use it privately — it needs no Postgres, Vercel Blob, Cron, or hosted auth. You just set your API keys, drop images in a folder, and run npm run taste. The web app version stores workflow state in Postgres and run artifacts in private Vercel Blob storage.

Who it’s for

• Designers or developers who want to capture a visual style from reference images and reuse it in AI‑powered front‑end generation.
• Teams that build with AI agents and want a repeatable way to make those agents follow a specific look.
• Anyone comfortable running Node.js scripts locally or who wants to try the hosted demo.

Highlights

• Turns images into a reusable, concrete skill – The generated SKILL.md is explicit about visual rules, not generic praise.
• Works with multiple AI providers – You can use OpenAI and Anthropic directly, or via OpenRouter or Vercel AI Gateway.
• Local‑first – The pipeline can run entirely on your machine, writing results to .taste/runs/... with no external services required.
• Validated with comparisons – The README shows side‑by‑side tests where the Taste skill outperformed a base model and a hand‑crafted frontend‑design skill in generating dashboards and chat interfaces.

---

Note: All details above are taken directly from the project README.