You know the look.

Same gradients. Same cards. Same three-column layout with the little icon on top. Every AI-built site on the internet is starting to look like it came off the same assembly line — because it did.

That’s not a model problem. The model is fine. It’s a taste problem: your agent has never been handed a point of view, so it reaches for the average of everything it has ever seen. And the average looks like a template.

taste-skill, by Leonxlnx, exists to fix exactly that. It’s a skill file that gives coding agents design taste — art direction they can check their own work against. It had 100,902 stars as of October 2026, which makes it one of the most-starred agent skills on GitHub. It’s MIT licensed and free.

Here’s what it actually does, why references beat prompts, and how it fits into a bigger design stack for your agent.

What taste-skill does: design taste for coding agents, written down

Designers carry taste in their heads. They can’t always explain it, but they know when a page feels right: the composition holds together, the hierarchy tells your eye where to go, the whitespace gives everything room to breathe.

taste-skill writes that judgment down in a form an agent can use. It gives your coding agent:

Art direction. A clear visual point of view for the project, instead of whatever default styling the agent reaches for every single time.

Reference images. Real visual references the agent works from, so design decisions are grounded in something concrete — not guessed from training averages.

Layout judgment. Composition, hierarchy, and whitespace rules — the gap between “a page with stuff on it” and a page that feels designed by a person with intent.

Rules it checks against. The references and rules live inside the skill, and the agent checks its own work against them as it builds. That’s the key mechanic: taste stops being a vibe and becomes a checklist the build gets measured against.

References beat prompts — every single time

Here’s the core lesson of this repo, and it’s worth more than the skill itself.

People try to fix AI design with longer prompts. It doesn’t work, because adjectives aren’t a brief. This is the prompt most people write:

Make it look premium

That’s not a brief. “Premium” means something different to every model run, every day of the week. You’ve handed your agent a mood and asked for a design.

A reference is a brief. A reference image your agent can check against — this is the spacing, this is the hierarchy, this is the restraint we’re aiming for — gives the agent something to measure its output against while it works. taste-skill is built on that idea: before the agent designs, it studies the references, and taste goes into every build from the first line of code instead of getting patched on at the end.

Once you understand that, you understand why this repo blew up. Models copy patterns. Feed yours better patterns — award-site references instead of the same five templates — and every build levels up.

How you use design taste for coding agents in practice

The setup is simple: you install it as a skill in your coding agent, following the setup in the repo. Once it’s in, the agent carries that taste into every build in that project.

In practice it changes the workflow in one specific way: the design pass stops being a rescue operation. Without taste, you generate a page, squint at it, and start the “make the hero bigger, no, different font, less purple” loop. With a skill like this installed, the agent starts from references and rules, so the first draft lands closer to designed — and your feedback gets specific instead of desperate.

Try it honestly. Install the skill, rebuild a page you’ve already shipped, and compare. Then open your last AI-built page and be straight with yourself about which one looks like somebody meant it.

Who taste-skill is for — and its limits

This is for anyone whose AI-built sites all look the same: founders shipping landing pages, freelancers building client sites, vibe coders whose products work great and look generic. Anyone can generate a page now. Taste is the difference between a demo that looks generated and a product people trust enough to pay for.

The honest limit: a skill file gives your agent judgment to apply, not a brand to copy. taste-skill supplies the point of view and the references — your product still needs its own identity underneath. Taste tells the agent what good looks like. It doesn’t decide what your good is. That part is still your job, and that’s a good thing.

It also doesn’t replace a design system. Taste judges; a system specifies. In the Repo Week lineup, this is Day 2 of a design trilogy — taste first, then a DESIGN.md design system, then a full design language — and the three stack together. The full series is rounded up here: https://bigjermsays.com/how-to-make-money-with-ai-build-a-system-not-a-side-hustle/, and the tools and templates I build around these workflows live in the shop: https://bigjermsays.com/shop/

The repo, in detail

Repository: github.com/Leonxlnx/taste-skill
Site: tasteskill.dev
License: MIT
Stars: 100,902 as of October 2026

What it is: an agent skill that hands your coding agent art direction — reference images, composition, hierarchy, and whitespace rules — so its design output stops defaulting to templates. You install it as a skill, and before the agent designs, it studies the references in the skill and checks its work against them as it builds.

Who it’s for: builders whose AI-generated pages all look identical and who want the fix to live in the workflow, not in a longer prompt. Read the skill file before you install — that’s true for any repo — then steal the approach even if you end up writing your own version: references and rules your agent can check against beat adjectives, every time.

Note: this article covers the install approach at a workflow level because the source material for the original Repo Week posts describes installation as “install it as a skill in your coding agent, following the setup in the repo” — there is no single install command in that source material, so follow the repo’s own setup steps rather than a command copied from a blog.

FAQ

Is taste-skill free?

Yes. It’s MIT licensed and free to use. You can read the entire skill on GitHub before installing anything.

What coding agents does taste-skill work with?

It’s an agent skill for coding agents — install it as a skill in the agent you use, following the setup in the repo, and the agent applies it to the builds in that project.

Can a skill file really fix AI design?

It fixes the root cause for most generic output: the agent was designing from averages because nobody gave it references or rules. A skill that supplies both — and makes the agent check its work against them — changes the first draft, not just the revision loop.

The bottom line

Your AI’s design has no taste because nobody ever gave it any. This skill is how you hand it some: a point of view, real references, and rules it checks against while it builds. Save this before your next build — your future landing pages will look like it.

Big Jerm