taste-skill is a free, MIT-licensed collection of SKILL.md files that gives AI coding agents a sense of visual taste - stopping them from generating the generic, template-looking interfaces that teams currently spend $3,000 to $15,000 a month to fix with a designer retainer. If your company uses Claude Code, Cursor, or Codex to build product, this repo is a direct line-item threat to one of your recurring vendor invoices.
taste-skill is a free, MIT-licensed file you drop into any AI coding agent - Claude Code, Cursor, Codex - that tells it to stop building ugly interfaces. The paid workflow it replaces is the UI polish retainer: a recurring engagement with a designer or boutique agency that currently runs $3,000 to $15,000 a month for startups and scale-ups that want their AI-built product to stop looking like every other AI-built product.
The problem taste-skill is solving is real and widely felt. When a coding agent builds a user interface from scratch, it defaults to statistically average choices: centered layouts, generic spacing, predictable card patterns, Bootstrap-era visual rhythm. The interface works. It passes QA. It just looks like it was built by a machine that has seen ten thousand websites and averaged them into one. This is what the community has started calling "AI slop," and it is the reason many teams keep a designer on retainer even after they have shipped to AI-assisted development.
The repo, which has accumulated more than 30,000 GitHub stars and gained over 1,400 in a single day this week, does not install software. There is no server. There is no dashboard. A SKILL.md file is a plain-text instruction set that a coding agent loads automatically before it writes any code. taste-skill's default file instructs the agent to read the brief, infer a design language appropriate to the product, and then apply three adjustable parameters: layout variance (how experimental the structure gets), motion intensity (how much animation it adds), and visual density (how much information fits in a single viewport). The agent is forbidden from using em dashes, from building symmetric layouts as a default, and from producing what the project calls "metronomic rhythm" - the tell-tale sameness of paragraphs that all feel the same length.
For a business leader this translates simply: you give the AI better instructions, and it ships better-looking interfaces without anyone doing a design review pass.
Installation takes about thirty seconds. You run npx skills add https://github.com/Leonxlnx/taste-skill in a terminal, or you copy the SKILL.md file into your project folder. Claude Code and Cursor pick it up automatically on the next build. There is no account to create, no API key to manage, and no subscription tier to unlock. The MIT license allows commercial use without restriction.
The repo ships several skill variants beyond the default. There is a redesign-skill for teams that already have a running product and want to improve existing interfaces without rebuilding them. There is a soft-skill for products that need a premium, high-whitespace aesthetic - the kind of look that used to require a Framer agency engagement, which runs $5,000 to $35,000 per project for a polished marketing site or SaaS landing page. There is a minimalist-skill that targets the Notion/Linear aesthetic, and a brutalist-skill for teams that want something intentionally raw. Image-generation skills produce design reference boards that can be passed to Codex or Claude Code for implementation.
The honest limitation here is that a SKILL.md file is only as useful as the agent that reads it. If the codebase is large and the context window is full of other instructions, the design skill competes for attention. taste-skill does not fix bad component structure, accessibility gaps, or mobile responsiveness on its own - those require separate attention. And the aesthetic output depends heavily on how specific your prompt is. The skill gives the agent taste, not judgment. Judgment still lives with the person writing the brief.
What taste-skill does not do is eliminate the need for a designer entirely. Senior product design work - brand systems, user research, information architecture decisions - is not in scope here. What it eliminates is the specific line item that many teams add to their budget after they notice their AI-built interfaces look mediocre: a recurring engagement to review, tweak, and polish what the agent ships. That is the cost this repo directly pressures.
The project also ships image-generation skill variants for ChatGPT Images and Codex that produce design comp references - website frames, mobile screens, brand kit boards - before any code is written. This closes the loop on a workflow that agencies typically charge separately for: design exploration followed by implementation. With taste-skill, both steps sit inside a single agent session at zero incremental cost.
For any company currently paying a monthly retainer to make AI-generated interfaces look less AI-generated, this file is worth thirty seconds of install time and a single test build. The worst outcome is that it helps some and you keep the retainer for what remains. The likely outcome is that the scope of what requires human design review gets smaller.
The irony is that a SKILL.md file - a format invented to make AI agents more capable - is now the thing standing between a design agency and a renewal conversation. That is a sentence no agency roadmap would have predicted two years ago.