Leonxlnx/taste-skill

The problem

Coding agents can produce a working page in minutes, but left alone they converge on the same hero-plus-three-cards layout, stock gradients and placeholder imagery. The work is technically done and visually forgettable, and the developer still has to art-direct it afterwards. Taste Skill packages art direction, real reference images and a browser-based check as an installable agent skill so the agent does that work during the build rather than leaving it to you.

What's different

The usual approach is a prompt like "make it beautiful" plus a design system the model half-remembers. Taste Skill instead encodes a process: a script deals three of twenty researched directions, the agent commits to one idea tying the brief to that direction, pulls credited images from open archives, then reviews its own screenshots in Chrome at three widths and fixes the top problems over a few rounds. Compared with generic UI component kits or template generators, the output is not a fixed set of blocks; compared with asking the model for a style, the check step gives the agent a way to fail its own page before you see it. It also carries a separate route for dashboards and app screens with rules for charts, tables and mobile layout.

Who it's for

Frontend and full-stack developers using Claude Code, Codex, Cursor or similar agents to build marketing pages, landing pages and app shells who care about visual quality but do not want to hand art-direct each attempt. Teams standardising agent output across many small sites will get the most from it. It is a poor fit for design-system-heavy products that already have strict component libraries, for backends, and for anyone who needs deterministic, diff-stable output: the skill deliberately picks among directions, so results vary run to run.

Cost of adoption

Requires Node 22+ and an installed Chrome, Edge or Chromium; image sourcing needs network access. No API keys and nothing installed beyond the skill itself. Installs via the npx skills CLI or as a Claude Code / Codex plugin, and the README notes each build runs roughly 30-45 minutes and a few hundred thousand tokens, which is the real cost. MIT licensed. The README is unusually detailed about method but the project is young, and older skills have already been split into separate repos, so expect structure to shift. The README's own benchmark claims are not independently verifiable here.

Verdict

Taste Skill is a genuinely new project, created in February 2026, that is riding a sharp spike: roughly 12,500 stars on the monthly JavaScript board and 3,800 in a week, with 100k stars total but essentially no Hacker News discussion and no daily trending history. That pattern says a viral moment more than a settled tool. The idea, moving design judgement into an agent skill with a browser-based self-check, is the right shape for where coding agents are heading. Adopt it experimentally on low-stakes pages, verify the browser check works in your environment, and do not yet build a workflow that depends on its internal structure staying put.