The two fastest-growing Claude Code skill searches right now are both taste layers

Your agent just shipped a working dashboard. It compiles, the tests pass, and it looks exactly like every other AI-generated UI on the internet. Same card grid. Same almost-right spacing.

The model writes correct code all day. What it will not do, unprompted, is care how the result looks.

Developers have apparently noticed. On Skillselion's trending rail, which tracks rising skill searches against a live catalog of 59,000+ skills ranked by real installs, the top two risers this week are both breakout terms, and both of them are taste layers:

1. Ui Ux Pro Max (GitHub) - 275k installs. Instant UI/UX decisions, component recommendations, style guidance.
2. Design Taste Frontend (GitHub) - 271k installs. Its own pitch: "override generic LLM frontend output with intentional, metric-driven UI/UX taste."

Neither one gives the agent a new capability. They exist to strip away defaults.

Five risers, one pattern

Look further down the same rail and the pattern holds:

1. Ponytail (GitHub), +750%: force the simplest, shortest solution, eliminate unnecessary code and dependencies.
2. Stop Slop (GitHub), +130%: strip predictable AI patterns from prose.
3. Caveman (GitHub), +110% and 364k installs: cut token usage ~75% without losing technical detail.

All five risers do the same job: restraint. In pixels, in code, in prose, in tokens.

There is a second thing happening that makes this more than a curiosity. The New Stack recently argued that Cursor, Claude Code and Codex are converging into one composable stack: orchestration, execution, review, mix and match. If the agents are interchangeable, the model is no longer where your output gets its character. The constraint layer is, and unlike the model's defaults, it travels with you from agent to agent.

Two things worth acting on

Treat taste as a dependency, not a vibe. A taste skill is a reviewable rules file: spacing scales, type ramps, what never to generate. You can diff it, version it, and argue with it in a PR. That beats re-prompting "make it look less generic" forty times per project, and it is why the top riser sits at 275k installs: people got tired of writing prompt number forty-one.

Watch installs, not stars. The #1 riser has 108k stars and 275k installs. Stars measure applause. Installs measure need. When the skills whose whole job is saying no outrun everything else in a 59,000+ skill catalog, that is the market telling you what is actually scarce: judgment.

The trending rail refreshes daily at skillselion.com. If you want your agent to pull the right skill mid-task instead of shopping first, there is an MCP server for that: skillselion-mcp.

I work on Skillselion, the directory of Claude Code, Codex and Cursor extensions ranked by real installs. Data above is from the live catalog, refreshed daily from skills.sh, GitHub and MCP registries.