Yesterday I watched an agent take a one-line bug report and produce a 400-line pull request. New abstraction layer, two helper modules, a config option nobody asked for. The fix itself was four lines, buried somewhere in the middle.
That experience is apparently common enough to show up in search data.
The seven fastest-rising Claude Code skill searches this week are all skills that constrain agent output - less code, fewer tokens, better design judgment - not skills that add capability.
Google Trends (past-month window, worldwide, July 20) shows seven "claude code skills" searches rising fast. Here they are, with live install counts:
| Rising search | Trend | What it does | Installs |
|---|---|---|---|
| ui-ux-pro-max-skill | Breakout | design system for agent UI output | 275,112 |
| taste skill | Breakout | metric-driven design taste for frontend work | 271,494 |
| ponytail github | +750% | forces the simplest, most minimal solution | 17,617 |
| stop slop skill | +130% | strips AI writing patterns from prose | 7,686 |
| caveman skill | +110% | claims ~75% token savings without losing technical detail | 364,180 |
| impeccable github | +110% | design polish pass for shipped UIs | 199,134 |
| front end design skill claude code | +90% | Anthropic's own frontend design skill | 682,858 |
Install counts come from the live catalog, refreshed daily from skills.sh, GitHub and MCP registries, ranked by real installs.
Two camps, one complaint
Look at what these seven actually do. Four of them (ui-ux-pro-max, taste, impeccable, frontend-design) are design judgment skills: they exist because agent-generated UIs are functional but generic. The other three (ponytail, caveman, stop-slop) are restraint skills: less code, fewer tokens, less filler prose.
Not one of the seven makes an agent do more.
Earlier waves of agent tooling were about adding capability. This week's risers are all corrective: developers have the capable agent, and now they are searching for the skills that rein it in.
The restraint stack, ranked by installs
If you want to try the "do less" camp, here is the shortlist with real numbers behind it:
- caveman (364,180 installs) - its pitch is roughly 75% less context usage without dropping technical detail. The biggest of the three at 20x the next one, which makes sense: token budgets are a bill, not a preference. Catalog page · GitHub
- ponytail (17,617 installs) - a standing instruction to pick the simplest possible implementation and delete unnecessary code, dependencies and features. This is the one that would have turned my 400-line PR into four. Catalog page · GitHub
- stop-slop (7,686 installs) - removes the recognizable AI writing patterns from any prose your agent produces: docs, commit messages, PR descriptions. Catalog page · GitHub
And the taste camp, for when the output works but looks like every other agent's output:
- frontend-design by Anthropic (682,858 installs) - the most-installed skill in this week's risers. Catalog page · GitHub
- ui-ux-pro-max (275,112 installs, this week's Breakout) - Catalog page · GitHub
- design-taste-frontend (271,494 installs, also Breakout) - Catalog page · GitHub
- impeccable (199,134 installs) - Catalog page · GitHub
The install numbers put restraint in the mainstream: caveman alone is past 364,000 installs.
Why this matters if you run agents daily
The default failure mode of a capable coding agent in mid-2026 is too much code, wrapped in too many words, styled the same as everyone else's. The fix turned out to be standing instructions, packaged as skills, that encode judgment instead of capability.
If your agent's PRs read like my 400-line example, start with one skill from each camp: ponytail for the code volume, and whichever design skill matches your stack for the output quality.
The full ranked list, updated daily by real installs, is on Skillselion - the seven above are live on the homepage rail right now.
I run Skillselion, a directory of Claude Code, Codex and Cursor extensions ranked by real installs. The trend labels come from Google Trends (past month, worldwide, captured July 20); install counts are from the live catalog.
Top comments (0)