The fastest way to burn a Claude Code usage limit is not a hard problem. It is the agent narrating a plan, restating your context, explaining the diff it is about to write, writing it, then summarizing what it just did. Multiply that by every task in a session and the meter runs out while the actual work was maybe a quarter of the tokens.
Developers have started fixing this with skills, and the install numbers show it has quietly become its own category. Three of the risers on our trending rail this week are not capability skills at all. They are muzzle skills: their whole job is to make the agent produce less.
The numbers
caveman (GitHub) - 374,542 installs, 92,125 stars, search interest up 110% over the last 30 days. It compresses agent output across replies, commits, and reviews, claiming roughly a 75% token cut while keeping every technical detail. The pitch line that lands: answers that read like a senior engineer's Slack message, not a tutorial.
ponytail (GitHub) - 19,421 installs, 87,947 stars, and the sharpest momentum of the three: search interest for "ponytail github" is up 750%. It is a "lazy senior developer" skill that forces the simplest possible solution through a seven-rung decision ladder (do we need this? does it exist? is it in the stdlib?) before any code gets written. Less output here means fewer speculative features and smaller diffs, not just shorter prose.
stop-slop (GitHub) - 7,969 installs, 14,269 stars, search interest up 130%. Narrower target: the recognizable AI prose patterns themselves. Where caveman compresses everything, stop-slop rewrites the tics.
All three numbers come from our live catalog, refreshed daily from skills.sh, GitHub and MCP registries, ranked by real installs.
Why this category exists now
Two weeks ago the risers were taste skills: design-judgment layers like ui-ux-pro-max (now past 280,000 installs) that constrain what the agent builds. The token-diet skills are the same move one level down. Developers are not asking their agents for more capability. The models are capable enough. They are asking for less output per unit of work, because output is what they pay for, wait for, and have to read.
There is a real cost argument (a 75% token cut on a metered plan is not cosmetic) but I think the stronger driver is review fatigue. Every sentence the agent writes is something you have to at least skim. Verbose agents shift work from the model to you. A skill that deletes the narration deletes that review load with it.
One sentence to take away: the highest-leverage Claude Code skill right now might be the one that makes it do less, and 374,000 caveman installs suggest that is not a niche opinion.
If you want to watch this category move, the trending skills rail updates from live Google search data, and the Code Review & Quality category is where most of these land.
I work on Skillselion, a directory of Claude Code, Codex and Cursor extensions ranked by real installs.
Top comments (0)