Agent skills

Ponytail, Caveman, Grill Me and other viral AI agent skills

Ponytail, Caveman, Grill Me, and Find Skills each give an agent one recognizable working habit. Their names explain the behavior, installation takes one command, and users can see the effect in the next task or response.

Install data checked July 21, 2026 · 9 min read

Public skill directories now make adoption visible. The current Skills.sh leaderboard shows millions of cumulative installs for discovery skills and hundreds of thousands for focused workflow skills. GitHub stars, forks, derivative skills, and community discussion add more evidence of attention.

Install telemetry counts installations. Unique users, continued use, task quality, and security require separate evidence. Those limits matter when a leaderboard becomes the main signal for a new skill.

What makes an AI agent skill spread

Across these examples, the fastest-spreading skills give an agent one recognizable working habit.

• A clear job: the skill changes planning, communication, review, discovery, or implementation in a way users can describe in one sentence.
• A memorable name: names such as Caveman and Grill Me communicate the expected interaction before installation.
• A visible result: shorter answers, smaller diffs, or sharper planning questions appear during the first session.
• Portable instructions: a public SKILL.md folder can travel across Claude Code, Codex, Cursor, Windsurf, and other skill-aware agents.
• Simple distribution: a public repository and one install command make sharing easy in posts, chats, and team documentation.
• Easy remixing: users can inspect the instructions, adapt them, and publish a variation for another workflow.

Ponytail: smaller implementations

Ponytail guides an agent through a decision ladder before it writes code. The agent first checks whether the requested behavior already exists, then looks for reusable code, standard library support, native platform features, and installed dependencies. New code comes after those options.

This behavior targets a common coding-agent failure: implementing a full abstraction for a task that the current codebase already handles. The repository describes the persona as a “lazy senior developer,” which gives users a quick way to remember and request the behavior.

Ponytail's author reports about 54% less code across 12 feature tasks using Claude Haiku 4.5, with four runs per task. The repository also explains why an earlier single-shot result was replaced: its baseline generated padded output. Treat the 54% result as an author-run benchmark with a disclosed setup.

The repository had about 86,900 GitHub stars and 4,700 forks on July 21, giving Ponytail one of the strongest repository-attention signals among recent behavior skills.

npx skills add DietrichGebert/ponytail

Caveman: compressed agent responses

Caveman tells an agent to communicate with compressed grammar, symbols, fragments, and minimal repetition. It is useful for users who want short status updates and dense technical answers.

The author reports an average 65% reduction in output tokens across 10 prompts, with results ranging from 22% to 87%. The repository also documents the tradeoff: input and reasoning tokens remain, and loading the skill adds roughly 1,000 to 1,500 input tokens per turn. Short workloads can therefore save little total cost. The immediate effect is a shorter response.

Skills.sh showed about 365,900 cumulative installs for Caveman, while its GitHub repository had about 91,100 stars on July 21.

Ponytail reduces implementation scope. Caveman compresses the explanation around it. Users often combine them because the behaviors apply to different parts of the same coding session.

npx skills add JuliusBrussee/caveman --skill caveman

Grill Me: structured planning questions

Grill Me starts a focused interview about a plan or design. The underlying grilling workflow asks one question at a time, explores the codebase for factual answers, and leaves product or architecture decisions with the user. Each question includes a recommended answer.

The skill is useful before implementation when requirements still contain unresolved choices. A related grill-with-docs skill writes the evolving decisions to a document, while community variants apply the same interview format to learning and review.

Skills.sh showed about 614,600 cumulative installs and 10,700 installs during the previous 24 hours on July 21. The Matt Pocock repository contains many skills. Its GitHub stars describe the full collection; install telemetry supplies the skill-level metric.

npx skills add mattpocock/skills --skill grill-me

Find Skills: discovery as an agent workflow

Find Skills teaches an agent how to recognize a request that may already have a reusable skill. It searches the public directory, checks install activity and source reputation, presents suitable options, and offers the install command.

This meta-skill had about 2.6 million cumulative installs and 17,000 installs in the previous 24 hours. Its adoption reflects a practical problem: skill collections are growing faster than users can browse them manually.

Skillscout offers the same search at browsing time. Open a tool's website and the Chrome extension finds relevant skills for that context. It is handy for quick search when you already know which product or service you are working with.

npx skills add vercel-labs/skills --skill find-skills

Improve Codebase Architecture: guided refactoring

Improve Codebase Architecture helps an agent identify architectural friction, explore the current design, propose several interfaces, and prepare a refactoring plan or RFC. It uses John Ousterhout's deep-module principle to favor small interfaces backed by substantial implementation.

The skill had about 506,800 cumulative installs on July 21. It gives coding agents a repeatable process for architecture work that would otherwise depend heavily on the wording of each prompt.

npx skills add mattpocock/skills --skill improve-codebase-architecture

Other agent skills trending now

The 24-hour leaderboard shows current movement across media generation, design, merge-conflict resolution, database setup, and browser work. These numbers change daily.

How to evaluate a popular skill

Popularity helps with discovery. Installation should follow a source review.

1. Read the complete SKILL.md. Check triggers, commands, file access, external calls, and expected output.
2. Inspect scripts and dependencies. Follow every executable path and review packages added to the environment.
3. Check the publisher. Confirm the repository owner, update history, issue activity, and relationship to the named product.
4. Read benchmark details. Look for model, task set, baseline, sample size, and known limits.
5. Run a representative test. Use a disposable branch or sandbox and compare the result with your existing workflow.
6. Control updates. Pin a reviewed version when the skill can execute commands or access sensitive files.

The AI agent skills security guide covers source checks, local scanners, secrets, dependencies, and agent payment risks in more detail.

Frequently asked questions