ClawHub now hosts almost 50,000 community-built agent skills. For anyone arriving fresh, "most downloaded" is the obvious filter to start with. It's how app stores have worked for fifteen years. Popular things are popular for a reason, right?
We pulled the real download and install data for the top five skills on ClawHub, then spent an afternoon installing every one of them, reading the source code, and actually testing what each one does. The results were not what we expected.
Four out of the five are instruction-only files. Two are by the same developer. One has a typo in its install command (the very first command in its setup section). And exactly one of them did something useful when we typed a command into it.
Before we get to the skills themselves, there's a number worth understanding.
The Download-to-Install Ratio
ClawHub publishes two metrics for every skill: downloads and installs. They sound like the same thing. They aren't.
A "download" on ClawHub is closer to a page view. Someone landed on the skill, clicked into it, maybe read the SKILL.md. An "install" is when someone actually ran clawhub install on their machine. The gap between those two numbers tells you how many people looked at a skill versus how many actually committed to it.
Across the top five most-downloaded skills, there are over 1 million combined downloads and just 16,937 installs. That's an average ratio of 59 downloads per install. For every person who installed one of these skills, 59 others bounced.
Some skills have much wider gaps than others. The widest in the top five has a ratio of 144 to 1. The narrowest is 37 to 1. We'll come back to what that gap actually means after we've looked at each skill.
Here's the leaderboard.
| Rank | Skill | Downloads | Installs | Ratio |
|---|---|---|---|---|
| 1 | self-improving-agent | 358,904 | 5,815 | 61:1 |
| 2 | skill-vetter | 190,826 | 3,774 | 50:1 |
| 3 | ontology | 154,493 | 1,072 | 144:1 |
| 4 | gog | 147,941 | 3,311 | 45:1 |
| 5 | github | 148,893 | 3,969 | 37:1 |
Let's go through them.
1. self-improving-agent
Install: clawhub install self-improving-agent
Downloads: 358,904 | Installs: 5,815
The most downloaded skill on ClawHub is called "self-improving-agent" and bills itself as "a self-evolution engine for AI agents." That sounds remarkable. An agent that gets smarter on its own, learning from its own runtime history, applying protocol-constrained evolution to its own behaviour.
What it actually is: a markdown filing system.
The skill ships with a 500-line SKILL.md that instructs the agent to log errors to .learnings/ERRORS.md, log successes to .learnings/LEARNINGS.md, and log feature requests to .learnings/FEATURE_REQUESTS.md. Each entry gets categorised by priority (critical, high, medium, low) and area (frontend, backend, infra, tests, docs, config). Important learnings can be "promoted" to project config files.
It's a structured note-taking template. There's no runtime intelligence, no automatic analysis, no self-modification. The agent doesn't get smarter. The agent writes things down in a folder, and only if it remembers to.
That's not a fatal criticism. Structured learning capture for agents is genuinely useful, and the SKILL.md is one of the most thoughtfully written we've seen on ClawHub. The skill extraction script is well-coded, with proper input validation and a dry-run mode. The cross-platform support (Claude Code, Codex, Copilot, OpenClaw) is a nice touch. There's a real engineer behind this.
Then we hit the install command in the documentation. It says:
\
clawdhub install self-improving-agent
\\
The CLI is called clawhub, not clawdhub. That's the first command in the setup section, on the most downloaded skill on the platform, and it doesn't work. Anyone copy-pasting it will get a "command not found" error before they've installed anything.
This is the kind of thing that makes you question how many of those 358,904 downloads turned into successful installs. Probably the 5,815 in the install column.
The honest take: Good concept, good code, misleading name, broken install command. The gap between "self-evolution engine" and "markdown folder with templates" is wide enough to drive a truck through.
2. skill-vetter
Install: clawhub install skill-vetter
Downloads: 190,826 | Installs: 3,774
A security scanner for ClawHub skills. Use it before installing anything, the description says, to check for red flags, permission scope, and suspicious patterns. Given the ClawHavoc supply-chain incident earlier this year, when over 1,400 malicious skills were found on the registry, this sounds like exactly what every user should install first.
We installed it. Took about ten seconds. Read the entire skill in two minutes.
It's a checklist. A single SKILL.md file with no code, no scripts, no scanning capability. It tells the agent what to look for: external curl/wget calls, credential file access, eval/exec on untrusted input, obfuscated payloads. It provides a four-step vetting process and a risk classification matrix (low, medium, high, extreme). It includes a report template you can fill out after manually reviewing a skill.
What it does not do is vet anything. There's no automation. The agent still has to read every file in the skill it's vetting and apply the checklist by hand. The "tool" is the agent's existing ability to read code, plus a piece of paper telling it what counts as suspicious.
The irony of installing a security scanner that can't actually scan anything was not lost on us.
To be fair, the checklist itself is solid. The red flags are the right red flags. The risk matrix is clear. If you've never thought about skill security before, reading this is genuinely educational. But calling it a security tool is generous. It's a "How to Spot a Suspicious Skill" PDF in markdown form.
It's also stuck at version 1.0.0 with no updates since launch, and references "ClawdHub" (with an extra D) in one place, the same typo that appears in self-improving-agent. We suspect they share a template or an author.
The honest take: Useful awareness for first-time installers. Zero automation. An agent capable of installing skills unsupervised should already know to grep for suspicious patterns.
3. ontology
Install: clawhub install ontology
Downloads: 154,493 | Installs: 1,072
This is the one. The only skill in the top five where we typed a command and something actually happened.
Ontology gives your agent a memory that actually remembers things. Not chat history, not a notes file, but a structured map of the people, projects, tasks, and events your agent has come across, and how they all connect to each other.
Say you tell your agent: "John from Acme is leading the website redesign project, and the deadline is June 1st." Without ontology, that fact lives in the chat window and disappears the moment the conversation ends. With ontology, the agent stores John as a Person, Acme as an Organisation, the website redesign as a Project, John's role as a relationship between them, and the deadline as a property of the project. A week later you can ask "what's John working on?" or "what deadlines are coming up?" and the agent has real answers, because the facts are linked together rather than scattered through old chat logs.
Out of the box it understands more than fifteen types of things your agent might want to remember: people, organisations, projects, tasks, goals, events, locations, documents, messages, accounts, devices, and several more. You can also define your own rules. For example: "every Task must have a status, and the status can only be 'todo', 'doing', or 'done'." Try to create a task with a status of "maybe" and the skill will catch it.
We tested all of this. Created a Person, queried it back, defined a Task rule, then tried to break the rule on purpose by creating a task with an invalid status. The skill caught it and refused. Thirty seconds from install to working.
It's also genuinely well-built. The code is clean and stays out of trouble. The storage format keeps a full history of every change so you never lose data. It runs on basically anything that can run Python, including a Raspberry Pi, and nothing about it reaches out to the internet.
A few honest caveats. The skill checks your data against your rules when you ask it to, not automatically when you create something, so bad data can slip in if you're not careful. Some of the fancier features mentioned in the documentation (the agent reasoning about cause and effect across your knowledge graph, for instance) are described as future goals rather than working features today. And if your graph grows very large, performance will start to suffer because the skill loads everything into memory at once. None of these are dealbreakers for a personal or small-team setup.
Here's the surprising part. Ontology has the worst download-to-install ratio in the top five, at 144 to 1. For every person who installed it, 144 others looked at the page and left. The reason becomes obvious the moment you open the SKILL.md. It's dense, full of jargon, and assumes you already know what a "typed knowledge graph" is. Most people don't, and they bounce.
There's an irony in that. The most useful skill in the top five is the one most people give up on before installing, purely because the description scares them off.
The honest take: If you want your agent to actually remember the people, projects, and commitments in your work life rather than starting fresh every conversation, install this one. It's the only skill in the top five that earns its place.
4. gog
Install: clawhub install gog
Downloads: 147,941 | Installs: 3,311
Gog gets called the Google Workspace integration for OpenClaw, and we've recommended it ourselves in previous posts. So this one was the most surprising teardown of the five, because the skill itself is barely a skill at all.
It's a 44-line SKILL.md. That's the entire thing. No code, no scripts, no Go binary. The 44 lines are a command reference card for an external tool called gog that you have to install separately via Homebrew, configure with a Google Cloud Platform OAuth project, run through a browser-based consent flow, and authenticate per-account before you can do anything.
If you already have all of that set up, the SKILL.md is a tidy cheat sheet. It documents the syntax for gmail search, calendar create, drive get, sheets append, and so on. The examples are correct. The tip about GOG_ACCOUNT as an env var to avoid passing the account flag every time is useful.
If you don't already have all of that set up, this skill does nothing to help you. There's no setup walkthrough, no troubleshooting guidance, no platform-specific notes. The install command in the docs is brew install steipete/tap/gogcli, which means if you're on Linux or Windows or anywhere without Homebrew, you're on your own. We tested on a Raspberry Pi 5, and our agent couldn't run a single command because the binary wasn't available.
The skill is by Peter Steinberger (steipete on GitHub), the founder of PSPDFKit and a well-known iOS developer with a substantial following. His name carries weight in the developer community, and we strongly suspect that's where most of the 147,941 downloads come from. People see his name, recognise it, click. The actual skill they're downloading is a cheat sheet for someone else's binary that runs only on macOS.
We're not saying gog (the tool) is bad. The underlying CLI looks excellent and the homepage at gogcli.sh is well put together. We're saying that what's on ClawHub isn't really the tool. It's a thin pointer to the tool, masquerading as a skill, hosted on a registry whose download metric counts page views.
The honest take: If you're on macOS and willing to do the OAuth setup, this is your fastest path to using gog. If you're not, the skill is documentation you can't act on.
5. github
Install: clawhub install github
Downloads: 148,893 | Installs: 3,969
If gog is a 44-line cheat sheet, github is a 48-line cheat sheet. It documents five gh commands.
The skill, also by steipete, is a thin instruction wrapper for the GitHub CLI. It covers PR checks, CI run viewing, fetching failed step logs, and a couple of gh api examples with jq filtering. The examples are correct. The --log-failed tip for debugging CI is genuinely useful. The --jq filtering note is something casual gh users might not know.
Then it ends. Forty-eight lines. The skill omits issues, repos, releases, workflows, actions, gists, secrets, environments, and roughly forty other gh subcommands. It's been at version 1.0.0 since launch and has never been updated. Meanwhile the gh CLI itself has added features.
Here's the thing. Running gh --help gives you more than this skill does. Running gh pr --help gives you everything the skill covers plus everything it omits. The skill adds almost no information over what's already in the binary you're using.
So why does it have nearly 149,000 downloads? Same reason as gog. Same author, same pattern, same trust transfer from a reputable developer's name. It's also the safest skill to install in the top five (it just documents an existing tool everyone already trusts), which is probably why it has the best install-to-download ratio in the group at 37:1.
The honest take: It's not bad. It's just barely there. If you've never used the gh CLI, this is a 30-second introduction. If you have, you already know everything in it.
What the Leaderboard Is Actually Telling You
Step back from the individual reviews and a pattern emerges.
Four out of five are instruction-only. They contain no executable code. They are markdown files telling the agent what to do or how to think. That isn't inherently bad (a well-written instruction can shape agent behaviour effectively), but it means most of the "most downloaded skills" are documentation, not tools. ClawHub doesn't visually distinguish between a skill that ships with working code and a skill that ships with a checklist. Both look the same on the leaderboard.
Two are by the same author. Peter Steinberger has two skills in the top five (gog and github), both following the same pattern of thin instruction wrappers for external CLIs, both untouched since launch. This isn't a knock on him personally. PSPDFKit is a serious product and his reputation is earned. But it's a useful reminder that on a registry where a single click counts as a download, name recognition gets compounded into ranking.
The actual #4 on the leaderboard is a near-duplicate of #1. A skill called proactive-agent (135,774 downloads) is by the same author as self-improving-agent and does almost the same thing, with an added layer of proactive learning triggers. We skipped it for this review because reviewing essentially the same skill twice felt like padding. But the ClawHub leaderboard happily lists both as separate top skills.
Only one of them works the way you'd expect a tool to work. Ontology was the only skill where we ran a command and got a result without first installing a separate binary, configuring OAuth, reading 500 lines of templates, or filling out a checklist by hand. It also had the worst download-to-install ratio in the group, because its description scares off browsers.
The most useful skill in the top five is the least committed-to.
So What Should You Actually Install?
Honestly? Don't sort by downloads.
If you're new to ClawHub, the leaderboard will steer you toward skills that are popular for reasons that have nothing to do with whether they'll be useful to you. Reputation drives downloads. Catchy names drive downloads. Being early to the registry drives downloads. Whether the skill actually does something when you install it does not, apparently, drive downloads as much as you'd hope.
Better filters to start with:
Look at install counts, not download counts. If 144 people looked at a skill and 1 installed it, that's a signal. Sometimes it's a signal that the skill is too technical for casual browsers (ontology). More often it's a signal that people clicked, read the SKILL.md, and decided it wasn't worth the disk space.
Open the SKILL.md before installing. If it's 48 lines and just documents an existing CLI, ask whether you actually need the agent to have it pre-loaded. If it's 500 lines of templates, ask whether your agent will follow the instructions or just skim them. The SKILL.md is the skill. There's no hidden value behind it.
Check for actual code. A
scripts/folder with a real executable file means the skill does something deterministic. A solo SKILL.md with no scripts means the skill does whatever the agent's LLM decides to do when it reads the markdown. Both have their place, but they are very different categories of thing, and ClawHub doesn't help you tell them apart.Be sceptical of skills named after grand concepts. "Self-evolution engine" turned out to mean "folder of templates." "Security vetting" turned out to mean "checklist." Names on ClawHub are aspirational. Test what's underneath.
The most popular skill on ClawHub has a typo in its first command. The most useful skill in the top five is the one most people abandon before installing. The leaderboard is telling you something, just not what you think it's telling you.
Sort by downloads with caution. The number is honest about what it measures. It just doesn't measure what you probably wanted.
This post was originally published on The Skill Economy, a weekly look at the AI agent skills market.
Top comments (0)