OpenClaw went from a hobby project by an Austrian developer to the most-starred software repository in GitHub history in roughly four months. By March 2026, it sits at 250,000+ stars — more than React, more than Linux, more than anything. Jensen Huang stood on stage at GTC and asked the crowd, “What’s your OpenClaw strategy?” NVIDIA built an enterprise product on top of it. Tencent is using it for WeChat. And at the center of all this is a skills system so simple it’s almost embarrassing — a folder with a markdown file. That’s the thing you need to understand first.
What OpenClaw Actually Is (And Why the Skills System Is the Core of It)
OpenClaw — originally released by Peter Steinberger in November 2025 as Clawdbot, briefly renamed Moltbot after a trademark complaint from Anthropic, then settled on OpenClaw three days after that — is an open-source, locally-running AI agent framework. MIT licensed. Bring your own API key. The red lobster mascot is not ironic. The tagline, “The lobster way,” is not explained. Steinberger has never elaborated on it and that somehow feels right.
The basic architecture: OpenClaw runs on your machine and uses your existing messaging apps — WhatsApp, Telegram, Signal, Discord, iMessage — as the interface. You’re not installing a new chat client. You’re wiring an AI agent into the apps you already live in. It supports Claude, GPT-4o, DeepSeek, Gemini, or local models through Ollama. The model is pluggable. What makes the whole thing coherent is the skills system.
A skill is a directory. Inside that directory is a SKILL.md file. That file tells the agent what the skill does, when to use it, what inputs it expects, and what it should return. OpenClaw ships with 100+ built-in skills. There’s a community registry called ClawHub where people publish and discover skills built by others. The simplicity is deliberate — Steinberger’s background is in iOS development (he founded PSPDFKit, now Nutrient), and the approach feels like it borrowed some of the App Store’s distribution logic while ditching everything that made that ecosystem annoying.
How the Skills System Works in Practice
When you send a message to your OpenClaw agent — say, “Summarize everything in my Notion about Project Atlas and send the key points to the team Slack channel” — the agent doesn’t execute that as a monolithic instruction. It routes across skills. There’s likely a Notion-read skill, a summarization skill, and a Slack-post skill. The SKILL.md for each one defines the contract: what the skill can do, what it needs, what it hands back.
Here’s what a simplified SKILL.md structure looks like conceptually:
- Name: Human-readable label for the skill
- Description: Plain English explanation of what this skill does — this is what the LLM reads to decide whether to invoke it
- Trigger conditions: When should the agent reach for this skill versus another
- Inputs: What parameters or context the skill needs to execute
- Outputs: What the skill returns to the agent for further processing
- Execution: The actual script, API call, or subprocess that runs
The agent uses the LLM to read those skill descriptions and decide which combination to invoke. This means your skill descriptions are doing real work — a vague description will get ignored in favor of a clearer one. Writing good SKILL.md files is closer to writing good prompts than it is to writing good code.
One of the more useful things about the local-first design: skills can access your filesystem, your local apps, your calendar — things that a cloud agent simply can’t reach without complex OAuth flows and permission grants. A skill that reads your local mail client’s database and surfaces action items is something you can build in an afternoon. A skill that monitors a folder and triggers a workflow when a new file appears is five lines of bash and a markdown file.
What to Actually Build: A Skills Framework by Use Case
Not all skills are equal in terms of effort and payoff. Here’s a practical framework for thinking about what to build, ordered from lowest friction to highest complexity.
Tier 1: Information Retrieval and Summarization
These are the skills that deliver value immediately and are easiest to build. They read something and transform it. Examples: summarize your unread newsletters, pull the three most important things from a long PDF, surface key action items from a meeting transcript. The execution logic is usually a single API call or a file read. Almost all the work is in the SKILL.md description — being precise about what “summarize” means in your context.
Tier 2: Notification and Monitoring Skills
Skills that watch something and tell you when it changes. A price tracker for a SaaS tool you’re evaluating. A GitHub watcher that pings you when a specific repo tags a release. A job board scraper that filters by your exact criteria and drops matches into your Signal chat. These require a bit more — usually a cron-style trigger and some state management — but they deliver disproportionate value because they replace a category of tab-checking behavior entirely.
Tier 3: Multi-Step Action Skills
Skills that don’t just read but write, post, or trigger something in another system. Draft and send an email follow-up based on a meeting note. Create a Linear ticket from a Slack message you forward to the agent. These require API credentials, error handling, and — critically — some thought about what happens when the LLM misinterprets the intent. Build human-in-the-loop checkpoints for anything with real-world consequences. A skill that sends emails should confirm before sending, not after.
Tier 4: Agent-to-Agent Skills
This is where it gets interesting and also where you need to be careful. Skills that invoke other agents or hand off to specialist sub-agents. The Moltbook companion app — the social network for AI agents built by Matt Schlicht — was designed partly around this interaction pattern. Agents publishing capabilities that other agents can discover and call. The infrastructure for this exists. The reliability at scale is still genuinely uncertain. If you want to understand how agent-to-agent communication is being standardized at a protocol level, the Model Context Protocol is the piece worth understanding next.
Comparing Built-In vs. Custom vs. ClawHub Skills
| Source | Best For | Trust Level | Customization | Maintenance |
|---|---|---|---|---|
| Built-in Skills (100+) | Common tasks: calendar, file ops, web search, messaging | High — maintained by core project | Fork and modify locally | Updated with OpenClaw releases |
| ClawHub Registry | Specific integrations, niche tools, community workflows | Variable — audit before using | Full — it’s just a directory | Depends on author activity |
| Custom-Built Skills | Internal tools, proprietary systems, specific workflows |