OpenClaw + Skills

OpenClaw

• GitHub: openclaw/openclaw — “Your own personal AI assistant. Any OS. Any Platform. The lobster way. 🦞”

OpenClaw is a personal AI agent that showed up like a typhoon (?) in early 2026 (its GitHub star count is insane). Like Hermes Agent, Claude Code, and Codex, it is an AI agent, and a flood of follow-up “Claw” projects such as NemoClaw and NanoClaw have appeared since — it has practically built its own Claw ecosystem 😮.

From my own experience using it: AI agents already existed, but OpenClaw feels like an open-source project that did an exceptionally good job on compatibility with external apps. As for security concerns, I think it should be fine as long as you manage permissions carefully.

It is implemented in TypeScript. If the language is a barrier when reading the code, I also recommend NousResearch/Hermes Agent (“The agent that grows with you”), which is written entirely in Python. Hermes Agent is also one of the best-built open-source AI agent projects.

Plenty of blogs and YouTube videos already explain OpenClaw well, so here are a few posts I personally recommend (though the fastest way to understand it is to just use it):

• OpenClaw official blog: https://openclaw.ai/blog
• OpenClaw SkillHub: https://clawhub.ai/
• 갓대희의 작은공간 blog (Korean): https://goddaehee.tistory.com/504

How Does OpenClaw Select Skills?

My question: Users can install lots of skills as plugins or define their own under the /skills folder. So how does the AI agent pick the right skill to faithfully carry out a user’s query? (Is there some complex internal logic…?)

What I found: The system prompt is extremely long and elaborate. It contains a section labeled # Skills (mandatory), and the LLM in use (e.g., gpt-4.1-mini) looks only at the name and description of each listed skill, selects one, and then reads it. That is why writing a good name and description for each skill is so important.

# Skills (mandatory)
Before replying: scan <available_skills> entries.
If exactly one skill clearly applies: read its SKILL.md at with read, then follow it.
If multiple could apply: choose the most specific one, then read/follow it.
If none clearly apply: do not read any SKILL.md. Constraints: never read more than one skill up front; only read after selecting.
<available_skills>
<skill>
<name> Skill name </name>
<description> Skill description </description>
<location> Path to SKILL.md </location>
</skill>
...
<skill>
...
</skill>
</available_skills>

Seeing this, I now understand why the cloud LLM bill keeps growing when you use OpenClaw continuously. The system prompt fed in is unimaginably long. The volume of input tokens is huge, so a model that can’t handle it well probably won’t work properly…. :(