How Hermes skills work

Hermes skills are on-demand knowledge documents: Markdown files the agent loads only when it needs them, following a progressive-disclosure pattern to keep tokens down. They live in one directory under the Hermes home, every installed skill is a slash command, and the agent can write, patch and delete its own. That last part is what the docs call procedural memory, and it is the reason Hermes skills feel different from OpenClaw’s.

Three levels of loading

• Level zero: the list of names, descriptions and categories, a few thousand tokens.
• Level one: the full skill, loaded when it is relevant.
• Level two: a specific reference file inside the skill, loaded when a question needs it.

Every skill is a slash command, and up to five can be stacked at the start of one message; parsing stops at the first token that is not a skill, so a file path is never swallowed. Bundles put a repeated combination under one short command. A skill can restrict itself to an operating system, hide itself when a toolset is available, or show only when one is.

The SKILL.md a human writes

Frontmatter with a name, a description, a version and optional platform and Hermes metadata, then the four sections the house standard expects: when to use it, the procedure in order, the pitfalls, and how to verify it worked. A learn command turns anything you can describe, a local SDK, a doc page, the workflow you just walked through, a whole book, into a skill in that shape; a large source becomes a lean SKILL.md with one distilled reference file per chapter.

The skills the agent writes

When it works out a non-trivial workflow, hits a dead end and finds the path, or is corrected, the agent records the approach with its skill tool. The docs are precise about what a good entry is: lessons rather than logs, a rule plus one clause of why, no incident narration. A write-approval setting stages every skill write for review, in the CLI or on a messaging platform, before it lands.

A skill is the instructions for doing a class of task the most efficient and correct way, to your specifications.

The Hub

The Skills Hub browses, searches and installs from online registries, the skills.sh directory, well-known endpoints on any site, official optional skills and direct URLs, with a security scan on install and an update check for what came from upstream. Hermes vs OpenClaw compares the two engines’ skill systems alongside everything else.

On Diali

The catalog shows which skills run on Hermes, with the same review as for OpenClaw, and installs them into the assistant’s sandbox; the skills the agent writes for itself stay in its workspace, which the Backups add-on (included on Max) snapshots daily. Hosted Hermes on Diali is the hosted version; The Diali skills catalog is the catalog.

• Load the list, then the skill, then the file.
• Every skill is a command; five can stack.
• The agent writes its own, and you can gate the writes.