1. What Are Hermes Agent Skills?
Skills are versioned, self-improving procedures the agent stores as markdown files — one SKILL.md per skill, with YAML frontmatter at the top and the procedure written out below it, kept in ~/.hermes/skills/. Think of a skill as a saved "how I solved this last time" note that the agent can read back before it starts working. If you want the wider picture of where skills sit in the system, our complete guide to Hermes Agent by Nous Research walks through the architecture around them.
The reason this matters: most agent frameworks cold-start every task, re-deriving the same steps from scratch each time. Skills are Hermes's procedural memory — the differentiator that lets a repeat task reuse a known-good path instead of re-exploring it.
2. What Makes Skills "Self-Improving"?
Two behaviors earn the "self-improving" label. First, the agent auto-creates skills from completed workflows — when it finishes a task, the solution path is a candidate to be saved. Second, a background self-improvement review runs after a turn, distilling what just happened into a reusable skill without you asking for it.
The payoff is concrete: because the next similar task can load an existing skill, the exploration phase is short-circuited — fewer tokens, fewer wrong turns, faster and cheaper completion. That is the whole point of procedural memory, and it is the loop most serious agent stacks are moving toward. If the word "agentic" is doing a lot of work here, this behavior is a textbook example of it in shipping code.
3. Where Are Skills Stored, and How Are They Organized?
Skills live in ~/.hermes/skills/. Each skill is a directory containing a SKILL.md file, and you can group related skills into subdirectories to keep the folder legible as it grows:
Agent-created skills are written into ~/.hermes/skills/. Existing skills are modified in place — including any that live under configured external_dirs, so shared or external skill folders update where they already are rather than being copied.
4. How Do You Manage Skills? The skill_manage Actions
Both you and the agent manage skills through skill_manage actions. Here is what each one does:
| Action | What it does |
|---|---|
| write | Create a brand-new skill (the agent auto-writes these into ~/.hermes/skills/) |
| edit | Rewrite an existing skill's content in place |
| patch | Apply a targeted change to part of a skill without rewriting the whole file |
| delete | Remove a skill you no longer want the agent to reach for |
| write_file | Add a supporting file inside a skill directory (scripts, templates, notes) |
| remove_file | Delete a supporting file from a skill directory |
In practice, the agent leans on write to save what it just learned, then edit and patch to refine a skill over time as it hits edge cases — which is the "versioned" part of the definition doing its job.
5. How Do You Load a Skill?
When you want the agent to reach for a specific skill on the next task, load it explicitly. From the shell, pass the skill name with -s; inside the CLI, use the slash command:
This is how you steer the agent toward a proven procedure instead of hoping it rediscovers it. Haven't installed Hermes yet? Our step-by-step guide to installing Hermes Agent covers every platform, and skills work the same way regardless of how you installed it.
6. Manual vs. Auto Skill Creation
You do not have to wait for the agent to write a skill — you can author one by hand. Create the skill directory and a SKILL.md file with YAML frontmatter, or scaffold it with the CLI:
The two creation paths coexist: hand-written skills capture procedures you already know, while auto-created skills capture what the agent discovers while working. Both end up as the same SKILL.md format in the same folder.
7. Staying in Control: write_approval
A self-writing agent editing files on your machine deserves a safety valve, and Hermes has one. When write_approval is enabled, every skill_manage write is staged rather than committed — the change is held for your review instead of the agent writing it freely. You get to see exactly what the agent wants to save before it lands in ~/.hermes/skills/.
If you are still weighing whether Hermes's approach fits your workflow, our Hermes Agent vs OpenClaw comparison puts its memory and control model side by side with the alternative.
8. Sharing Skills: The Skills Hub and the Open Standard
Skills are portable. Because they are plain markdown files, you can share them with others through the Skills Hub, and they are compatible with the open standard at agentskills.io. That means a procedure your agent learned can be handed to a teammate's agent — the accumulated know-how is not locked inside one install.
9. Frequently Asked Questions
What is a Hermes Agent skill?
It is a versioned, self-improving procedure stored as a SKILL.md markdown file (with YAML frontmatter) in ~/.hermes/skills/. The agent can write skills automatically and reuse them on later tasks so it does not have to re-solve the same problem from scratch.
Does Hermes create skills automatically?
Yes. The agent auto-creates skills from completed workflows, and a background self-improvement review runs after a turn to distill what it just did into a reusable skill. You can also write skills manually.
How do I load a specific skill?
Launch with hermes -s "skill-name", or run /skill skill-name from inside the CLI. Skills you (or the agent) have saved live in ~/.hermes/skills/.
Can I stop the agent from writing skills without my approval?
Yes. Enable write_approval and every skill_manage write is staged for your review instead of being committed automatically, so nothing lands in your skills folder until you approve it.