01 / Discover
List and open the skill
Use the Skills System, a GitHub tap, ClawHub, a direct URL, or an external directory. Record the origin before copying anything.
Source-checked · July 31, 2026
Hermes Agent Skills: Create, Review, Install, and Share Workflows
A practical lifecycle for working with SKILL.md files: inspect the source, verify requirements, install deliberately, test in a narrow scope, approve agent-authored changes, and share only what your team can review.
CreateReviewInstallTestShare
The skills lifecycle
One reviewable path from discovery to team reuse
A skill is executable procedure, not just a prompt. Treat origin, requirements, permissions, and change history as part of the artifact.
Direct answer
Install only after you can explain what the skill reads, writes, calls, and asks the agent to remember.
02 / Inspect
Review source and requirements
Read SKILL.md, referenced files, commands, environment variables, network targets, and the tools the procedure expects.
03 / Install
Place it in a known scope
Hermes reads local skills from ~/.hermes/skills/. Prefer a narrow, reversible install before sharing it across machines.
04 / Test
Run one bounded task
Use the least-privileged environment that still exercises the workflow. Confirm both the expected output and the stop path.
05 / Approve
Review agent-authored changes
When write approval is enabled, proposed skill changes can remain staged under ~/.hermes/pending/skills/ before they become active.
06 / Share
Publish the evidence with the skill
Share the source, requirements, tested environment, and known limits—not just a Markdown file that happens to run.
SKILL.md anatomy
Readable procedure, explicit boundary
Hermes follows the Agent Skills-compatible SKILL.md pattern. Frontmatter makes the skill discoverable; the body should explain when to use it, what to inspect, which steps to follow, and where to stop.
---
name: review-repository-change
description: Inspect a proposed change before it is accepted.
---
1. Read the source and referenced files.
2. Identify permissions and external calls.
3. Test the smallest reversible case.
4. Stop when evidence is incomplete.Approval + safety
Do not confuse “installed” with “trusted”
Hermes can help write and improve skills, but the operator still owns the permission boundary and the final acceptance decision.
Agent-authored change flow
Stage first. Accept later.
When skills.write_approval is enabled, proposed changes can be held in ~/.hermes/pending/skills/ for review.
- 01
Proposed
The agent prepares a new or changed skill.
- 02
Staged
Files remain pending instead of active.
- 03
Reviewed
The operator checks the source, permissions, and diff.
- 04
Accepted
Only the approved version enters the active skills directory.
Supported source routes
Origin is part of the evidence
Hermes documents several ways to obtain skills. The route changes what you need to verify.
- Bundled or optional
- Check the installed version and local files.
- GitHub tap
- Inspect the repository, branch, and commit history.
- ClawHub
- Verify the publisher and downloaded artifact.
- Direct URL
- Treat the URL and fetched content as untrusted input.
- External directory
- Confirm who controls updates and filesystem access.
Security boundary
In-process checks are heuristics. Use OS-level isolation when the model or skill may be adversarial.
Primary sources + quick answers
Check the source date before you copy the workflow
Source-checked 2026-07-31 · official material, not copied vendor claims
Quick answers
Where do Hermes skills live?
The default local skills directory is ~/.hermes/skills/.
Can Hermes create or edit a skill?
Yes. The documented skill_manage tool supports creating and changing skill files. When write approval is enabled, proposed changes can be staged for review.
Does installing a skill mean it is safe?
No. Review the origin, referenced files, required tools, permissions, external calls, and isolation boundary before activation.
Agent Skills Hub · evidence-first guide