Skip to content

Skills System

Skills are reusable knowledge and workflow units for Hermes. Think of a skill as a task-specific operating manual that loads when relevant: it stays out of the way most of the time, then teaches the agent how to do a class of work.

A good skill makes repeated work steadier. A bad skill wraps an old workflow in authoritative language and keeps influencing future output.

A task is a good skill candidate when it:

  • Repeats often.
  • Has a clear trigger.
  • Has relatively stable steps.
  • Has describable inputs and outputs.
  • Has explicit things not to do.
  • Has a clear stop or escalation path when it fails.

Release notes, PR review, fixed-format research reports, migration checks, and log analysis are better skill candidates than “write code more intelligently.”

Do not turn a skill into an encyclopedia. It should focus on one repeatable action:

---
name: release-notes
description: Use when preparing release notes from merged pull requests.
---
1. Read merged PR titles and labels.
2. Group user-facing changes first.
3. Mention breaking changes clearly.
4. Keep private review notes out of the changelog.
5. Return a draft; do not publish it.

The important part is the boundary: when to use it, what to read, what to produce, and what not to do.

MechanismBest content
Context filesRules that should apply every time Hermes enters the project.
MemoryStable personal preferences across sessions.
SkillA workflow needed only for a class of tasks.
Project docsFacts, contracts, and decisions humans need to review.

If a rule applies every time, put it in context. If it applies only to one task type, use a skill. If it is a personal preference, use memory. If the team must review it, use project docs.

Temporary research, unreviewed judgments, current task plans, and one-off workarounds should not become skills directly. They may expire tomorrow, but a skill will keep presenting them as workflow.

Before publishing a skill, check:

  • Does it contain stale version numbers?
  • Does it contain private paths, tokens, accounts, or internal links?
  • Could it make the agent perform high-risk actions automatically?
  • Did it turn a one-task constraint into a general rule?

When Hermes becomes stubborn in a surprising way, skills are one of the first places to inspect. A stale skill can be worse than no skill because it looks official.

Practical maintenance habits:

  • Name skills specifically.
  • Keep each skill focused on one task class.
  • Delete unused or stale skills.
  • Put human confirmation points into high-risk skills.
  • After editing a skill, test it with a small task.