Hermes Agent skills: how they work and how to add more

What a skill is

A skill is a folder with one required file, SKILL.md. The top of the file holds a name and a one-line description of when to use it; the rest is the method itself, usually in the same order: when to use it, the steps, the pitfalls, and how to check the result. A skill can also carry extra files in references/, templates/, scripts/ and assets/ folders, such as a style guide, a report template or a small script.

Hermes does not read every skill all the time. It first sees only the list of names and descriptions, opens the full skill when a job matches, and pulls in an extra file only if a step needs it. So a large library costs far less than its full size would suggest, though the list of names and descriptions itself still grows with every skill you add.

Skills live in ~/.hermes/skills/ on the machine that runs Hermes, one folder per skill, optionally grouped into category folders. The format follows the open Agent Skills standard (agentskills.io), so skills written for other agents that use the same standard generally work too. In the terminal and in connected chat apps, every installed skill also becomes a slash command: typing / and the skill’s name runs it directly.

The skills that come built in

A fresh install of Hermes Agent 0.21.5 comes with 58 bundled skills, sorted into categories such as productivity, email, research, note-taking, media, social media, web, software development and DevOps. A few that matter for everyday business work:

• Documents: reading and writing PDF, Word, Excel and PowerPoint files.
• Meetings and planning: turning a document or meeting notes into action items, and a weekly review and planning routine.
• Email: sorting an inbox by priority and drafting replies for you to check.
• Research: monitoring competitor news and writing answers with sources attached.
• Apps: Google Workspace (Gmail, Calendar, Drive, Docs, Sheets), Notion, Airtable and Microsoft Teams meeting summaries.

On top of those, Nous Research maintains about 150 optional official skills that are not installed by default, in areas such as finance, health, security, smart home and payments. Any of them installs with one command, for example hermes skills install official/<category>/<name>.

How Hermes writes and improves its own skills

This is the part people mean when they say Hermes “learns”. The agent has a tool for creating, editing and deleting its own skills, and its instructions encourage it to save a reusable method in three situations: after a job that took several steps, after it fixes a mistake or finds a way round a dead end, and after you correct how it did something. It decides case by case, so not every job turns into a skill.

It also checks itself. After a stretch of tool use without saving anything, Hermes runs a short review in the background once it has sent its reply, so you never wait for it. The review looks back at the work and decides whether a new skill, or a fix to an existing one, is worth keeping. When a skill leads it astray, it is told to patch that skill rather than start a new one, so a method can get better with use.

To stop old skills piling up, a curator checks about once a week while the agent is idle. By default it only looks after skills the agent saved on its own in the background review: one that nobody has used, opened or edited for 14 days is marked stale, and after 30 days it is archived. Archived skills are moved aside, not deleted, so they can be brought back. Skills you taught it yourself are left alone unless you hand them to the curator with hermes curator adopt, and pinned skills are always skipped.

If you would rather approve every new skill first, set skills.write_approval: true. New and changed skills then wait in a pending queue, and you review them with /skills pending, /skills diff, /skills approve and /skills reject.

Teaching it a skill on purpose

You do not have to wait for Hermes to decide a job is worth saving. The /learn command turns almost anything into a skill:

• /learn how I just did the monthly invoice run saves the steps you just walked it through.
• /learn https://example.com/our-style-guide reads a web page and turns it into a skill.
• /learn filing an expense: open the portal, New, Expense, attach the receipt, submit turns notes you paste into a skill.

A large source, such as a book, a long manual or a folder of documents, becomes a short skill file plus reference files the agent opens only when needed, rather than one giant page. Plain words work as well: “save how you wrote that client report as a skill” has the same effect.

Installing skills from the Skills Hub

Hermes can search for and install skills written by other people. The commands are hermes skills search, browse, inspect, install, update and uninstall, run in the terminal (or as /skills inside the terminal chat; in connected chat apps /skills only handles approving new skills). Sources it knows out of the box include:

• Nous Research’s own optional skills,
• skill repositories on GitHub from OpenAI, Anthropic, Hugging Face and NVIDIA, among others,
• public skill directories such as skills.sh, ClawHub and LobeHub,
• websites that publish their own skills, and a direct link to any SKILL.md.

A skill is instructions your agent will follow and sometimes code it will run, so treat one from a stranger like any download. Hermes helps here: every hub install is first unpacked into quarantine and scanned for things like attempts to send your data elsewhere, hidden instructions and destructive commands. What happens next depends on where the skill comes from. Nous Research’s own skills count as built-in. Skills from the OpenAI, Anthropic, Hugging Face and NVIDIA repositories count as trusted, and everything else counts as community, except that a skills.sh entry takes the trust of the GitHub repository behind it. A third-party skill judged dangerous is refused, and --force cannot override that; for community skills it only gets past milder warnings.

Two habits worth keeping: run hermes skills inspect before installing a community skill, and remember that hermes skills update leaves alone any skill you have edited yourself.

Turning skills off and other settings

• skills.disabled lists skills the agent should ignore, and skills.platform_disabled switches skills off for one chat app only.
• skills.creation_nudge_interval sets how much tool use passes before the background review runs. Set it to 0 to turn off the automatic background review; the agent can still save a skill during a conversation.
• hermes skills opt-out stops Hermes adding bundled skills in future without touching the ones you have; add --remove to also delete the bundled skills you have not edited. hermes skills opt-in --sync brings them back.
• skills.external_dirs adds skill folders shared with other tools; if two skills share a name, the one in Hermes’s own folder wins.

Skills on ReadyHermes

ReadyHermes runs your own Hermes Agent on our servers, always on, without a server for you to look after. For skills, that means:

• Your agent starts with the bundled skills, and the skills it saves belong to your agent only, not to other customers’ agents.
• On a paid plan the background review is on, so it can save and improve skills as it works. During the free trial hour that background review is paused, because each review uses AI from the trial allowance; you can still teach it a skill on purpose.
• To teach it on purpose, ask in plain words anywhere, including the workspace chat: “save how you did that as a skill”. From Telegram, WhatsApp or another connected chat app, /learn works too.
• Your skills are included in backups: in Bot tools, open Backup and tick Skills to download them with your conversations and files.

New to Hermes? Start with what Hermes Agent is, or see how to install it yourself if you would rather run it on your own computer.