The Hermes /learn command teaches your agent a new skill from any source in one line: you type /learn, point it at the thing you want mastered, and the agent writes itself a reusable skill.md file that it keeps forever. There is no training run and nothing extra to install — Hermes reads the source with the tools it already has and does the writing on its own. This page is the hands-on usage guide, from the update you run first through to loading the finished skill on every surface.
📺 Watch: Learn Anything Engine for Hermes AgentOS
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
If you want the bigger picture of why teaching beats re-briefing every session, my /learn overview makes the full case. This page stays practical, built directly on the demos from my own June sessions.
Before You Run the Learn Command: Update Hermes First
One prerequisite, and it takes under a minute: run hermes update in your terminal before anything else. The command ships with recent builds, so an out-of-date install simply will not have it — if you type /learn and nothing happens, this is almost always the reason. Update, start a fresh session, and the command is there in every chat from then on. It is the dullest step on this page and the one people most reliably skip.
How to Use the Hermes /learn Command: Point It at Almost Anything
Usage really is one line. In any Hermes chat, type /learn followed by the source you want turned into a skill. Plain English is fine — you are describing where the knowledge lives, not writing syntax. The sources that work:
- A URL — a guide, a documentation page, any article worth keeping.
- A document or a PDF — the same reading ability I cover in my PDF reader guide applies here.
- A folder of code — point it at a project and let it work out how the pieces fit together.
- Pasted notes in plain words — your own rough process notes are a perfectly good source.
- A job the agent just did — reference the task it has only just finished and it distils that work into a repeatable skill.
Whatever you point it at, the mechanics are identical: Hermes reads the source using the tools it already has — there is no separate learning engine bolted on and no extra installs — and then writes a clean skill.md file on its own. You brief, it studies, it documents.
My first demo: one URL, one saved skill
In my live demo I typed /learn followed by the URL of my Agent OS guide. It read the page, broke the system down properly — the dashboard, the six pillars, the surfaces, the setup path — and saved the skill first time. Nothing came back for babysitting; it went from link to finished file in a single pass.
My second demo: the terminal and the win log
The command line works just as well as the chat window. I ran /learn on my daily win log workflow — the routine I use to record each day's results — and it simply learned it, writing the full skill file without me touching a line. Two completely different sources, the same one-line usage, the same finished file at the end.
Want to watch this done before you try it? Inside my free community I share the full teaching walkthrough — the sources I use, the skills I have built, and the ones members are publishing for each other. Join AI Profit Lab and teach your agent its first skill alongside people doing the same thing this week.
The Approval Gate: Seeing a Skill Before /learn Saves It
What happens at the save depends on one setting. With the approval gate on, the skill only saves after you approve the write — Hermes drafts the file, shows it to you, and waits. You see exactly what it intends to keep before it keeps anything, which is the right arrangement while you are still building trust in the output. I run without the gate, which is why the skills in my demos saved instantly; once you have reviewed a dozen drafts and changed nothing, the gate becomes ceremony.
Either way, what lands on disk is a readable markdown file, and knowing how to read it makes you a better teacher. My skill.md deep dive walks through the file itself, so you know precisely what you are approving.
📺 Watch: Hermes AI Just Learned to Read Books
Loading a Skill: /skill and the Auto-Load Triggers
A saved skill loads two ways. The deliberate way: type /skill followed by the skill name and it loads on demand. The effortless way: skills also auto-load when you mention related terms in conversation. My Agent OS skill loads the moment I mention agent OS, mission control or memory galaxy — no command, no ceremony, the knowledge just arrives.
After that, you simply ask for the task by name. The agent already knows how, because everything /learn captured is sitting in the loaded skill. That is the payoff moment: a request that used to need a paragraph of briefing shrinks to a few words.
Learn Once, Use It on Every Surface
A skill is not tied to the chat where you taught it. Learned once, it is available everywhere Hermes runs: the dashboard chat, the command line, voice through Hermes Jarvis — my voice agent guide covers that setup — and messaging, which my bot mode guide handles in full. Teach the win log skill at your desk and ask for it by name from your phone; it is the same skill answering. That is the quiet superpower of the /learn command: one command, learned once, available on every surface you use.
Tips for Better Skills from /learn
- A clear, specific source beats a vague one. Point the command at the exact page, the exact folder, the exact notes. Fuzzy input produces a fuzzy skill, and you will only end up re-teaching it later.
- Route tricky jobs to a stronger model. For a genuinely complex workflow, have /learn run on the strongest model available to you — the quality of the written skill follows the quality of the model doing the studying.
- Do not assume it needs one. In my Goldie Bench testing, a free local model wrote a skill in about 25 seconds. This is not a heavy job, so save the big model for the genuinely hard teaches.
The habits around the edges — clean briefs, session hygiene, when to double-check output — are in my best practices guide, and all of them apply on teaching days.
When /learn Gets It Wrong: Fixing and Re-Learning
Sooner or later a skill will come out wrong — too shallow, slightly off, or built from a source that was messier than you realised. The fix is refreshingly boring, because the skill is just a markdown note. Edit the line that is wrong. Delete the file if it is beyond saving. Or re-run /learn with a clearer source and let it write the lesson properly this time. Nothing locks in, and a wrong skill cannot break anything else — the worst case is a bad note, corrected in less time than it took to read this paragraph.
📺 Watch: Hermes3D Is INSANE: AI Agents Now Have a 3D Office
What to Teach First
My recommendation is unglamorous and specific: point /learn at the workflow you re-explain most. If there is a job you repeat weekly and find yourself briefing from scratch every session, that is the one — it currently costs you a few minutes every single run, and teaching it costs five minutes once. That was the logic behind my win log demo, and it is the fastest way to feel the compounding start.
I pair this with the seeding routine in my memory setup guide: memory carries who you are, skills carry how your jobs are done, and an agent holding both stops needing the same briefing twice. Skills travel, too — publish yours, or install someone else's with a single command, so the best version of a workflow only has to be taught once by anyone.
The /learn Workflow at a Glance
| Step | What you type or do |
|---|---|
| Update your install | Run hermes update in the terminal so the command exists |
| Teach the skill | In any chat, type /learn followed by the source — a link, a file, a folder or plain notes |
| Approve the save | With the gate on, read the drafted file and approve the write |
| Load it deliberately | Type /skill followed by the skill name |
| Or let it auto-load | Mention a related term and the skill loads itself |
| Use it | Ask for the task by name — the agent already knows how |
The /learn Command FAQ
Do I need to update Hermes before the command works?
Yes. Run hermes update first — the command ships with recent builds, and an older install will not have it. It is the only setup step there is, and it takes under a minute.
What sources does /learn actually accept?
A URL, a document, a PDF, a folder of code, pasted notes in plain words, or a reference to a job the agent has just done. Plain English descriptions are fine — there is no syntax to memorise.
Does every new skill need my approval?
Only if the approval gate is on. Gated, the skill saves after you approve the write, and you see the full file before it saves. I run without the gate, which is why mine save instantly.
Can I use a learned skill by voice?
Yes. A skill learned once is available on every surface, including voice through Hermes Jarvis — teach it at the keyboard, then ask for it out loud.
Is it safe if a skill comes out wrong?
Yes. A skill is a markdown note: edit it, delete it, or run the teach again with a clearer source. Nothing locks in, and a bad skill cannot break anything beyond its own output.
My Verdict on the /learn Command
This is the rare feature where the honest description sounds like the sales pitch: one line in, a permanent capability out. The usage barely needs a guide — update, point, approve, load — which is exactly why the leverage sits in what you choose to teach rather than how you type it. Pick the workflow you are sick of explaining, run the teach tonight, and ask for the job by name tomorrow.
Want my first five skills to copy? Inside my free community I share the exact teaches I ran first — the sources, the wording and the finished files — so you can skip the trial and error entirely. Join AI Profit Lab and have your agent trained on your best workflow by the end of the week.
