Feed Hermes Agent Your Docs — It Builds Its Own Skills

Nous Research added a /learn command to Hermes Agent, their open-source self-improving agent framework. Feed it any source material — a folder of API docs, a URL, a past conversation, or pasted notes — and the agent reads through everything using its own tools, then writes a reusable SKILL.md file that becomes a slash command. You don't write a template; the agent does.

Before /learn, the only way to give Hermes a new skill was to author a SKILL.md file by hand: a structured spec describing what the skill does, how to invoke it, and what tools it uses. That was manageable for one or two things, but onboarding a whole internal API, capturing a multi-step deployment procedure, or documenting a recurring agent workflow meant real manual work every time. /learn hands that job back to the agent. It uses read_file, search_files, and web_extract tools to work through whatever you point it at, then produces a standards-compliant SKILL.md without you touching a template.

What skills actually are in Hermes

Skills in Hermes live in ~/.hermes/skills/. Each one is a directory with a SKILL.md file at its core. The skills index — the catalog of all available commands — stays around 3,000 tokens and is always loaded. Full skill content only loads when the skill is relevant to the current task, so you're not paying for what you're not using. Once installed, a skill becomes a slash command that Hermes can invoke automatically in the middle of a longer task without being asked.

There are now four ways to create a skill: writing SKILL.md by hand, running /learn on material you already have, letting the agent auto-capture after a complex workflow of five or more tool calls, and installing from the community Skills Hub registry. The auto-capture mode is the most interesting: if Hermes finishes a complicated multi-step task, it can optionally write down what it did as a reusable skill — no human involvement.

The broader direction here is worth noticing: agent frameworks are starting to work the way good human teams do. You don't re-train a new teammate from scratch every morning — you give them documentation. /learn is the beginning of that pattern. If the agent can learn from docs today, the next step is the agent updating those docs on its own when the process changes.

The use cases Nous highlights are practical ones: onboarding a private API by running /learn on its documentation, capturing a deployment walkthrough by letting the agent guide you through it once, and grouping recurring tasks into skill bundles that combine multiple skills. All three describe the kind of institutional knowledge that normally lives in someone's head and gets lost when they're not available.