Jonathon's AI Wiki

Hermes /learn — Turn Any Source into a Skill

5 min read

Source: raw/hermes-learn-skill-from-sources-docs.md (official Nous docs — Skills System) · raw/x-account-nousresearch-2069526242236182697.md (@NousResearch announcement, 2026-06-23) · raw/Hermes_Agent_learn_Skill_BEST_Use_Cases_Bookmark_THIS.md (creator walkthrough, “Ron”)

/learn is a one-command source-to-skill converter in Hermes Agent (announced 2026-06-23, post-v0.17): point it at a source — a local code/doc directory, a doc URL, a PDF, a YouTube/Instagram video, an X post, or a procedure you just walked the agent through — and the agent gathers the material with its existing tools and authors a standards-compliant SKILL.md without you writing one by hand. It’s the automatic counterpart to the manual skill-creator and the conversational “hand Hermes the docs” flow.

How it works

  • Command syntax (verbatim from the docs): 
    /learn the REST client in ~/projects/acme-sdk, focus on auth + pagination
/learn https://docs.example.com/api/quickstart
/learn how I just deployed the staging server
/learn filing an expense: open the portal, New > Expense, attach the receipt, submit
  • Accepted sources: a local SDK/doc directory (read_file / search_files), an online doc page (web_extract), the workflow you just performed in-conversation, or pasted notes. On a VPS, /learn uses yt-dlp to fetch transcripts/metadata from YouTube/Instagram (the agent reads the transcript, it doesn’t watch the video); on a local machine with login creds or an Instagram MCP/OAuth, the agent can actually watch the video for a fuller skill.
  • Output discipline: the generated skill adheres to the documented standards — “≤60-char description, the standard section order, Hermes-tool framing, no invented commands.”
  • Live, no ingestion engine. Because sourcing happens live in-agent, /learn behaves identically in the CLI, TUI, dashboard, and every messaging gateway, on any backend (local/Docker/remote).
  • Persistence + approval. The result is saved via the skill_manage tool, so it follows the same write-approval gate as agent-created skills (skills.write_approval; staged under ~/.hermes/pending/skills/ when the gate is on).

The load-bearing pattern: skill.md as a PRD

The highest-value use from the creator walkthrough: treat the generated skill.md as a PRD and hand it to a coding harness (e.g. Claude Code) to build specs — instead of retyping a plan in a tool’s plan mode, /learn captures ~100% of a source’s takeaway into a structured file you pass downstream. Demoed sources included an Instagram-Reel pairs-trading strategy, an X post on a Kimi agent-swarm self-improving loop, a YouTube video on Karpathy’s loop-engineering method, and a brand-guidelines doc combined with a research skill to draft a script.

Don’t fear skill sprawl: the skill curator consolidates near-duplicate skill.md files into one on a roughly weekly cron, so liberal /learn use stays organized. Pairs with skill bundles (group skills under one slash command) and skill_manage (the agent’s procedural memory) to form a near-automatic skill pipeline.

Community pattern — research before you learn (2026-06-25). A widely-shared refinement (@jumperz) is to add a review space in front of /learn: create a private “Learnables” channel (a Discord thread or Telegram forum-topic per subject), drop a messy topic (e.g. “how should my agent handle Stripe disputes?”), let Hermes assemble a research/context pack first (docs, examples, common mistakes, commands, open questions), then add your own taste/constraints, and only then say “learn this” so it learns the reviewed version, not the first messy thought. The loop is topic → research room → your correction → learn the cleaned version — the review step is where the skill gets good. ^[inferred] (single-creator tip, not official docs)

Key Takeaways

  • /learn <source> turns a URL, PDF, code/doc directory, video, X post, or just-performed procedure into a standards-compliant SKILL.md — no hand-authoring.
  • Sourcing is live in-agent (no separate ingestion engine), so it works identically across CLI / TUI / dashboard / gateways and on any backend; on a VPS it falls back to yt-dlp transcripts, locally it can watch the video.
  • Output is saved through skill_manage and respects the skills.write_approval gate (~/.hermes/pending/skills/).
  • Best use: skill.md as a PRD handed to a coding agent — captures a source’s full takeaway without retyping a plan.
  • The skill curator merges near-duplicate skills weekly, so generating many /learn skills is safe.

Try It

  1. Learn a doc site: /learn https://docs.<tool>.com/quickstart then skills_list() to confirm the new skill landed.
  2. Capture a workflow you just did: after walking Hermes through a multi-step task, run /learn how I just did that to freeze it as a reusable skill.
  3. PRD handoff: /learn a tutorial video, then feed the generated skill.md to Claude Code as the spec for a build — compare against typing the plan by hand.
  4. Turn on the approval gate (skills.write_approval: true) if you want to review every /learn output before it’s saved.

Open Questions

  • Quality ceiling of yt-dlp-transcript-only skills (VPS) vs watched-video skills (local) — how much fidelity is lost?
  • How aggressively does the skill curator merge — can it collapse two genuinely distinct skills that happen to look similar?

Graph View

Backlinks