Hermes Agent /learn Skill BEST Use Cases (Bookmark THIS) — Superbash Learn

Superbash Learn

Click to load the YouTube player

Key takeaways

Use `yt-dlp` as the fallback fetcher when running Hermes on a VPS — the agent cannot natively watch Instagram Reels or YouTube videos in that environment.
Treat every `/learn` output as a PRD and hand it to Claude Code for indicator and trading-strategy builds rather than invoking the skill inside Hermes.
Prefer `/learn` over Cursor plan mode when the source is a Reel or short video — you keep 100% of the captured info instead of paraphrasing it in natural language.
Don't manually deduplicate overlapping `skill.md` files; the skill curator runs roughly weekly and merges duplicates like the Karpathy loop and Kimi K2.6 swarm loop skills without breaking the folder.
Chain `/learn` with `skill bundle` — feed a brand guideline (e.g. HKUST) plus a research skill to auto-generate copy that follows the brand's selling points.
Use `/learn` on every bookmark the moment you save it — Hermes behavior is identical across CLI, TUI, dashboard, and messaging gateways, so there is no separate ingestion step.
Skip `/learn` for empty or low-signal bookmarks — Ron's rule is "preferably if there's a meaningful outcome," otherwise you just bloat the skill folder.

Video summary

Companion notes

/learn turns any URL, PDF, or reel into a skill.md you can hand straight to a coding agent.

What the command actually does

You run /learn, point it at a source — Instagram Reel, X post, YouTube video, PDF — and Hermes fetches it and writes a skill.md file. On a VPS the agent can't watch video natively, so it falls back to yt-dlp to pull transcripts and metadata. With local login, MCP, or OAuth for Instagram it can actually view the video. The same behavior is consistent across CLI, TUI, dashboard, and messaging gateways — there is no separate ingestion engine or offline index.

The PRD shortcut

Ron treats each skill.md as a PRD rather than invoking the skill directly inside Hermes. He feeds it into Claude Code to build initial indicator specs. Reels or X posts with dense info — like a co-integrated pairs trading quant strategy or Kimi K2.6 self-improving agent swarm loops — get captured verbatim instead of being paraphrased into Cursor's plan mode with grammar mistakes. Result: less manual typing, full fidelity to the source.

Real workflows shown

IG Reel → trading PRD: yt-dlp transcribed the audio, Hermes output a Pairs Trading Basics skill with prerequisites, procedure, and quick reference, then handed off to Claude Code.
X post → production loop: A post about Kimi K2.6 self-improving loops was applied to the existing AI research pipeline on the server.
YouTube → Karpathy loop engineering: yt-dlp pulled metadata from a Karpathy video and the agent mapped it onto current Boxmining AI production.

Cleanup and overlap

The skill curator runs roughly every week to merge similar files — overlap between the Karpathy loop skill and the Kimi swarm loop skill is expected. Don't over-curate manually; let the curator combine duplicates.

Beyond trading

Feed a university brand guideline (e.g. HKUST) into /learn, then chain it with a research skill via skill bundle to output a script or article that follows the brand's voice and selling points.

Bottom line

Clear your Instagram bookmarks and YouTube Watch Later. Run them through /learn. Hand the resulting skill.md to Claude Code.

TL;DR from Ron

I'm just using it as a PRD, that way I can hand it off to a coding agent and then build the specs upon that.

Watch on YouTube

Prefer the native player? Open it on YouTube: https://www.youtube.com/watch?v=Tosd0yLSopo