Source: raw/How_To_Build_A_Hermes_Agent_Skill_In_15_Minutes.md (auto-caption transcript)

Creator: Sharbel (YouTube @sharbelxyz) · URL: https://www.youtube.com/watch?v=YgCUJgqoIcE · Duration: 15:50 · Uploaded: 2026-08-30 · No sponsor named. He promises the skill-audit skill from the video in his pinned comment and on GitHub.

Sharbel audited 55 popular skills from GitHub and argues that a skill you write yourself beats most downloads. He walks through four decisions that make a skill work, using a skill-audit skill as the example. His audit numbers are his own and unverified, but the rules match Anthropic’s published skill guidance in this wiki.

Key Takeaways

  • His audit of the top 55 GitHub skills:
    • 27 of 55 (49%) describe what the skill is but never say when to use it.
    • Median size: 1,665 tokens.
    • Only 1 of 55 (2%) uses a references folder.
  • Source material first. A skill is worth building when it holds something the agent cannot work out: your method, your corrections, your mess. Either do the job once by hand and log every correction, or ask the agent to read the corrections out of past work: “Go back through the last three times we did this … and write down every single thing I corrected you on. Do not summarize it.” He calls this step “80% of what makes a good skill.”
  • The description is the trigger. Hermes loads only each skill’s name and description, then picks from those. Say what it does and when to use it. His fix for his own “Nova” skill: add “Use the skill whenever we’re working on a YouTube video.” Models under-trigger, so write the description more forcefully than feels comfortable.
  • Size costs you every time it fires. His Nova skill is 19,000 tokens against published guidance of under 500 lines or about 5,000 tokens. He says it sent 6.19 million extra input tokens over the period he measured, counted against his Codex plan’s 5-hour limit rather than billed.
  • Steps in the file, everything else in references. A skill holds the procedure and the material the procedure needs. Material used only sometimes goes in a references/ folder with a pointer. Prompt: “Read the skill and tell me which parts only get used some of the time. Move those into a references folder and leave a pointer.”
  • One right way means a script. If a step allows many right answers, write instructions. If it has exactly one (arithmetic, counting), write code. He asked Hermes to count tokens three times from instructions and got three answers. Scripts are not loaded into context, so they are cheaper as well as reliable. Say explicitly whether the agent should run the file or read it.
  • Mark consequential skills user-only. A skill that sends, deploys or spends should be invocable only by you.
  • The deletion test. Delete a paragraph, rerun, and if the output did not change the paragraph was costing you for nothing. Hunt for the same rule in two places, material that piled up, and text that sounds like instruction but changes nothing.
  • Grow it from failures. After each imperfect run, ask whether the fix is one-time or permanent; permanent fixes go in the file. Prompt: “look at the back and forth we just had, what should I add to the skill so we never have this conversation again.”
  • Supply-chain warning. A skill can ship code that runs. Researching the video, he found a well-known memory tool (27,000+ stars, MIT) whose obvious install command pulled a different package because someone had squatted the name on the public registry. Copy the install command from the repository itself, then ask your agent to scan the link for anything malicious.

Try It

  1. Ask Hermes: “What are the things that I ask you for repeatedly that you think we should turn into a skill?”
  2. Open one skill you installed last week and read the description aloud. If you cannot tell when it fires, add a “Use when …” sentence.
  3. Run the references prompt above on your largest skill and compare /context all before and after.
  4. Move any counting or calculation step into a script and tell the skill to run it, not read it.

Open Questions

  • He does not name the squatted memory tool or the registry.
  • The source of the “under 500 lines or roughly 5,000 tokens” guidance is not stated; it matches Anthropic’s skill-authoring advice
  • The audit-skill repo had not been checked at ingest.