32dots HEIDELBERG AI
Curriculum › Hermes course › hermes-05
Session 5 medium

Let it write its own skill

USE 0 - 25 min

Solve it once, save it forever

Here is what makes Hermes genuinely different from a chatbot: when it works out how to do something, it can write that procedure down as a reusable 'skill' — its own procedural memory. You will watch it turn a one-off solution into a permanent capability, with you approving what it saves.

  1. 1 Walk Hermes through a fiddly, multi-step task once — cleaning a messy file, formatting a report, a repeatable lookup. Get it fully right.
  2. 2 Say: Save that as a skill called so we can reuse it.
  3. 3 In chat, run /skills pending to see the staged skill, then /skills diff to read exactly what it wants to save.
  4. 4 If it looks right, /skills approve . The skill is now stored under ~/.hermes/skills/.
  5. 5 Start fresh and invoke it by name (run on this). Confirm it does the whole procedure without you re-explaining it.
✓

A task you solved once now runs from a single instruction, because the agent saved it as a skill you approved.

UNDERSTAND 25 - 50 min

Procedural memory and the compounding loop

Lesson 2 gave the agent memory of facts about you. Skills are memory of how to do things — and together they are why an assistant gets more useful the longer you use it. In the desktop app, the Skills & Tools screen has two tabs — and the difference between them is the whole point of this lesson.

a) Skills — procedures the agent has learned or that you authored. Each row is a SKILL.md it loads on demand; the toggle arms it. The skill you save in this lesson lands right here.
  • Each row is one skill (a SKILL.md): a name, what it does, and a toggle to enable it.
  • Grouped by category — the number beside each (e.g. Creative 16) is how many it holds (81 total here).
  • A skill the agent writes, or one you drop into ~/.hermes/skills/, appears here ready to switch on.
b) Toolsets — the built-in capabilities the agent can call, distinct from skills. 16 of 27 enabled here. Each bundles concrete tools — e.g. File Operations → read_file, write_file, patch, search_files.
  • A toolset groups built-in tools the agent can call — e.g. Browser Automation bundles browser_click, browser_navigate, web_search.
  • Configured = ready to use; Needs keys = add credentials first (e.g. Home Assistant, Mixture of Agents).
  • Several map straight onto this course: Memory (lesson 2), Cross-Platform Messaging (lesson 3), Cron Jobs (lesson 6).
Key concept

A skill is an on-demand instruction document (a `SKILL.md`: frontmatter plus When-to-use / Procedure / Pitfalls / Verification) stored in `~/.hermes/skills/`. The agent writes them via its skill tool after it solves something non-trivial, and — with approval on — every write is staged for your `/skills approve` or `/skills reject`. This is the compounding loop closing: facts make answers personal, skills make actions repeatable, and approval keeps you the editor of both. Skills follow the open agentskills.io standard, so a good one is shareable, not locked to your machine.

  1. ?What is the difference between memory (lesson 2) and a skill — and why do you need both?
  2. ?Open an approved SKILL.md. Which of its sections (when-to-use, procedure, pitfalls, verification) matters most for it being reused correctly?
  3. ?Why route skill writes through an approval gate instead of letting the agent save whatever it likes?
  4. ?What makes a skill worth saving versus a one-off you should just let the agent redo from scratch?
BUILD 50 - 75 min

Author or sharpen a skill by hand

You do not have to wait for the agent to propose skills — you can write or edit them yourself. You will take a starter and make it yours.

Your task

Either refine the skill the agent wrote, or hand-author one from a starter (the course Downloads include tidy-csv, daily-brief, inbox-triage). Then prove it gets used.

  1. 1 Download a starter SKILL.md from the course Downloads section, or open the one the agent saved under ~/.hermes/skills/.
  2. 2 Edit it: tighten the Procedure to your workflow, and make the Verification section concrete (how do you know it worked?).
  3. 3 Place it in the right ~/.hermes/skills/// folder.
  4. 4 Trigger it by name and confirm the agent follows your edited procedure.
  5. 5 Break it on purpose (give it bad input) and check the Pitfalls/Verification section actually catches the problem.
Deliverable

Your edited SKILL.md, a run showing the agent used it, and one sentence on what your Verification section now guarantees.

Self-check · tick before you mark done

✎Every skill you approve is a capability your assistant keeps. After a year of this, what does your `~/.hermes/skills/` folder say about how you work — and would you trust a colleague to inherit it?