Skills that get stronger with use
Hermes auto-creates Skills from repeated tasks and improves them based on your feedback. This is what makes it different from every other agent.
Step 12 taught you how to write Skills by hand and install them from the hub. But the most powerful part of Hermes's Skill system is what you don't have to do — it creates Skills on its own and improves them over time.
Skills that grow on their own
Have you noticed yourself asking Hermes to do the same kind of thing over and over? Maybe "convert this Markdown to clean HTML," or "summarize what changed in this PR," or "clean up this CSV and import it into the database."
The first two or three times, Hermes figures it out from scratch. But around the third or fourth time, it does something in the background: it distills the solution into a Skill file and saves it to ~/.hermes/skills/.
You didn't teach it to do this. It decided on its own that this task was worth remembering how to do.
See it happen
Let's trigger an automatic Skill creation. Give Hermes the same kind of task three or four times in a row:
First time:
Convert this Markdown to clean HTML. Keep bold and code block styling. Remove all id attributes from headings.
# Test heading
This is some **bold** text with some `inline code`.
Second time, different content but the same request:
Convert another one:
## Another heading
A list:
- First item
- Second item with `code`
Third and fourth time, keep the same pattern.
After you're done, check your Skills directory:
ls ~/.hermes/skills/
You'll likely see a new folder — something like markdown-to-html. Open the SKILL.md inside — it records the input format, conversion rules, and your preferences (like removing heading id attributes).
Checkpoint — A new Skill appeared in
~/.hermes/skills/that you didn't create manually. That's the Learning Loop at work.
Skills improve themselves
Auto-creation is just the start. The more interesting part is improvement.
Next time you ask Hermes to do the same conversion but you're not happy with the result — say:
Next time you do this conversion, add a blank line between paragraphs and use a dark background for code blocks.
Hermes doesn't just apply this feedback once. It goes back and edits the Skill file, writing your feedback into the rules. Next time this Skill runs, the new rules are the default.
Verify it:
cat ~/.hermes/skills/markdown-to-html/SKILL.md
Your feedback is now a rule in the Skill.
Why this matters
Traditional AI tools are stateless. You re-explain your preferences every single time — "I like this format," "don't use that word," "follow this convention."
Hermes's Skill system is stateful and grows:
- Time 1: Starts from scratch, result might be rough
- Time 3: Auto-creates a Skill, remembers the basic workflow
- Time 10: After several rounds of feedback, the Skill is highly precise
- Time 20: You barely need to say anything extra — the result is just right
This is the "flywheel effect" — the more you use it, the stronger every step gets, and they accelerate each other.
How to give good feedback
Skill improvement quality depends on your feedback quality.
Good feedback: "Don't use inline styles on code blocks, use class names instead." — Specific, actionable.
Bad feedback: "This doesn't feel right." — Hermes doesn't know what to change.
The more specific you are, the more accurately the Skill evolves.
Manual override
Skill files are just Markdown. You can edit them by hand anytime. If Hermes's auto-improvement drifts in the wrong direction, open the file and fix it. Hermes treats your manual edits as the highest-priority signal and bases future improvements on your corrected version.
Next: deploy Hermes to a server so it's online around the clock.