← all posts site ↗

how hermes skills and petdex work

skills in hermes are essentially markdown files you drop into the skills directory that the agent reads at startup and appends to its system prompt. they're structured instruction bundles — a yaml or md file with a name, optional description, and a block of text that teaches the agent how to handle a specific domain. you activate them with skill_view() and the agent will remember the context within that session. it's not vector search or RAG, it's literal prompt injection with file-based persistence. the elegance is that anyone can write a skill without touching the agent's core code.

the directory layout is dead simple. every file in ~/.hermes/skills/ ending in .yaml or .md gets loaded. the filename becomes the skill identifier, so naming conventions matter. a skill for home assistant might be home-assistant.yaml and the agent will know to call it by that name. skills can reference each other — you can have a base skill that defines common tools and domain skills that extend it. the memory aspect is handled by a separate .hermes/memories/ directory where the agent can write persistent context that survives across sessions. so you could have a skill called "user-tools" that loads your common toolchain and a memory file that tracks your current project state, and the agent will merge both into its context window.

petdex is the terminal pet system built into hermes. you invoke it with skill petdex, pick a creature from the built-in catalog — void dragon, basilisk, slime, a few others — and it spawns an ascii-animated pet that lives in your terminal. the pets have idle animations, mood states, and some of them interact with the terminal cursor. the rendering is pure unicode frame-based animation with no external dependencies, so it works in any terminal emulator. i've been running the basilisk for a few days because its idle loop doesn't stutter like the dragon's. it doesn't do anything useful, but having a little animated creature sitting in the corner of your terminal while you work is surprisingly pleasant.

the real win is combining both features. you can stack skills to build a domain-aware agent — a retrieve skill for web searches, a code skill for your language preferences, a travel skill for itinerary planning — and have a pet sitting alongside as a persistent companion. the skills make the agent smarter about your specific workflow, the pet makes the experience feel less like a command-line utility and more like a workspace. it's the kind of small design detail that compounds into a tool you actually enjoy using.