Hermes Agent is an AI agent runtime built by NousResearch — it has its own memory system, tool routing, and cross-session learning capabilities. The original design runs it as an independent process with its own runtime.
The author of hermes-CCC made an interesting observation: Claude Code is already an agent runtime. There’s no need to run a separate process.
So they broke down every Hermes Agent capability into 46 Skills and loaded them into Claude Code’s native Skill system. No OAuth, no extra accounts, no external processes — restart Claude Code and everything is available.
Procedures-as-prompt
The core design philosophy here is worth examining on its own.
Each Skill is a SKILL.md file written in structured natural language that specifies: trigger conditions, execution procedures, output format, failure modes, and recovery actions. Claude Code reads the description field to decide whether to trigger, and reads the ## Checklist to step through execution.
No code. No function calls. Pure markdown instruction scripts driving Claude’s behavior.
This isn’t laziness — it’s a deliberate design choice. Traditional agent frameworks require Python, dependency management, API handling, and runtime maintenance. Procedures-as-prompt moves the “behavioral specification” from the code layer up to the text layer. Anyone can read it, anyone can modify it, and the AI itself understands what it’s doing.
Manual vs Automatic Triggering
There’s an important mental model to establish when working with this kind of Skill system.
Manual triggering: You type /hermes-route, /systematic-debugging, or another slash command directly. Guaranteed execution, no ambiguity.
Automatic triggering: Each SKILL.md has a description field in its frontmatter. Claude reads every Skill’s description on each conversation turn and decides whether the current context matches — if it does, the Skill triggers automatically. This is an expectation, not a guarantee.
The more precisely the description is written, the more reliable the automatic triggering. But LLM judgment is inherently non-deterministic.
Manual triggering is a guarantee. Automatic triggering is an expectation. Keep this in mind or you’ll develop incorrect expectations about how “automated” the system actually is.
The Most Notable Skills
With 46 Skills, I can’t cover them all, but a few stand out for their design thinking:
/hermes-route: Task Router
Classifies each task before execution, determining how to approach it. Outputs a routing decision block including: task class (lightweight / standard / deep), execution mode, files to read first, whether parallelism is possible, and required reasoning depth.
The key insight is “evaluating the cost of getting the first step wrong” — for a deep task, a wrong first move costs far more to recover from than for a lightweight one.
/hermes-memory: Layered Cross-Session Memory
Manages persistent memory across sessions with four operations: Prefetch (pull relevant memories before work), Sync (merge new knowledge after work), Nudge (suggest things worth saving without auto-saving), Compress (clean up duplicates and delete stale memories).
It forces a distinction between “persistent knowledge” (architecture decisions, tool pitfalls, preferences) and “temporary state” (branch names, single-task progress) — a distinction most people miss when managing memory manually.
/hermes-compress: Context Compression into Six Buckets
When the context window is nearly full, it extracts the current conversation into a structured summary stored in memory, freeing up context space. Six buckets: decisions, artifacts_created, problems_solved, facts_learned, open_issues, next_steps.
This is more organized than manually running /compact — instead of condensing conversation into prose, it forces structure.
The Real Value of This Repo
hermes-CCC is worth studying not just because you can use 46 Skills out of the box.
More importantly, this repo demonstrates something: Skill systems can be architecturally designed. Task routing, memory layering, context management, sub-agent coordination — all the familiar patterns from agent frameworks can be expressed in SKILL.md language.
If you’re designing your own Skill system, hermes-CCC is currently the most complete reference implementation of “how to architect a Skill system” that the community has produced.
GitHub: https://github.com/AlexAI-MCP/hermes-CCC (run /security-scan before installing)