Hermes Skill System
blueif16/hermes-skill-system
Framework disciplinado para evoluir skills/agents com commits rastreáveis e generalizáveis
Ver no GitHubSkill para Claude Code que implementa um sistema disciplinado de evolução contínua de skills e agents. Organiza mudanças em commits atômicos e rastreáveis, mantém um mapa centralizado do sistema, e aplica três procedimentos (INIT/OPERATE/CONSOLIDATE) para garantir que cada modificação seja generalizável e revertível.
Instalação
npx convez-ai blueif16/hermes-skill-system1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Estruturar e mapear sistemas complexos de skills e workflows
- Rastrear e revisar mudanças em agents com auditoria completa
- Diagnosticar falhas em sistemas orquestrados de forma sistemática
- Consolidar duplicações e manter clareza nas responsabilidades de cada nó
Veredito

Ferramenta sólida para times que precisam governar sistemas de agents em escala. Exige disciplina rigorosa, mas paga dividendos em rastreabilidade e confiabilidade — ideal para quem constrói workflows complexos que evoluem continuamente.
README
A tradução para o português deste README ainda está na fila.
hermes-skill-system
A Claude Code skill for setting up and continuously evolving a skill/agent system the disciplined way. It does not grade the artifacts a system produces — the human is the eye. Its whole job is to make every change to the system land in the right place, generalize across all future runs, and stay trackable as one atomic, revertible commit.
Informed by Nous Research's "Hermes Agent" research (2026-06-08).
The model — one orchestrated system, mapped completely
A skill system is not a pile of skills; it's a workflow that orchestrates skills. A central
orchestrator (e.g. .claude/workflows/lesson-build.js) calls each node, and each node reads part of
the skill system. So the map records everything — skills, the orchestrator/workflow file(s),
nodes, subagents, kits, the capability registry, the governing CLAUDE.md, and which node relies on
what. The map is also the diagnostic surface: it records what is responsible for what, where each
run's real logs + artifacts live, and a running diagnostics log of what past fixes concluded. Read
together with the live problem, it points at the top candidate to fix. It gets more certain with
every run — a stale map is the one real failure mode of this method.
The three procedures
| Command | File | When |
|---|---|---|
| INIT — map the system | references/init.md | Once per repo, and whenever the system's shape changes (new node, new skill, re-wired orchestrator). Also injects the ambient stewardship hook into the project's CLAUDE.md/AGENTS.md and installs the skillsys(...) commit convention. |
| OPERATE — the daily loop | references/operate.md | A flaw was spotted, a node misbehaved, or a finding recurred. capture → route → edit → verify → commit. |
| CONSOLIDATE — on demand | references/consolidate.md | Merge duplicated/conflicting guidance into the canonical owner; keep skills inside their disclosure budget. |
Install
This is a Claude Code skill. Make it discoverable by placing (or symlinking) the folder under your skills directory:
ln -s "$(pwd)" ~/.claude/skills/hermes-skill-system
Claude Code will surface it when you spot a flaw in a skill or workflow, when a finding recurs across runs, when bootstrapping the convention into a repo, or when reviewing/consolidating recent edits.
The laws (the point)
- Diagnose from real evidence, not guesses. Read the user-surface problem, the skill composition, the workflow composition, and the actual runtime logs + artifacts together.
- Generalize or don't ship. Every edit must be reliable across all future runs. A fix that only helps the case in front of you is a bug, not a lesson learned.
- Verify by intent, never by hard-coded test. Verification is an instruction the next session carries out — never a prediction of the exact next output. The human is the eye for visuals.
- Prefer fixing the chain over fixing one skill. Improve a wave by editing its SKILL; improve the chain by editing the workflow/orchestrator.
- One canonical home, no duplication. Refine the rule in place; don't restate it.
- Smallest durable edit. Patch a section > add a
references/file > create a new skill. - Every change is one atomic, revertible commit —
skillsys(<owner>): <rule>with why/lesson/verify in the body. Record it in the map's diagnostics log. - Immediate, on demand — no timers. Fix flaws the moment they're spotted.
- Concise. A lean system is a usable one.
Contents
| Path | What it is |
|---|---|
SKILL.md | The skill — the model, the laws, the three procedures. |
references/init.md | INIT: build/refresh <repo>/.agents/skill-system-map.md and inject the ambient hook. |
references/operate.md | OPERATE: the capture → route → edit → verify → commit loop. |
references/consolidate.md | CONSOLIDATE: merge drift into the canonical owner. |
scripts/review-edits.sh | Lists every skillsys(...) commit in a span, grouped, with its why-line — the periodic review. |
The per-repo map (<repo>/.agents/skill-system-map.md) is data that lives in each consuming
repo; this repo is the portable method.
License
MIT