Hermes Agent Skills Advanced Guide: от SKILL.md до GEPA self-evolution

Getting started решает install; advanced — как сделать agent stronger over time. Четыре measurable advantages:

• On-demand loading: zero tokens до activation — Progressive Disclosure держит cost под контролем.
• Open standard: agentskills.io — cross-platform reuse в Hermes, Claude Code, Cursor.
• Composability: Skill Bundles — один slash command загружает entire workflow.
• Evolvability: GEPA парсит execution traces и патчит SKILL.md без fine-tuning model weights.

Четыре типичных pain point у power users:

• Token bloat: все SOP в system prompt — тысячи tokens каждую сессию впустую.
• False activation: размытый description — LLM подгружает skill в irrelevant context.
• Workflow fragmentation: PR review, TDD, deploy по одному через /skill-name.
• No team sharing: skills в personal dirs — высокий onboarding cost при смене машины.

Memory hook: Prompt = sticky note (single session); Memory = notebook (permanent notes at hand); Skill = SOP manual (step-by-step, open when needed).

Skills и MCP complementary: MCP даёт tool interfaces (DB access и т.д.); Skill учит agent как правильно использовать tool для migration и подобных задач.

Все Hermes Skills следуют open standard agentskills.io. Базовый frontmatter:

---
name: my-skill
description: |
  Use when the user needs to [...].
version: 1.0.0
license: MIT
compatibility: Requires git, docker
allowed-tools: Bash(git:*) Read
metadata:
  hermes:
    tags: [devops, automation]
    category: software-development
    related_skills: [github-pr-workflow]
    requires_toolsets: [terminal]
    fallback_for_toolsets: [web]
---

# My Skill Title
## Overview / When to Use / Procedure / Common Pitfalls / Verification Checklist

Recommended directory layout:

my-category/my-skill/
├── SKILL.md              # Core steps, рекомендуется ≤500 строк
├── references/           # API refs, on-demand load
├── templates/            # Reusable templates
└── scripts/              # Executable scripts для agent

Writing rules: description — вся Level 0 info; «when to use» важнее «what it is»; SKILL.md: Overview, When to Use, Procedure, Common Pitfalls, Verification Checklist. Validate: skills-ref validate ./my-skill.

Skill Bundles (Hermes 2026) — lightweight YAML упаковывает несколько skills в один slash command. При /bundle-name все listed skills загружаются simultaneously.

Path: ~/.hermes/skill-bundles/<slug>.yaml

name: backend-dev
description: Full backend feature workflow — code review, TDD, and PR management.
skills:
  - github-code-review
  - test-driven-development
  - github-pr-workflow
instruction: |
  Always write failing tests first before implementation.
  Never push directly to main.

Advanced scenarios: research employee workflow — arxiv, deep-research, plan, excalidraw; MLOps deploy — vllm, llama-cpp, github-pr-workflow, systematic-debugging.

Priority rules: Bundle beats same-named Skill; missing skills skipped с warning; Bundles не трогают system prompt — Prompt Cache остаётся valid.

hermes bundles create backend-dev \
  --skills github-code-review,test-driven-development,github-pr-workflow \
  --instruction "Always write failing tests first"

В metadata.hermes — четыре activation rules; skills show/hide по tool availability:

Classic case: DuckDuckGo search skill с fallback_for_tools: [web_search] — при FIRECRAWL_KEY / BRAVE_SEARCH_KEY активируется paid web_search, DuckDuckGo скрывается (token savings); при API outage fallback resurfaces.

Platform awareness: telegram-notify: requires_toolsets: [messaging], platforms: [telegram, discord]; через TUI hermes skills — independent toggles для CLI, Telegram, Discord.

Official install channels:

hermes skills install official/research/arxiv
hermes skills install https://example.com/SKILL.md --name my-skill
hermes skills install github:openai/skills/k8s
hermes skills tap add github:my-org/my-skills

agentskills.io — skills portable между Hermes, Claude Code, Cursor, OpenCode; assets не locked к одной platform.

GitHub repo как Tap — team/community subscribes к вашему skill set. Recommended structure:

my-skills-tap/
├── skills.sh.json          # Category config (optional)
├── mlops/vllm-deploy/SKILL.md
├── research/paper-summarizer/SKILL.md
└── README.md

1. Plan categories: MLOps, Research и т.д.; skills.sh.json для Hub grouping.
2. Write SKILL.md: per-skill directory; skills-ref validate.
3. Push GitHub: public или private (token для private).
4. Team subscribe: hermes skills tap add github:your-org/your-skills-tap.
5. Regular updates: hermes skills tap update.
6. Version control: ~/.hermes/skills/ в Git; cross-device sync: git pull && hermes skills reset.

hermes skills tap add github:your-org/private-skills --token $GH_TOKEN
hermes skills tap list
hermes skills tap update

GEPA (Genetic-Pareto Prompt Evolution) — ICLR 2026 Oral, integrated в hermes-agent-self-evolution. No weight fine-tuning: parse execution traces, generate variants, multi-objective Pareto optimization skill text. Cost per run: ~$2–10 (pure API, no GPU).

Five-phase pipeline: (1) trace collection (SQLite); (2) reflective failure analysis (actionable side info); (3) targeted mutation (10–20 SKILL.md variants); (4) multi-objective Pareto eval (success rate × token efficiency × speed); (5) human PR review перед production.

export HERMES_AGENT_PATH=~/.hermes
python -m evolution.skills.evolve_skill \
    --skill github-code-review \
    --iterations 10 \
    --eval-source sessiondb

Four safety guardrails: full test suite 100% pass; Skills ≤ 15KB, tool descriptions ≤ 500 chars; Prompt Cache compatible; semantic preservation check.

agentskills.io compliant — feed Claude Code или Gemini CLI traces: --eval-source mixed --trace-dirs ~/.claude/traces,~/.hermes/sessions.

Plugins pack skills как namespace plugin:skill: absent из default skills_list, activate только на explicit invoke; sibling skills внутри plugin cross-reference. skill_view("superpowers:writing-plans") показывает sibling skills того же plugin.

description drives activation precision: avoid «Helps with code»; explicit trigger conditions и exclusion scenarios.

Pitfalls — quality differentiator: concrete failure modes, root cause, fix steps (CSS selector fragility, GitHub API rate limit, large diff token overflow).

Scripting: reference executable scripts в scripts/; on failure fallback к references/manual-extract.md.

Agent может dynamic patch/create через skill_manage; в config.yaml: skills.agent_writes_require_approval: true для human approval gate.

Build blog-workflow Bundle — one-shot load SEO research, outline, code validation, bilingual check, publish:

name: blog-workflow
description: Full tech blog writing workflow.
skills:
  - seo-keyword-research
  - outline-generator
  - code-example-validator
  - bilingual-checker
  - publish-to-platform
instruction: |
  Always research SEO keywords before writing.
  Ensure all code examples are tested and runnable.
  Generate both Chinese and English title options.

Custom seo-keyword-research: set requires_toolsets: [web]; flow: identify topic → RU long-tail («как использовать X», «X tutorial») → EN long-tail («X tutorial», «X vs Y») → cross-ref Habr/Dev.to/HN → output 3–5 head terms + 10–15 long-tail matrix. Search behavior и terminology validation per target platform.

• Skills vs MCP? Skills = procedural knowledge docs; MCP = tool interface — complementary.
• Skill patched но old version active? Current session unaffected — /reset или install с --now (invalidates Prompt Cache).
• GEPA safe? Four guardrails + human PR review — still review every diff.
• Claude Code reuse? Copy SKILL.md to ~/.claude/skills/ или ai-agent-skills multi-platform install.
• Non-EN content и tokens? Non-English description может degrade LLM matching — EN или bilingual recommended.

Further reading: official docs, GEPA algorithm, DSPy framework.

• GitHub Stars: Hermes Agent early 2026 — 160K+ за два месяца.
• Level 0 tokens: all name+description ~3K tokens/session.
• GEPA cost/run: ~$2–10, pure API, no GPU.
• GEPA size limit: Skills ≤ 15KB, tool descriptions ≤ 500 chars.
• Community scale: kevinnft/ai-agent-skills — 191 skills, 28 categories; hermeshub 166 stars с security scan.

Hermes Agent + GEPA evolution pipeline требует 24/7 online, low-latency macOS host. Raspberry Pi (insufficient RAM), oversubscribed shared VPS (long-connection drops), flaky home broadband — всё это режет trace collection и Gateway uptime.

Для stable production — Hermes Gateway, continuous sessiondb trace collection, GEPA iterations — JEXCLOUD multi-region bare-metal Mac: dedicated Apple Silicon, 24/7, monthly elastic scaling, 120-second node delivery. Config и pricing: JEXCLOUD pricing.