Если Hermes Agent у вас уже в проде и хочется не «просто юзать», а крутить skills, которые реально эволюционируют, копать надо не в bigger model, а в Skills subsystem Nous Research: open standard agentskills.io, Progressive Disclosure (L0/L1/L2), Skill Bundles, conditional activation и GEPA + DSPy self-evolution. Гайд: полный SKILL.md spec, Tap publish pipeline, OSS ecosystem, Plugin namespace, skill_manage, blog workflow bundle — с comparison tables, six deploy steps и KVMNODE cloud Mac host. Cross-read: install guide, three-layer memory.
01

Зачем в 2026 отдельно разбирать Hermes Skills: Prompt vs Memory vs Skills

В начале 2026 Hermes Agent за два месяца пробил 160k+ GitHub Stars. Философия "the agent that grows with you" реализована через Skills — standardized, evolvable, cross-session procedural memory по agentskills.io, portable между Hermes, Claude Code и Cursor.

DimensionPromptMemorySkills
PersistenceCurrent chatCross-session permanentCross-session permanent
Load timingAlways in contextAuto-inject per sessionOn demand
Token costEvery turnSmall, stableZero until activated
Content typeAny intentUser prefs/factsProcedural steps
ShareabilityHardPrivateCommunity Tap

Mnemonic: Prompt = sticky note · Memory = notebook · Skill = SOP manual.

01

Skill as prompt dump: full inject каждую session — token explosion, Progressive Disclosure dead.

02

description слишком generic: LLM грузит skill в irrelevant context.

03

Ignore conditional activation: free/paid search coexist — два skill'а жрут Level 0 slots.

04

Skill bloated: >15KB — GEPA guardrails reject, split в references/.

05

Laptop host: SKILL.md patched без /reset, Gateway sleep, team Tap out of sync.

02

SKILL.md deep dive: agentskills.io spec и Progressive Disclosure L0–L2

Все Hermes Skills следуют agentskills.io. Frontmatter required: name (lowercase+hyphen, ≤64) и description (≤1024, best practice «Use when...»); recommended: version, license, compatibility, experimental allowed-tools.

yaml
---
name: my-skill
description: |
  Use when the user needs to [...].
  Handles [...] and [...].
version: 1.0.0
license: MIT
compatibility: Requires git, docker
metadata:
  hermes:
    tags: [devops, automation]
    requires_toolsets: [terminal]
    fallback_for_toolsets: [web]
---
# My Skill Title
## Overview / When to Use / Procedure / Common Pitfalls / Verification Checklist

Modular layout (~/.hermes/skills/my-category/my-skill/): SKILL.md core (≤~500 lines); references/ API docs; templates/ reusable templates; scripts/ agent-executable scripts.

LevelPayloadTriggerToken cost
Level 0name + descriptionSession start, all skills~3K total
Level 1Full SKILL.md body/skill-name or LLM matchFile size dependent
Level 2references/, scripts/Runtime LLM decisionPer file on demand

Writing rule: description — entire Level 0 signal; «when to use» beats «what it is». Validate: skills-ref validate ./my-skill.

Size control: <500 lines → all in SKILL.md; 500–1000 → offload to references/; >1000 → split or two skills; >15KB → GEPA hard limit, mandatory split.

03

Skill Bundles и conditional activation: one slash, full workflow

Skill Bundles (~/.hermes/skill-bundles/<slug>.yaml) — Hermes 2026 feature: /bundle-name loads listed skills simultaneously. Bundle beats same-name Skill; missing skills skipped silently; no system prompt mutation — token-friendly.

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.

CLI: hermes bundles create backend-dev --skills github-code-review,test-driven-development --instruction "Always write failing tests first". Research stack: arxiv, deep-research, plan, excalidraw; MLOps: vllm, llama-cpp, systematic-debugging.

Conditional activation via metadata.hermes:

FieldLogic
requires_toolsetsHide if toolset missing
requires_toolsHide if tool missing
fallback_for_toolsetsHide if toolset present (fallback path)
fallback_for_toolsHide if tool present

Classic: duckduckgo-search + fallback_for_tools: [web_search] — with FIRECRAWL_KEY/BRAVE_SEARCH_KEY DuckDuckGo drops from L0; API down → fallback resurfaces. Platform-aware: requires_toolsets: [messaging], platforms: [telegram, discord]; TUI hermes skills per-channel toggles.

04

Skills Hub, Tap publish и Plugin skills: install → team share

Official install channels:

shell
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
hermes skills tap update
hermes skills tap list
RepoDescriptionHighlight
ChuckSRQ/awesome-hermes-skillsProduction-grade picksDeep Research, MLOps, 23 skills + Copilot
amanning3390/hermeshubCommunity registrySecurity scan, API, marketplace
kevinnft/ai-agent-skills191 skills, 28 categoriesHermes/Claude/Cursor cross-platform
NousResearch/hermes-agentOfficial repoBuilt-in skills, spec source of truth

Team Tap: GitHub repo with skills.sh.json + category SKILL.md files. Team: hermes skills tap add github:your-org/your-skills-tap; private repo → --token $GH_TOKEN. Git-track ~/.hermes/skills/; post-sync hermes skills reset rebuilds builtins.

Plugin skills namespace plugin:skill (e.g. skill_view("superpowers:writing-plans")), opt-in only, absent from default skills_list. Declare skills path in plugin.yaml.

Deploy

Six steps: solo Skill → team Tap + cloud Mac host

01

Verify env: hermes doctor — Gateway, toolsets, ~/.hermes/ paths.

02

First SKILL.md: agentskills.io template — description, Procedure, Pitfalls, Verification Checklist.

03

Create Bundle: hermes bundles create blog-workflow --skills seo-keyword-research,outline-generator --instruction "Research SEO first".

04

Conditional activation: free/paid search, messaging platforms — separate fallback/requires rules.

05

Publish Tap: push GitHub, team hermes skills tap add; agent_writes_require_approval: true approval gate.

06

Cloud Mac 7×24: KVMNODE dedicated Mac Mini M4 — no laptop sleep, stable GEPA trace accumulation.

05

GEPA + DSPy self-evolution, advanced authoring и hard numbers

GEPA (Genetic-Pareto Prompt Evolution) — ICLR 2026 Oral, shipped in hermes-agent-self-evolution. Core idea: no weight fine-tuning — execution trace analysis, variant generation, multi-objective Pareto optimization on SKILL.md text. Cost: $2–10/run, pure API, GPU not required.

Five-phase pipeline: ① SQLite trace collection → ② reflective failure analysis → ③ targeted mutations (10–20 SKILL.md variants) → ④ Pareto eval (success × token efficiency × speed) → ⑤ human PR review before ship.

shell
git clone https://github.com/NousResearch/hermes-agent-self-evolution
export HERMES_AGENT_PATH=~/.hermes
python -m evolution.skills.evolve_skill --skill github-code-review --iterations 10 --eval-source synthetic
python -m evolution.skills.evolve_skill --skill github-code-review --iterations 10 --eval-source sessiondb
python -m evolution.skills.evolve_skill --skill github-code-review --eval-source mixed --trace-dirs ~/.claude/traces,~/.hermes/sessions

Four guardrails:pytest tests/ -q 100% green; ② Skills ≤15KB, tool descriptions ≤500 chars; ③ Prompt Cache intact; ④ semantic preservation check. Roadmap: Phase 1 Skill files (✅ DSPy+GEPA) → Phase 2 tool descriptions → Phase 3 system prompt → Phase 4 tool implementation (Darwinian Evolver) → Phase 5 fully automated continuous improvement.

Advanced authoring: Pitfalls with concrete failure modes + root cause + fix; Procedure references scripts/ with fallback to references/manual-extract.md; runtime maintenance via skill_manage(action='patch'|'create').

Blog workflow bundle: blog-workflow packs seo-keyword-research, outline-generator, code-example-validator, bilingual-checker, publish-to-platform — instruction: SEO research first, validate runnable code, bilingual title options.

A

GitHub Star velocity: Hermes Agent early 2026 160k+ in two months — fastest-growing OSS AI agent project.

B

Level 0 token budget: all skills name+description ~3K tokens — first cost gate.

C

GEPA per-run cost: $2–10 pure API; sessiondb traces beat synthetic eval.

D

Cross-platform library: kevinnft/ai-agent-skills — 191 skills, 28 categories, Hermes/Claude/Cursor.

E

SEO long-tail: skip saturated «Hermes tutorial» — target GEPA prompt evolution, agentskills.io SKILL.md, hermes skills tap publish.

Note: keep description in English or bilingual — LLM matching точнее; body может быть на русском без существенного token penalty (~1–1.5 token/char).

Alternative matrix: MacBook + GEPA — lid closed = trace gap, SQLite bloat → swap thrash; Linux VPS — no official macOS path, no Metal local inference; Skill Tap + Xcode CI same box — reboot roulette. Production с 7×24 Gateway, stable traces и continuous GEPA evolution → KVMNODE dedicated Mac Mini M4/M4 Pro monthly rent: launchd daemon, six regions, day/week/month flex. Цены, Оформить заказ, Центр помощи; closes loop с OpenRouter CLI Hermes #1 ranking.