菜单
Orchestrate parallel AI agent workstreams — multi-agent deep research (3 parallel experts), research orchestration Trigger phrases: orchestration/multi, agent, orchestration. Converted from Hermes-Skill (Yuno's mastery library).
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Funny animal podcast creation assistant. Given a topic, automatically generates a comedic podcast video hosted by two animal "hosts." Full pipeline: character design, comedy script, character portraits, TTS voice + Seedance 2.0 video generation, subtitle/pop-text/sound-effect post-production. Trigger words: animal podcast, funny podcast, pet podcast, animal talk show, pet talk show.
Professional anime/2D art style generation skill. Covers 14 sub-styles (modern Japanese anime/moe, retro Japanese cel-shading, Japanese shonen, Japanese shojo, Ghibli, Makoto Shinkai, Chinese xianxia/ink wash, modern Chinese anime, Chinese 3D fantasy, Korean webtoon, Korean impasto, Western cartoon, chibi/moe, 2D cyberpunk) + 5 anti-failure iron laws + cross-style shared rules (character lock / facial proportion spec / stroke consistency / universal negative). Core capabilities: precise style targeting, consistent character identity, cross-style conversion. Trigger: "anime", "2D art", "manga", "illustration", "Japanese anime", "Chinese anime", "Korean webtoon", "webtoon", "Western cartoon", "ghibli", "shinkai", "ufotable", "trigger style", "cel-shading", "impasto", "chibi", "moe", "catgirl", "Chinese 3D fantasy", "xianxia", "ink wash", "hanfu character", "cyberpunk anime", "draw an anime character", "make an anime avatar", "anime character", "anime style". NOT for: photorealistic (use image agent default) / s
Specialized in anime/2D/character stylization for image generation and conversion. Covers Japanese, Chinese, Korean, and Western art style families. Uses provenance analysis to trace reference images' style DNA, performs a 10-dimension analysis → 3-dimension collapse to precisely lock the style's essence, then matches the optimal tool and prompt approach for generation. Trigger on: "anime-ify", "2D style", "convert to anime", "cel-shading", "ghibli style", "Korean watercolor", "fantasy 3D", "chibi", "Japanese anime style", "style conversion", "manga style", "character illustration", "anime style", "webtoon style", or any request involving converting content into a specific anime/2D art style. Key distinction: User requests generation or conversion to a specific anime/2D art style. Do NOT trigger for: photorealistic photography style, pure logo design, general image editing (crop/background removal etc.).
Audiobook creation assistant. Converts book text into multi-character narrated audio, supporting audiobook production, multi-character voiceover, novel narration, TTS voiceover, and read-aloud scenarios. Automatically identifies dialogue and narration, assigns a distinct voice to each character, intelligently adds pause markers, and generates natural, fluent audiobook audio. Trigger phrases: audiobook, read aloud, TTS book, multi-character voiceover, novel narration, book narration, voice acting, narration, 有声书, 朗读, 读书, 多角色配音, 小说朗读, 读书配音. Supports chapter-level generation — user confirms the first chapter, then remaining chapters continue sequentially.
Beat-sync video editing skill. Input music (URL / local file / AI-generated) + video or image assets, automatically performs energy-tension analysis → smart trimming → beat detection → beat-synced timeline generation → asset matching → ffmpeg concatenation, outputting a beat-synced video perfectly aligned to the music. Supports image slideshows, video clips, and mixed assets as input. Supports every-N-beat asset switching, automatic intro/chorus segmentation, user-annotated keypoints, and other beat-sync modes. Trigger words include: music beat sync, beat-sync editing, beat sync, beat-sync video, beat-synced editing, beat-sync video, music rhythm editing, rhythm beat sync, beat detection, edit to the beat, auto beat sync, music sync edit.
Video project push tool. Programmatically creates editing projects via Python scripts, with one-click push to JianyingPro or CapCut. Direct mode creates project files directly in the local JianyingPro/CapCut draft directory — open the app and start editing. Built on pyJianYingDraft (PyPI). Trigger words include: JianyingPro, CapCut, push, push to JianyingPro, capcut, export to JianyingPro.
| name | orchestration/multi-agent-orchestration |
| trigger-words | ["orchestration/multi","agent","orchestration"] |
| description | Orchestrate parallel AI agent workstreams — multi-agent deep research (3 parallel experts), research orchestration Trigger phrases: orchestration/multi, agent, orchestration. Converted from Hermes-Skill (Yuno's mastery library). |
Orchestrate parallel AI agent workstreams for research, audits, and development.
set -euo pipefail
Phase 1: Parallel Research (3 experts, ~5-10 min each)
Phase 2: Immediate Fixes (parent acts while reviewing results)
Phase 3: Synthesis (merge reports, prioritize) + 3.5 Cross-Check + 3.6 User-Decision
Phase 4: Execute & Document (P0 fixes first, write report to ~/docs/)
Phase 5: Retrospective (what worked, what didn't)
For recurring health checks, use no_agent: True shell scripts that exit silently on success. Only output when there's a problem.
Use delegate_task(tasks=[...]) with 3 parallel tasks. Full briefing template, scope assignment, toolset table, and parent-direct fallback: see references/phase-1-spawn-experts.md.
Typical scopes (adapt to actual system): Expert 1 Auth/Security/Credentials, Expert 2 Tools/APIs/GPU-Compute, Expert 3 Infra/Messaging/Monitoring.
Critical: Experts without terminal + file produce estimates, not measurements.
While experts run, execute obvious quick wins in the parent: re-enable deactivated toolsets, fix permissions, clear caches, verify subagent claims (don't trust "✅ done").
Deduplicate, conflict-check, verify subagent claims, identify blocking deps, prioritize P0/P1/P2/P3. Two proven sub-procedures are required (full detail in references/phase-3-synthesis.md):
Both sub-phases include the 3-Tier Verifikations-Matrix (Datei-Existenz + Content-Validierung + Realitäts-Check) and 4 Recovery Options for timed-out subagents.
Execute P0 fixes first in dependency order. Write report to ~/docs/builds/ or ~/docs/system/. Update ~/docs/builds/README.md. Report template: references/report-template.md.
Key questions: (1) Were the 3 scopes correct? (2) Did any expert produce estimates instead of measurements? (3) Were CLI commands validated before execution? (4) What would a 4th "Validator" expert have caught?
Die Bienenkönigin orchestriert einen Schwarm aus Subagenten. Das Delegation-Modell ist unabhängig vom Parent-Modell — die einzige Steuerung ist delegation.model in ~/.hermes/config.yaml (Pitfall #10: model-Param wird ignoriert).
Rollen: Königin (stark, höher Kosten, Strategie/Synthese) → Scouts (free tier, parallel, Messungen) → Arbeiter (Cron-Script, no_agent=True, recurring checks). Bei Free-Tier Rate-Limits → Parent übernimmt direkt (Fallback Phase 2).
Full config blocks, free-model recommendations (Owl Alpha, Nemotron, Qwen3 Coder), Token-Plan-aware multi-provider routing (Profile + MoA-Preset patterns), and anti-patterns: see references/queen-bee-configuration.md.
| Criterion | Subagents | Python Batch |
|---|---|---|
| File count | 1-10 per expert | 20-1000+ |
| Check type | Needs reasoning | Deterministic |
| Context cost | High | Low |
| Risk | Phantom-fixes | Parent sees exact output |
The full 20-pitfall detailed list with examples lives in references/pitfalls-detailed.md. For symptom-based trigger-watchlist (load BEFORE every delegate_task): see companion skill multi-agent-pitfalls-cheatsheet.
Top 10 one-liners (must defend against):
--help first — never assume CLI syntax~/docs/system/NAME-DATE.md, never defaults.env/config.yaml writesmodel param ignored — only delegation.model config controls routingCritical advanced pitfalls (full detail in reference):
delegate_task blindly)ls — user-path is a hint, not a contractSeit 2026-06-27 existiert hermes-orchestration als dedizierter Ausführungsskill — dieser Skill bleibt das Pattern-Repository, der Ausführungsskill ist hermes-orchestration.
| Feature | multi-agent-orchestration (Pattern) | hermes-orchestration (Runtime) |
|---|---|---|
| 3-Expert Deep Research | ✅ Phase 1 Pattern | ✅ Task-Typ research |
| Queen-Bee Delegation | ✅ Konfig + Owl Alpha | ✅ Hermes Subagent Bridge |
| Parent-Direct-Fallback | ✅ Pitfall #11 | ✅ Automatisch + --manual |
| Heuristic Extraction / Mnemosyne / Cron | ❌ | ✅ |
Empfehlung: Für neue Runs hermes-orchestration nutzen (es läuft). Diesen Skill für das Pattern-Wissen und die documentierten Pitfalls laden.
For the canonical code pattern (Worker-Role → Toolset mapping, dispatch with parent-direct fallback, worker-contract parsing): see references/hermes-subagent-bridge.md. Used by hermes-orchestration (V2.1) and the-dmz-transfer.
hermes-orchestration (same category) — the runtime with orchestrator.py, metrics_extractor, heuristic_aggregator, mnemosyne_importer, weekly-pipeline cron. Use this for actually running orchestration.multi-agent-pitfalls-cheatsheet — load BEFORE every delegate_task for the top-20 trigger-watchlistsubagent-driven-development — execution pattern using subagentsskill-library-maintenance — class-level skill for slim-down patterns and 4-Tier VerificationFor creating/improving README, DESCRIPTION, NAVIGATION docs across a directory tree. Spawn 4 parallel sub-agents (README, Description, Navigation, Control) + Queen consolidation. Full role definitions, context templates, toolset assignments, risk categories, verification scripts: see references/documentation-planning-pattern.md.
multi-agent-orchestration (this skill) for the patternmulti-agent-pitfalls-cheatsheet BEFORE every delegate_task for trapsreport-template.mdLive demos & known-good baselines: see multi-agent-pitfalls-cheatsheet references (demo-session-2026-07-02.md, 2026-07-02-additions.md).
Converted from Hermes-Skill by Yuno. Original SKILL.md preserved in references/original-hermes.md.