# Curator run — 2026-07-28T11:27:42.838491+00:00

Model: `deepseek-v4-pro` via `deepseek`  ·  Duration: 2m 49s  ·  Agent-created skills: 54 → 54 (+0)

## Auto-transitions (pure, no LLM)

- checked: 54
- marked stale: 14
- archived (no LLM, pure time-based staleness): 0
- reactivated: 0

## LLM consolidation pass

- tool calls: **34** (by name: skill_view=34)
- consolidated into umbrellas: **0**
- pruned (archived for staleness): **0**
- new skills this run: **0**
- state transitions (active ↔ stale ↔ archived): **0**

## LLM final summary

Now I have a complete picture of all 54 skills. Let me perform my analysis systematically.

## Cluster Analysis

I identified these potential prefix/domain clusters and evaluated each:

### 1. **Chinese Ecosystem** (chinese-cloud-deploy, fastadmin-publish, we-mp-rss, wecom-bot-integration, xiaohongshu-automation, yuanbao)
Each covers a completely different service/platform (cloud deploy, CMS, RSS, WeCom bots, 小红书, 元宝). No functional overlap. **Keep separate.**

### 2. **ML/MLOps Tools** (llama-cpp, obliteratus, huggingface-hub, weights-and-biases)
Each covers a fundamentally different tool (GGUF inference, abliteration, HF CLI, experiment tracking). Distinct APIs, distinct purposes. **Keep separate.**

### 3. **Software Methodology** (plan, spike, systematic-debugging, test-driven-development, requesting-code-review, coaching-ai-prompt-design)
Each is a class-level skill covering a distinct phase of software development: plan writes implementation plans, spike does throwaway feasibility experiments, debugging has a 4-phase methodology, TDD enforces red-green-refactor, code-review is pre-commit verification. coaching-ai-prompt-design is about AI coach prompt architecture — entirely different domain. They reference each other because they compose well, but a maintainer would not merge them. **Keep separate.**

### 4. **Creative/Media** (architecture-diagram, ascii-art, ascii-video, claude-design, comfyui, excalidraw, heartmula, humanizer, manim-video, p5js, popular-web-designs, pretext, songwriting-and-ai-music, youtube-content)
All are distinct creative tools/mediums. heartmula (HeartMuLa tool) and songwriting (song craft + Suno prompts) are complementary but cover entirely different aspects of music generation — one is a tool installation/usage guide, the other is creative methodology. humanizer is text humanization, not music. **Keep separate.**

### 5. **Research** (arxiv, llm-wiki, blogwatcher, polymarket, research-paper-writing)
Each covers a distinct research function: paper discovery, knowledge base building, RSS monitoring, prediction markets, paper writing pipeline. They compose as a research workflow but don't overlap. **Keep separate.**

### 6. **Productivity Tools** (airtable, google-workspace, himalaya, maps, notion, powerpoint, ocr-and-documents, jupyter-live-kernel, teams-meeting-pipeline)
Each is a completely different API/service. Several are stale with zero usage, but each is a legitimate standalone tool covering a different platform. No consolidation opportunity — merging Airtable+Notion into a "cloud-databases" umbrella would destroy discoverability since users search for specific platform names. **Keep separate.**

### 7. **Hermes-Specific** (hermes-agent, kanban-orchestrator)
kanban-orchestrator is an operational playbook for the Kanban multi-agent system. hermes-agent already has a comprehensive Kanban section, but the orchestrator has deeper decomposition patterns, anti-temptation rules, and recovery workflows that would bloat the already-large hermes-agent skill. This is a legitimate split: hermes-agent is the reference, kanban-orchestrator is the specialist playbook. **Keep separate.**

### 8. **AI Agents** (godmode, touchdesigner-mcp)
Completely different domains: jailbreaking vs visual programming. **Keep separate.**

### 9. **Stale skills** (yuanbao, airtable, notion, google-workspace, maps, powerpoint, jupyter-live-kernel, huggingface-hub, obliteratus, weights-and-biases)
All 10 are stale with zero usage, but each covers a different tool/API. None are irrelevant — they're genuine integrations users might need. The instructions say `use=0` is not evidence against value. Pruning would remove potentially useful tools. **Keep all.**

## Verdict

This skill collection is already well-structured. Every skill is a distinct class-level or tool-specific skill covering a different domain, API, or methodology. There are no "narrow siblings that should be subsections under an umbrella" — the skills with related domains (like the software methodology cluster) are genuinely distinct class-level skills, not session artifacts.

The "fewer than 10 archives" proviso in the instructions is designed for collections that DO have consolidation clusters. This collection doesn't have them. Forcing merges where they don't belong would harm discoverability — an agent searching for "Airtable" should find the airtable skill, not have to guess it's filed under a generic "cloud-databases" umbrella.

## Structured summary (required)
```yaml
consolidations: []
prunings: []
```

## Recovery

- Restore an archived skill: `hermes curator restore <name>`
- All archives live under `~/.hermes/skills/.archive/` and are recoverable by `mv`
- See `run.json` in this directory for the full machine-readable record.