comfyui-research（舒适研究）

ComfyUI Research Skill

Keeps the VideoAgent knowledge base current by monitoring sources, extracting knowledge, and flagging stale information.

Triggers

• On demand: User invokes /research comfyui or asks to check for updates
• Session start: Quick staleness check (read references/staleness-report.md)
• Before workflow generation: Targeted check if the specific model/technique is >2 months old

Research Pipeline

1. Discovery

Search for new developments across three source types:

YouTube Channels (technique extraction):

GitHub Repositories (node/model updates):

HuggingFace (model releases):

• Trending models: huggingface.co/models?sort=trending
• Filter: diffusers, video, audio, text-to-image
• Watch for: New FLUX variants, Wan updates, identity models, voice models

2. Extraction

For each discovery, extract structured knowledge:

From YouTube videos: Use youtube-video-analyst or youtube-chapter-clipper skills to:

1. Get transcript
2. Identify new models mentioned (name, source, requirements)
3. Identify new techniques (workflow changes, parameter discoveries)
4. Identify new nodes (package name, what it does)
5. Note any benchmark results or comparisons

From GitHub releases:

1. Check release notes for breaking changes
2. Identify new features or supported models
3. Note version requirements
4. Check if deprecated features affect our workflows

From HuggingFace:

1. Model card: architecture, training data, requirements
2. VRAM requirements and quantization options
3. ComfyUI compatibility (is there a node?)
4. Quality comparison vs current recommendations

3. Integration

Update reference files with timestamped entries:

<!-- Updated: 2026-02-06 | Source: Pixaroma Ep 45 -->
### New Model: XYZ
- Download: {url}
- VRAM: {requirement}
- Notes: {key findings}

Which file to update:

4. Staleness Audit

Scan all reference files for <!-- Updated: --> timestamps:

Thresholds:

Generate report: Update references/staleness-report.md with current status.

Research State

Maintained in skills/comfyui-research/state/:

state/
  last-research-run.json    # When, what was checked, what was found
  watch-list.yaml           # Sources to monitor (see references/)

last-research-run.json

{
  "last_run": "2026-02-06T12:00:00Z",
  "sources_checked": 15,
  "findings": 3,
  "updates_made": [
    {"file": "models.md", "entry": "Added ModelX", "source": "HuggingFace"},
    {"file": "workflows.md", "entry": "Updated Wan settings", "source": "Pixaroma Ep 50"}
  ],
  "stale_entries": 0
}

Quick Staleness Check (Session Start)

Fast check without full research:

1. Read references/staleness-report.md
2. If last_run is >2 weeks ago:

- Notify: "Research data is {N} days old. Run /research comfyui for updates."

1. If any entries are beyond their threshold:

- Notify: "Stale items found: {list}. These may have newer alternatives."

Full Research Run

When user triggers /research comfyui:

1. Check all YouTube channels for recent uploads (last 30 days)
2. Check all GitHub repos for new releases
3. Check HuggingFace trending for relevant models
4. Extract findings using youtube-video-analyst
5. Update reference files with timestamped entries
6. Run staleness audit on all references
7. Update staleness report
8. Summarize findings for user

Targeted Research

When asked about a specific topic:

1. Search YouTube for recent videos on the topic
2. Search HuggingFace for related models
3. Check GitHub for related node packages
4. Synthesize findings
5. Update relevant reference files if significant discoveries

Integration with Other Skills

Reference

• references/staleness-report.md - Current freshness status
• references/evolution.md - Update protocol and changelog
• skills/comfyui-research/references/watch-list.yaml - Monitored sources