Agent-Skills.md

Agent Skills: Search Comfy Workflows

Search for ComfyUI workflow precedents, VibeComfy ready templates, node wiring examples, Banodoco/Hivemind community workflow knowledge, and concrete graph patterns before editing or adding a workflow. Use when the user asks to find workflows, examples, precedents, ComfyUI node combinations, Hivemind/Banodoco evidence, or which workflow to start from.

UncategorizedID: peteromallet/vibecomfy/search-comfy-workflows

Repository

peteromallet/vibecomfy
peteromalletLicense: MIT
15119

Install this agent skill to your local

pnpm dlx add-skill https://github.com/peteromallet/VibeComfy/tree/HEAD/docs/agent-skill/skills/search-comfy-workflows

Skill Files

Browse the full folder contents for search-comfy-workflows.

Download Skill

Loading file tree…

docs/agent-skill/skills/search-comfy-workflows/SKILL.md

Skill Metadata

Name
search-comfy-workflows
Description
Search for ComfyUI workflow precedents, VibeComfy ready templates, node wiring examples, Banodoco/Hivemind community workflow knowledge, and concrete graph patterns before editing or adding a workflow. Use when the user asks to find workflows, examples, precedents, ComfyUI node combinations, Hivemind/Banodoco evidence, or which workflow to start from.

Search Comfy Workflows

Use this to find evidence before editing or adding a graph. The result should tell the next agent what to open, why it matches, which node classes matter, and what is still uncertain. A found JSON workflow is a precedent to import and inspect; it is not itself an execution instruction.

Fast Path

Start local. Escalate to Hivemind when local search does not have the precedent, or when the user asks for community/Banodoco practice.

vibecomfy workflows list --ready
vibecomfy workflows list
vibecomfy search "wan i2v controlnet" --task i2v --limit 10 --json
vibecomfy inspect <workflow_id_or_path>
vibecomfy analyze info <workflow_id_or_path>

Run vibecomfy sources sync only when indexes are stale and generated index updates are acceptable.

Hivemind

Use Hivemind for current ComfyUI practice, Banodoco workflows, Kijai/Ablejones node usage, settings, model notes, and community workflow examples. Local search is for the selected imports and curated adapters already present in this checkout.

Use the deployed Hivemind pack through Astrid. This is the single supported route; it pins the installed pack and applies the runtime's retry/error handling:

python3 -m astrid hivemind search "wan animate workflow openpose" \
  --kinds workflow --limit 10 --json

Fetch the exact accepted resource/revision through the Astrid Hivemind SDK (hivemind.get_item) after search. Do not call Hivemind's PostgREST tables or the retired unified_feed/contribute-resource paths directly from a workflow task.

Favor topic-specific sources such as wan_comfyui, wan_resources, ltx_chatter, ltx_resources, comfyui, resources, and daily_summaries.

One workflow lifecycle

Hivemind is the discovery and provenance layer. VibeComfy owns the local workflow artifacts:

  1. Search Hivemind and pin the accepted resource/revision.
  2. Pull that one workflow on demand with the normal importer, for example vibecomfy import hivemind:external_resources:<id> or vibecomfy import hivemind://resource/<id>[/revisions/<revision>]. The importer writes the single local bundle under workflows/<source-id>/ and retains the Hivemind evidence ID, any provider revision, and hash in bundle metadata; when the provider exposes no revision, the hash is recorded as the local snapshot pin.
  3. Edit, validate, and run that local bundle (or copy it to recipes/).
  4. Use vibecomfy templates create ... --id ... only when deliberately promoting a validated workflow into a curated ready_templates/ adapter. Any source snapshot shipped with that adapter is generated from the pinned Hivemind input; it is not a second source of truth.

Do not bulk-mirror Hivemind into external_workflows/. Hivemind is the canonical shared workflow catalogue; workflows/ is the single local on-demand ingestion/cache location. ready_templates/ contains only curated executable adapters.

Evidence Standard

  • A broad message hit is not enough. Extract exact workflow, node class, field, socket, model, or pack evidence.
  • Prefer local ready templates when the user wants something runnable immediately.
  • Prefer Hivemind resources/workflows over chatter; use chatter for clues, not final wiring.
  • If no precedent exists, say that directly and describe the next safest inspection step.
  • Keep the handoff short: best candidates, match rationale, node/model requirements, and blockers.