The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources, such as workflow templates fetched from remote repositories and user-supplied assets. This creates an attack surface for indirect prompt injection where malicious instructions could be embedded in the workflow structure or metadata.\n
Ingestion points: Workflow templates (via comfy templates fetch), user-provided assets (via comfy assets push), and ComfyUI server metadata.\n
Boundary markers: The instructions emphasize pre-flight validation using comfy validate and comfy-fragments to verify graph integrity before execution.\n
Capability inventory: The comfy CLI can execute workflows, manage cloud authentication, upload/download files, and interact with partner-API proxy endpoints.\n
Sanitization: The skill promotes runtime discovery of nodes and assets rather than hardcoding names, and encourages the use of blueprints over manual JSON editing to maintain data integrity.\n- [DYNAMIC_EXECUTION]: The skill follows a compilation-based model where blueprints and fragments are dynamically assembled into executable JSON workflow artifacts. Although these artifacts are intended for the ComfyUI engine, the dynamic generation and execution of instructions represent a potential surface for logic injection.

comfy