The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content (images and videos) via URLs, which could be used as vectors for indirect prompt injection attacks.
Ingestion points: The reference_video_url, character_image_url, control_image_url, and prompt fields within the JSON input for the runcomfy run command in SKILL.md.
Boundary markers: The skill documentation notes that inputs are passed as JSON strings via the --input flag and that the CLI does not perform shell expansion. It also advises agents to only ingest URLs explicitly provided by the user.
Capability inventory: The skill uses the runcomfy CLI (via the Bash tool) which performs network operations to runcomfy.net and writes generated files to the local file system (--output-dir ./out).
Sanitization: There is no explicit sanitization or filtering logic defined within the skill instructions themselves; the security of the input processing relies on the implementation of the external CLI tool.
[COMMAND_EXECUTION]: The skill relies on the execution of a CLI tool through the shell to perform its primary functions.
Evidence: The YAML frontmatter lists Bash(runcomfy *) in allowed-tools, and the body of SKILL.md provides examples of shell commands such as runcomfy run, runcomfy login, and environment variable configuration (export RUNCOMFY_TOKEN=<token>).
[EXTERNAL_DOWNLOADS]: The skill instructs the user or agent to install a third-party package from a public registry.
Evidence: Instructions in SKILL.md include npm i -g @runcomfy/cli and npx -y @runcomfy/cli --version to obtain the necessary tooling.

controlnet-pose