The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements an iterative workflow that is vulnerable to indirect prompt injection, as it relies on processing and acting upon text-based critiques fetched from external tool outputs.
Ingestion points: The agent is instructed to use asset_download to read verdicts and failure reasons generated by the asset_analyze and asset_quality_gate_run tools.
Boundary markers: The skill attempts to manage the data by prescribing a specific output shape (<index>: pass|fail, <the failed line>) for the critic tool, though this does not fully isolate the instructions from the content.
Capability inventory: The agent can execute powerful tools including model_run, scenario-image (for masked inpainting), and scenario-image-editing (for deterministic passes) based on the ingested text.
Sanitization: There is no mention of sanitization, filtering, or escaping of the critique text before it is interpolated into the agent's decision-making logic.
[EXTERNAL_DOWNLOADS]: The skill references external packages and provides instructions for the user to install them.
Evidence: The documentation directs the agent to ask the user to execute npx skills add scenario-labs/skills --skill <name> if required sibling skills are missing. These resources originate from the skill author's official ecosystem.

scenario-refine-loop