The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the local workspace to inform the design process.
Ingestion points: The skill reads project context from files such as README, product docs, package or app metadata, landing-page copy, manifests, and design tokens.
Boundary markers: There are no instructions defining delimiters or warning the agent to ignore instructions embedded within the processed workspace files.
Capability inventory: The gathered context is used to drive the image_generate tool and potentially delegate_task for parallel candidate generation.
Sanitization: No validation or sanitization of the content from the external repository files is described.- [PROMPT_INJECTION]: The skill contains instructions that explicitly mandate the bypass of typical agent oversight and output filtering.
Evidence: The document states "do NOT auto-inspect/retry/filter candidates" and "Do not inspect outputs to block delivery, classify them as recommended or non-recommended, retry them automatically, or repair them with post-processing."
Description: This directive forces the agent to skip safety and quality checks on generated images, which could lead to the delivery of unverified or inappropriate content produced by the image generation backend.

ip-as-logo