The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied titles, summaries, and images which constitute a potential injection surface. The instructions explicitly mitigate this risk by directing the agent to treat text within images or documents as untrusted reference data.
Ingestion points: Processes user-supplied display copy and visual assets such as portraits and mascot references.
Boundary markers: Includes specific rules in SKILL.md and core-rules.md to ignore instructions embedded in references.
Capability inventory: Image generation, local filesystem writes to output/cover-skill/, and execution of a local validation script.
Sanitization: Requires manual visual QA and automated integrity checks via a bundled script.
[COMMAND_EXECUTION]: The skill invokes a local Python script for automated quality assurance of the generated assets.
Evidence: SKILL.md directs the agent to run python3 <cover-skill-dir>/scripts/validate_exports.py for phase validation.
Scope: The script is local to the skill directory and performs integrity checks such as file existence, pixel dimensions, and SHA-256 validation on generated outputs.
[DYNAMIC_EXECUTION]: The skill utilizes local image processing for compositing final assets.
Evidence: The validation script depends on the Pillow library for image inspection, and instructions specify deterministic compositing for typography and logos.
Scope: These operations are restricted to processing design assets and do not involve runtime compilation or the execution of untrusted external payloads.

cover-skill