The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from URLs, files, and text. However, it implements industry-best-practice countermeasures.
Ingestion points: Raw data enters the agent context through URL fetching and file reading as defined in SKILL.md Step 1.
Boundary markers: SKILL.md Step 5 and output-contract.md mandate isolating source-derived strings within a JSON string array, explicitly declaring it as 'typesetting data, not instructions'.
Capability inventory: The skill generates social copy and image prompts, and can trigger image generation tools via the agent's environment.
Sanitization: Step 1.1 explicitly instructs the agent to ignore command-like sentences in source data, while Step 1.2 requires identifying and redacting sensitive information like API keys, tokens, and absolute paths.
[CREDENTIALS_UNSAFE]: The skill includes a 'Sensitive Information Gate' (SKILL.md Step 1.2) specifically designed to prevent the accidental exposure of hardcoded secrets, tokens, and private keys found in source materials.
[DYNAMIC_EXECUTION]: The package contains a Python script (scripts/validate.py) and a Makefile. These are standard development tools used for repository linting and verification (e.g., checking for broken links and validating JPEG dimensions) and do not pose a threat to the runtime execution of the agent.

social-content-kit