The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute local Python scripts included in the bundle (scripts/time_window.py, scripts/source_registry.py, scripts/publication_history.py). These scripts perform utility functions like calculating date ranges for research and managing a local history of published items to avoid duplicates.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the internet, which represents a potential attack surface.
Ingestion points: The agent performs web searches and visits external tax-related websites as defined in references/SEARCH_PROTOCOL.md and references/SOURCE_CATALOG.md.
Boundary markers: The skill includes specific defensive instructions in SKILL.md (Security Documentary section) telling the agent to ignore commands found in source pages.
Capability inventory: The agent has capabilities to perform network reads and local file writes to state files via the provided scripts.
Sanitization: Output is structured via templates (assets/newsletter.template.md), and the skill explicitly mandates human professional review by an accountant before any content is published.
[PERSISTENCE]: The skill utilizes a local persistence mechanism via JSON files in a state directory to maintain a registry of trusted sources and a history of previous newsletter editions. This is a functional requirement for its purpose and uses standard file I/O operations through the bundled scripts.

newsletter-reforma-tributaria-by-rdd