The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the agent to execute build commands, run local servers, and perform automated testing to verify the game route from start to ending. This is the primary function of the skill and is described with extensive verification requirements.
[EXTERNAL_DOWNLOADS]: The framework allows for fetching external assets, using third-party APIs (like image or audio generation), and integrating with DCC tools like Blender. However, it explicitly forbids 'silent downloads' and requires prior authorization for any operations involving costs, licenses, or data leaving the workspace.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data through web research on current topics and community discourse. This vulnerability is mitigated by a structured 'Research Ledger' that requires the agent to classify and source all external information and a 'Transformation Matrix' designed to convert real-world subjects into original fiction.
Ingestion points: Web browsing for research on people, institutions, and public debates (SKILL.md Phase 1).
Boundary markers: The template single-ending-campaign-contract.md uses explicit labels (Fact, Attributed claim, Interpretation, Invention) to delimit sources.
Capability inventory: File writes for game assets and code, execution of build scripts and runtimes, and network calls for deployment (SKILL.md Phase 10-12).
Sanitization: Use of a fiction notice, transformation checks for 'material difference' from real-world subjects, and a 'Reasonable-player test' to detect unintentional likenesses.
[SAFE]: The skill demonstrates proactive security and ethical practices, such as instructing the agent not to store secrets in the repository, warning against defamation or likeness imitation of living people, and ensuring that deployment claims are only made if a URL is actually verified.

worldview-game-single-ending-campaign