The Agent Skills Directory

[SAFE]: The skill explicitly defines a 'read-only' rule that prohibits the agent from performing state-changing operations (writing files, committing to git, etc.) by default. These capabilities are only unlocked when the user explicitly names a file for a specific action.
[SAFE]: The skill leverages standard research tools such as WebSearch, WebFetch, and Read for its primary purpose of information gathering, with no evidence of unauthorized data exfiltration or credential harvesting.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest data from repository files and web sources, creating a potential surface for indirect prompt injection.
Ingestion points: Repository content retrieved via Read, Grep, and Glob tools; external web content retrieved via WebSearch and WebFetch.
Boundary markers: No explicit delimiters are specified for external data, but the skill instructs the agent to 'Compress the result' and provide 'conclusions and evidence' rather than raw file dumps.
Capability inventory: The agent has access to file system operations (read/write), shell execution (via Bash), and network tools.
Sanitization: The mandatory summarization of subagent reports into two or three sentences acts as a filter that reduces the effectiveness of adversarial instructions embedded in external content.

discuss