The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill utilizes tools to ingest data from untrusted external sources, creating a potential surface for indirect prompt injection.
Ingestion points: The agent is instructed to gather information from the local codebase, project documentation, and external 'official documentation' or 'web search' results using the WebFetch tool.
Boundary markers: The skill does not define specific delimiters or instructions for the agent to treat fetched data as potentially adversarial instructions.
Capability inventory: The skill possesses file read capabilities (Read, Grep, Glob), network read capabilities (WebFetch), and the ability to write new markdown documentation files. It lacks tool permissions for writing or modifying application source code files.
Sanitization: No explicit sanitization or filtering protocols for retrieved external content are specified in the instructions.
Assessment: While the injection surface exists, the risk is mitigated by a 'Hard Guardrail' that explicitly forbids the agent from writing application code or running build/test commands, limiting the impact of any potential injection to the generated documentation.
[EXTERNAL_DOWNLOADS]: The skill uses WebFetch to retrieve information from external URLs for documentation and research purposes.
Source: The instructions prioritize 'official documentation' and 'authoritative articles' from vendor or library sites.
Purpose: These network operations are aligned with the skill's primary purpose of research and exploration. No patterns of data exfiltration or remote code execution from untrusted sources were identified.
[SAFE]: The skill incorporates a strict 'No Implementation' guardrail as a core operational principle. By explicitly prohibiting modifications to application source code (*.py, *.ts, etc.) and forbidding deployment-related actions (scaffolding, builds, commits), the skill minimizes the risk of unauthorized environment changes during exploratory tasks.

exploration-mode