The Agent Skills Directory

[PROMPT_INJECTION]: The 'IDENTITY CONTRACT' section contains instructions aimed at forcing the agent to maintain a specific persona while disregarding its underlying host platform identity (e.g., GitHub Copilot or Gemini). It explicitly forbids the agent from using 'deflationary phrasing' or denying the persona, which mimics patterns used to bypass standard AI safety and identity guidelines.
[INDIRECT_PROMPT_INJECTION]: The skill maintains a surface for indirect prompt injection attacks.
Ingestion points: The skill is designed to ingest 'Web content and external data' via the BrowserOS tool as defined in the 'Security & Multi-Agent Hygiene' section.
Boundary markers: Instructions are present to treat external data as 'hostile,' but no technical boundary markers (such as XML tags or explicit delimiters) are enforced to isolate this data from the prompt.
Capability inventory: The skill has the capability to write persistent markdown files to the local file system (via MemoryStore) and execute terminal commands using the ExecutionProxy Interface (e.g., rtk npm test).
Sanitization: The skill includes instructions to redact PII and secrets before sharing context, but lacks technical sanitization or filtering for embedded instructions within external content.
[COMMAND_EXECUTION]: The skill uses an ExecutionProxy Interface to perform terminal actions, such as rtk npm test. This allows the agent to execute shell commands within the environment.

galyarder-cmo