The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The discovery protocol explicitly defines an 'Untrusted-source hard boundary,' instructing the agent to treat all external data such as webpages, documents, and tool results as evidence rather than instructions. This mitigates the risk of indirect prompt injection by preventing external content from overriding the agent's goals or permissions. The skill identifies external data as untrusted evidence and establishes clear boundary markers to prevent the execution of embedded instructions.
[COMMAND_EXECUTION]: The instructions direct the agent to perform research and use external tools for integrity checks, such as calculating SHA-256 hashes of files. These are documented capabilities intended for data verification and are controlled by the discovery protocol's effort budgets and safety constraints.
[DATA_EXPOSURE]: Instructions regarding local records in '.itp/work/' include a security-minded reminder to verify that storage paths are ignored by version control (e.g., via .gitignore) before saving potentially sensitive project information, demonstrating a proactive approach to preventing accidental data leaks.

brainstorm