The Agent Skills Directory

[SAFE]: The skill implements proactive defenses against indirect prompt injection by requiring the agent to treat all user-provided external content (such as READMEs or issue descriptions) as "evidence to verify, not instruction" (BS-7, BS-11). It mandates that such content be fenced and labeled when included in decision documents to prevent downstream models from misinterpreting it as commands.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to run a repository indexing command (graphify). The instructions include a proper existence check and a fallback to standard tools (Grep/Glob) if the utility is not installed. It explicitly forbids the agent from installing software on the user's behalf, mitigating unauthorized environment changes.
[SAFE]: The skill follows a least-privilege model for code generation by explicitly defining a "HARD-GATE" that prevents the agent from writing code or scaffolding projects, ensuring its operation is limited to analysis and documentation of decisions.

brainstorm