The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the Bash tool to perform experiments during its execution phase. This is the intended primary function of the skill. Security is maintained through a mandatory Phase 1 qualification step where the agent must summarize constraints and steps for the user's explicit approval before any execution occurs.\n- [INDIRECT_PROMPT_INJECTION]: The skill is exposed to untrusted input through $ARGUMENTS, stateful handoff files, and project code read during probing. Evidence chain: (1) Ingestion points: Untrusted data enters via $ARGUMENTS and handoff file loading in Phase 1; (2) Boundary markers: Mandatory plan presentation (step 1.4) and a human-in-the-loop entry gate (step 1.5) serve as checkpoints; (3) Capability inventory: Uses Bash, Write, Task, and WebFetch for operation; (4) Sanitization: Relies on human review of the plan at the gate before Phase 2 begins. This structure reduces the likelihood of the agent autonomously executing malicious instructions embedded in project data.\n- [SAFE]: Documentation includes comprehensive security orientations, prohibiting production traffic usage and irreversible actions. Functional anti-patterns are clearly defined with 'WHY' explanations. A defensive check for the AskUserQuestion harness bug is implemented to prevent the agent from proceeding without valid user input. The skill's external references point to official vendor repositories on GitHub.

probe