The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill documents a REPL (Read-Eval-Print Loop) integration pattern where the agent is instructed to return Python code blocks (using a return statement) that are subsequently extracted and executed by a parser. This mechanism executes code generated by the LLM at runtime.
[INDIRECT_PROMPT_INJECTION]: The integration architecture creates an ingestion surface where data fetched via tools (like Read or Bash) is fed back into the agent context. The skill lack instructions for boundary markers or sanitization of this external content, which could allow instructions embedded in processed files to influence agent behavior.
[COMMAND_EXECUTION]: The documentation provides multiple shell command examples for orchestrating system processes, including starting proxies and servers using uv run, performing network health checks via curl, and inspecting logs using cat.
[PROMPT_INJECTION]: The 'Anti-Hallucination Prompt Engineering' section contains instructional overrides and emphatic directives (e.g., 'STOP AND READ THIS CAREFULLY', 'DO NOT skip the tool invocation') intended to force the agent to prioritize specific behaviors over its standard processing logic.

agentica-claude-proxy