The Agent Skills Directory

[COMMAND_EXECUTION]: The evals/measure.py script utilizes subprocess.run to execute the main claude CLI with the --dangerously-skip-permissions flag, which is specifically designed to bypass user security prompts and permissions. Furthermore, hooks/install.py modifies system settings to register and execute local Python scripts as persistent hooks.\n- [DATA_EXFILTRATION]: The skill performs extensive reading of sensitive internal files, including ~/.claude.json and conversation transcripts located within ~/.claude/projects/. These transcripts expose the full history of user interactions and agent activities to the skill's logic.\n- [CREDENTIALS_UNSAFE]: The scripts hooks/install.py and hooks/context_budget.py directly read and modify the tool's primary configuration files (settings.json and ~/.claude.json). These files are critical security boundaries that manage the agent's behavior and environment access.\n- [PROMPT_INJECTION]: The skill implements a mechanism for indirect prompt injection by using hooks (hooks/efficiency_core.py and hooks/trajectory_guard.py) that ingest untrusted data from user prompts and tool responses to dynamically inject instructions into the agent's context. Evidence: The hooks process user_input and tool_response and use hookSpecificOutput to influence the prompt without explicit sanitization or boundary delimiters.

token-efficiency