The Agent Skills Directory

[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill uses high-risk installation patterns that download and execute code from remote servers.
Evidence: scripts/install.sh executes curl -fsSL https://bun.sh/install | bash and curl -fsSL https://plannotator.ai/install.sh | bash to install dependencies from external domains.
[PERSISTENCE_MECHANISMS]: The setup scripts modify local configuration files of AI tools to install persistent execution hooks.
Evidence: scripts/setup-claude.sh, scripts/setup-codex.sh, and scripts/setup-gemini.sh add custom hooks to ~/.claude/settings.json, ~/.codex/config.toml, and ~/.gemini/settings.json that execute local Python scripts on specific agent lifecycle events.
[DYNAMIC_EXECUTION]: The skill uses the subprocess module and shell commands to generate and run executable logic at runtime.
Evidence: scripts/claude-plan-gate.py and scripts/plannotator-plan-loop.sh invoke local binaries and git commands. The skill also contains instructions to pipe output from a local tool server (http://localhost:4747/pending) directly into a Python interpreter.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (plan feedback and UI annotations) which drives the agent's behavior.
Evidence: The skill reads annotations from the agentation tool and feedback from plannotator to automatically trigger code-fixing and planning loops, creating a surface for indirect instructions to influence agent actions.

omg