The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill provides shell script templates (ralph-safety-check.sh and ralph-tier1-check.sh) in references/permission-profiles.md and instructs the user to save them locally and configure platform-specific hooks to execute them at runtime.
[INDIRECT_PROMPT_INJECTION]: The provided safety scripts are designed to ingest and analyze untrusted data (command strings generated by the AI agent) during execution.
Ingestion points: The safety scripts in references/permission-profiles.md read from the CLAUDE_TOOL_INPUT environment variable and stdin.
Boundary markers: The instructions recommend project-local configuration files (.claude/settings.json, config.toml) to limit the scope of automation.
Capability inventory: The scripts utilize bash, python3, and grep to evaluate commands.
Sanitization: The scripts implement regex-based filtering to identify and block a list of Tier 1 dangerous operations.
[COMMAND_EXECUTION]: The skill guides the user to configure platform hooks such as PreToolUse and BeforeTool to execute arbitrary local shell scripts.
[PRIVILEGE_ESCALATION]: Instructions in references/permission-profiles.md direct the user to apply chmod +x to the generated scripts to enable their execution on the system.

ralphmode