The Agent Skills Directory

[COMMAND_EXECUTION] (SAFE): The skill includes several shell scripts (activator.sh, error-detector.sh, extract-skill.sh) intended to be used as hooks or helper utilities.
extract-skill.sh properly validates the SKILL_NAME input using a strict regex (^[a-z0-9]+(-[a-z0-9]+)*$), preventing shell injection attacks.
error-detector.sh reads the CLAUDE_TOOL_OUTPUT environment variable to check for error patterns but only outputs static reminder text without executing the content of the variable.
[PROMPT_INJECTION] (SAFE): The skill uses XML-tagged reminders (e.g., <self-improvement-reminder>) to guide the agent's behavior. These are meta-instructions for workflow management (logging learnings) rather than attempts to bypass security filters or override the agent's core safety guidelines.
[EXTERNAL_DOWNLOADS] (SAFE): No automatic network requests (curl, wget) were found in the scripts. The documentation references clawdhub for installation, which is a standard package management practice for the targeted platform (OpenClaw).
[DATA_EXFILTRATION] (SAFE): The skill only interacts with local project files (specifically the .learnings/ directory) and does not contain any code for external data transmission.
[CREDENTIALS_UNSAFE] (SAFE): No hardcoded API keys, tokens, or sensitive credentials were found in any of the 13 files.

self-improvement