The Agent Skills Directory

[PROMPT_INJECTION]: The skill facilitates Indirect Prompt Injection by instructing the agent to capture untrusted data (user corrections, tool errors, and external feedback) and store it in .learnings/ log files. These logs are then intended to be "promoted" into system-level instruction files such as CLAUDE.md, AGENTS.md, and SOUL.md. This architectural pattern allows external content to influence the agent's long-term behavioral rules and project conventions without automated sanitization. Ingestion points include the .learnings/ directory files; boundary markers are absent during the promotion process; and the capability inventory includes shell script execution and persistent file modification. Sanitization is limited to instructional warnings for manual redaction.\n- [COMMAND_EXECUTION]: The skill provides and encourages the use of multiple shell scripts (activator.sh, error-detector.sh, extract-skill.sh) for initialization and automated reminders. These scripts are configured to run via agent hooks (e.g., in .claude/settings.json), resulting in persistent command execution whenever the agent is bootstrapped or tools are used. The error-detector.sh script specifically processes the CLAUDE_TOOL_OUTPUT environment variable to search for error patterns.\n- [EXTERNAL_DOWNLOADS]: The documentation directs users to install components from a third-party GitHub repository (github.com/peterskoett/self-improving-agent.git) and uses the clawdhub package registry for installation.

self-improvement