Tmp Self Improving Agent

Captures learnings, errors, and corrections to enable continuous improvement. Use when: (1) A command or operation fails unexpectedly, (2) User corrects Clau...

This skill appears to do what it says: inject lightweight reminders and provide helper scripts to log and promote learnings locally. Before enabling hooks or running the scripts: 1) Inspect scripts (activator.sh, error-detector.sh, extract-skill.sh) yourself — extract-skill.sh creates files on disk; 2) Note the error-detector expects CLAUDE_TOOL_OUTPUT from the host hook environment (this env var is not listed in the skill metadata); 3) When enabling hooks, prefer the activator-only setup first (read-only injection) and avoid auto-running extract-skill.sh unless you intend to let the skill create files; 4) If you install from the GitHub repo, review that remote repo for any changes; 5) If you have strict security requirements, run the skill in a confined user account or sandbox until you are comfortable.

Type: OpenClaw Skill Name: tmp-self-improving-agent Version: 1.0.0 The skill bundle implements a 'self-improvement' loop that instructs the agent to capture errors, user corrections, and observations into markdown files, which are then 'promoted' into core instruction files such as SOUL.md, AGENTS.md, and CLAUDE.md. This architecture creates a significant attack surface for indirect prompt injection; an attacker could potentially poison the agent's future behavior by triggering specific errors or providing 'corrections' that contain malicious instructions. While the included utility scripts (e.g., extract-skill.sh in the scripts/ directory) demonstrate good security practices like path traversal protection and input validation, the core design of persisting untrusted data into system-level prompts is inherently high-risk. No evidence of intentional malice was found, but the self-modifying nature of the skill warrants caution.