OpenClaw Self-Improvement

A reusable operator-guided workflow improvement skill for OpenClaw and ClawLite that turns repeated failures into logged learnings, binary eval loops, SOPs,...

This skill appears to do what it says: local logging, experiments, and promotion into local docs. Before installing or running it, confirm the workspace path it will use (WORKSPACE env var or default ~/.openclaw/workspace). Avoid setting WORKSPACE or OBSIDIAN_LEARNINGS_DIR to system-critical or sensitive directories (e.g., /, /etc, or project roots you don't want modified). Use promote-learning --dry-run to preview promotions; remember log-learning and log-experiment will append to files immediately. If you want extra safety, run the scripts in a dedicated, write-isolated directory or with a disposable user account to limit accidental modifications.

Type: OpenClaw Skill Name: openclaw-self-improvement Version: 0.2.9 The skill provides a mechanism for an AI agent to modify its own core instruction and configuration files (AGENTS.md, TOOLS.md, SOUL.md) based on 'learnings' derived from its operations. While the scripts (e.g., scripts/promote-learning.mjs and scripts/log-learning.mjs) are simple file-appending utilities and the documentation emphasizes operator review, the design inherently enables persistent prompt injection. If an agent is tricked into 'learning' a malicious instruction and then 'promoting' it, the agent's future behavior could be permanently compromised. No evidence of intentional malice or data exfiltration was found, but the capability for self-modification of instructions is a high-risk pattern.