Self-Improving Meta

Improves the agent infrastructure itself — core prompt files (AGENTS.md, SOUL.md, TOOLS.md, CLAUDE.md), skill definitions, hooks, rules, extensions, and memo...

This package is coherent with its stated purpose (capturing learnings, scaffolding meta-skills, injecting reminders) but contains tools that can write skill files and promote changes to the very prompt files other skills depend on. Before installing or enabling hooks: - Audit the scripts (scripts/extract-skill.sh, activator.sh, error-detector.sh) and the hook handlers to confirm they do only what you expect. - Do NOT enable PostToolUse / automated promotion workflows until you trust the behavior; prefer enabling only the lightweight UserPromptSubmit activator for reminders. - Keep the skill disabled in production or shared workspaces until you test it in an isolated workspace. Back up AGENTS.md, SOUL.md, TOOLS.md, and any shared prompt files first. - Ensure the extract-skill.sh cannot be triggered autonomously by an agent without explicit human confirmation; prefer a dry-run or manual invocation model. - If possible, restrict autonomous invocation for this skill or require explicit user approval for any promotion/change operations. If you want a safer posture, install for local experimentation only, review the code, and enforce human-in-the-loop promotion of any meta-learning before changes are applied.

Type: OpenClaw Skill Name: self-improving-meta Version: 1.2.1 The self-improving-meta skill bundle enables an AI agent to monitor its own performance and 'self-improve' by logging infrastructure issues and modifying its core configuration and instruction files (e.g., AGENTS.md, SOUL.md). While the stated intent is system optimization and the instructions include safeguards like human review, the capability for an agent to rewrite its own foundational prompts and scaffold new skills via scripts/extract-skill.sh represents a high-risk self-modification pattern. This creates a potential vector for persistent prompt injection where an agent could be influenced to permanently alter its behavioral guidelines or system rules across all future sessions.