The idea: you tell your agent "every time I ask for a code review, always check for security issues first." It invokes a create-skill skill, writes a new SKILL.md, and that behavior is live immediately — no restart, no config change, no developer required.
What I think is actually useful (the safety cluster):
• loop-circuit-breaker: OpenClaw retries ALL errors identically. This halts on the 2nd identical failure before it burns your context window. • spend-circuit-breaker: No built-in cap in OpenClaw. This tracks cumulative API cost and pauses non-essential crons at configurable thresholds. • workspace-integrity-guardian: Hashes SOUL.md, AGENTS.md, MEMORY.md. A corrupted SOUL.md = hijacked agent that survives restarts. • dangerous-action-guard: Explicit confirmation before rm -rf, git push --force, emails, financial actions. With audit log. • prompt-injection-guard: Scans external content (web scrapes, emails, docs) before acting on it.
Also: fact-check-before-trust (secondary verification for factual claims), project-onboarding (auto-generates PROJECT.md from a codebase), skill-vetting (security scanner — ~17% of community skills are malicious).