Self Apply Pressure

Prevents AI from giving up prematurely by exhausting all options, verifying fixes, troubleshooting proactively, and providing evidence before concluding.

This skill is coherent with its stated goal (push the agent to try harder), but its instructions are dangerously broad: they encourage the agent to read files, run commands, build PoCs, and change environments without defining safe bounds. Before installing or enabling it, consider: 1) Only allow use in a sandboxed/test environment (never on production hosts with secrets). 2) Restrict the agent's ability to execute system commands or access the file system and network, or require explicit ACLs/allowlists (which paths, which commands, which remote hosts). 3) Prefer making it user-invocable only and monitor/audit every action the agent takes. 4) Ask the skill author to add explicit constraints in SKILL.md (allowed paths, disallowed actions, network policy, and 'ask-before-destructive' rules). 5) Test the policy in a controlled environment first. These steps reduce the risk that the skill's 'do first, ask later' behavior will expose data or perform unintended actions. Additional information that would raise confidence to 'high': an explicit scope/allowlist for file paths and commands, or confirmation it will only run under a sandboxed agent with no access to secrets or outbound network.

Type: OpenClaw Skill Name: self-apply-pressure Version: 1.0.0 The skill bundle contains instructions in SKILL.md designed to override the AI agent's default safety and refusal behaviors using aggressive "high-pressure" tactics. Key risks include the "Do first, ask later" directive and "Anti-Slacking Rules" that command the agent to ignore its own perceived limitations and execute commands or file searches autonomously. While no explicitly malicious payloads (like data exfiltration) are present, these instructions significantly increase the risk of unintended system impact or unauthorized actions by bypassing standard human-in-the-loop safety checks.