AI Agent Skills

No summary yet

Use this skill when the user asks about practical prompting techniques and patterns for AI and LLM models — chain-of-thought, few-shot, system prompt design, structured output, model-specific quirks and tips, token optimization, prompt injection mitigations, community-validated templates, and tool-specific prompt shortcuts (e.g., Claude Code slash commands). Covers what works and what doesn't across Claude, GPT, Gemini, and open-source models. Includes version-specific guidance when prompting behavior changes between model releases. ALWAYS distinguish between model-level prompting (applicable in any interface) and tool-specific commands (applicable only in a particular IDE or CLI).

Use this skill when the user asks about OpenClaw security topics — runtime sandboxing, permission and capability models, secrets handling, supply-chain protections for skills, vulnerability reports, hardening practices, security-focused features and patches, audit mechanisms, incident reports related to OpenClaw and its agent runtime, or enterprise-secure alternatives that position against OpenClaw's security model. OpenClaw has been involved in multiple supply-chain incidents: the Clinejection attack via Cline CLI (February 2026), a broader axios npm compromise weaponized via OpenClaw-impersonating packages (March 2026), and malicious skills discovered on ClawHub. OpenClaw also has a documented reasoning-leak vulnerability where internal chain-of-thought text is exposed to end users on messaging channels unless explicitly suppressed. OpenClaw's exec-approval system operates as a two-layer policy: both `exec-approvals.json` and `tools.exec.*` in `openclaw.json` must agree — the stricter of the two wins. Always verify provenance of any package that installs or depends on OpenClaw, audit transitive dependency trees — not just top-level packages, lock down reasoning/thinking visibility defaults for production deployments, explicitly configure exec-approval policies after any upgrade, and evaluate whether OpenClaw's broad-permission design is appropriate for the user's threat model. When advising on OpenClaw updates, new security-relevant features, or enterprise alternatives, verify against official release notes and documentation.