Skill-Eval

Autonomous engine that systematically evaluates and ranks agent skills across models using rubric grading, error taxonomy, and improvement feedback loops.

This skill contains a detailed engine design but includes no code, no install, and no declared credentials—yet its instructions assume running scripts, writing files, and calling multiple external LLM providers. Before installing or enabling it: 1) Ask the publisher for the missing artifacts (scripts, eval configs) and for a clear list of required API keys and network endpoints. 2) Confirm where it will write outputs and whether it will read other skills or registries; restrict its filesystem scope in a sandbox if possible. 3) Require explicit declarations of any model provider credentials (OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.) and limit which credentials it may use. 4) If you must run it, do so in an isolated environment with restricted network and filesystem access and audit its operations and generated files. The source is unknown—treat it as untrusted until those gaps are resolved.

Type: OpenClaw Skill Name: skill-eval Version: 0.4.0 The `skill-eval` bundle describes an autonomous system for evaluating and improving AI agent skills. It features high-risk capabilities, including spawning subagents for comparative testing (Phase 3) and a 'self-evolving' mechanism (Phase 9 and 12) that directs the agent to modify its own instructions in `SKILL.md` based on evaluation results. While these functions are consistent with the stated purpose of a benchmarking engine, the combination of self-modification and sub-agent execution constitutes a high-risk operational profile. No evidence of intentional malice, such as data exfiltration or unauthorized access, was identified.