senior-prompt-engineer

The skill consists of markdown documentation and three Python scripts: agent_orchestrator.py, prompt_optimizer.py, and rag_evaluator.py.

1. SKILL.md, references/agentic_system_design.md, references/llm_evaluation_frameworks.md, references/prompt_engineering_patterns.md: These are documentation files. They describe various prompt engineering patterns, LLM evaluation frameworks, and agentic system designs. While they contain example prompts and pseudocode, none of these examples are malicious or attempt to inject prompts into the AI's own behavior. They serve an educational purpose. No obfuscation, data exfiltration, privilege escalation, or persistence mechanisms were found.

2. _meta.json: This file contains standard metadata. The commit URL points to a trusted GitHub organization (github.com/openclaw/skills). This is noted as an informational reference but does not pose a security risk.

3. scripts/agent_orchestrator.py: This Python script is designed to parse agent configurations, validate tool registrations, visualize workflows, and estimate token usage. It uses only standard Python libraries (argparse, json, re, sys, pathlib, typing, dataclasses, enum) and includes a custom, simple YAML parser. It reads local configuration files but does not perform any network requests, access sensitive file paths (e.g., ~/.aws/credentials), or execute arbitrary commands. It is a local analysis tool.

4. scripts/prompt_optimizer.py: This Python script analyzes prompts for token efficiency, clarity, and structure. It estimates token counts and costs, identifies issues, and suggests optimizations. It uses only standard Python libraries. It reads local prompt files but does not perform network requests, access sensitive file paths, or execute arbitrary commands. It is a local analysis tool.

5. scripts/rag_evaluator.py: This Python script evaluates Retrieval-Augmented Generation (RAG) systems by measuring context relevance, answer faithfulness, and retrieval metrics. It uses only standard Python libraries. It reads local JSON files containing contexts and questions but does not perform network requests, access sensitive file paths, or execute arbitrary commands. It is a local analysis tool.

Overall Assessment: All files were thoroughly reviewed for prompt injection, data exfiltration, obfuscation, unverifiable dependencies, privilege escalation, persistence mechanisms, metadata poisoning, indirect prompt injection, and time-delayed attacks. No such threats were identified. The Python scripts are local, self-contained analysis tools that do not interact with external systems or the agent environment in a dangerous manner. The skill's purpose is educational and analytical, and its implementation reflects a safe design.