The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Automated security scanners have identified the documentation URL https://jeffallan.github.io/claude-skills/skills/data-ml/prompt-engineer/ as malicious. This URL is a primary external link for documentation within the SKILL.md file.
[PROMPT_INJECTION]: The skill’s reference documentation, specifically references/system-prompts.md and references/evaluation-frameworks.md, contains explicit prompt injection payloads. These include jailbreak patterns like 'DAN' (Do Anything Now), system prompt extraction attempts, and instructions to 'Ignore all previous instructions'. Although these are presented as educational test cases for evaluating LLM robustness, they represent active injection vectors.
[INDIRECT_PROMPT_INJECTION]: The skill creates a significant surface for indirect prompt injection by design, as it facilitates the optimization and evaluation of prompts that incorporate untrusted data.
Ingestion points: The skill uses placeholders such as {review}, {document}, {article}, {transcript}, and {code} to interpolate external, potentially attacker-controlled content into prompts.
Boundary markers: Documentation in references/context-management.md recommends using XML-style tags to separate data from instructions; however, these markers provide limited protection against sophisticated adversarial inputs.
Capability inventory: The skill provides templates for 'Tool Use' and ReAct reasoning loops, which allow an LLM to invoke external tools and functions based on its internal logic.
Sanitization: The skill advocates for specific mitigations such as instruction hierarchy and canary tokens to identify and block injection attempts at runtime.
[REMOTE_CODE_EXECUTION]: The reference guide for 'ReAct' patterns in references/prompt-patterns.md describes an iterative reasoning loop where the agent takes actions using external tools. If the tools available to the agent include capabilities for system command execution or sensitive network requests, this pattern establishes a path for remote code execution.
[COMMAND_EXECUTION]: The evaluation framework and optimization references include Python and TypeScript code snippets that demonstrate how to run tests and analyze outputs. Depending on the environment, these snippets may execute code or call subprocesses, posing a risk if the inputs being evaluated are not strictly sanitized.

prompt-engineer