The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides robust defenses against prompt injection. It includes a 12-layer security framework in reference/agentic-guardrails-security.md with specific regex patterns for detection and instructions for techniques like the 'Input Sandwich' and 'XML Delimiter' in reference/agentic-prompt-engineering.md to prevent user input from overriding system instructions.
[CREDENTIALS_UNSAFE]: The skill demonstrates secure secret management. reference/agentic-config-project.md and reference/agentic-deployment.md provide templates for using .env files and pydantic-settings to manage API keys securely. It explicitly warns against logging secrets or hardcoding them in source code.
[EXTERNAL_DOWNLOADS]: The skill references standard, well-known libraries and services. It suggests using uv to manage dependencies from official registries (PyPI) and provides a Dockerfile that fetches the uv binary from a trusted source (ghcr.io/astral-sh/uv). All referenced domains are official documentation or well-known service providers.
[COMMAND_EXECUTION]: The skill avoids unsafe command execution. In reference/agentic-templates-tools.md, it specifically instructs users to avoid eval() and use ast.literal_eval for safe mathematical expression evaluation. It also provides patterns for parameterized database queries to prevent injection attacks.
[DATA_EXFILTRATION]: No exfiltration patterns were found. The skill includes PII (Personally Identifiable Information) detection and redaction patterns in reference/agentic-guardrails-security.md to ensure sensitive data is masked before logging or storage.

agentic-ai-dev