The Agent Skills Directory

[SAFE]: The skill provides normative guidelines and best practices for prompt engineering. It does not include any executable scripts, binary files, or suspicious command-line instructions.
[INDIRECT_PROMPT_INJECTION]: The skill instructions acknowledge the risk of processing untrusted data and provide defensive guidelines. 1. Ingestion points: Instructions for processing diverse AI-facing artifacts like plans, skills, and memory (SKILL.md, Section 20). 2. Boundary markers: Recommends the use of semantic delimiters and tags to separate instructions from data (references/input-representation.md, Section 12). 3. Capability inventory: No code execution or network tools are included in the skill; it provides guidance for external tool design (references/tools-and-security.md, Section 15). 4. Sanitization: Explicitly mentions redaction and egress controls for personal or confidential data (references/tools-and-security.md, Section 16).
[CREDENTIALS_UNSAFE]: The skill mandates that secrets and personal data must never be placed in prompts, examples, logs, or generated artifacts (references/tools-and-security.md, Section 16).
[REMOTE_CODE_EXECUTION]: Includes supply-chain security guidelines, advising that instruction and skill packages be treated as untrusted inputs requiring review and pinning (references/tools-and-security.md, Section 16).
[DATA_EXFILTRATION]: Recommends controls such as least-privilege credentials and sandboxed execution to prevent unauthorized data movement (references/tools-and-security.md, Section 16).

prompting