The Agent Skills Directory

[SAFE]: The skill instructions focus on aligning agent behavior with user intent and objective evidence. It explicitly warns agents against 'inventing the fact you were sent to retrieve,' which serves as a defensive measure against hallucinating sensitive data like API contracts, prices, or legal text.
[COMMAND_EXECUTION]: The skill provides examples of Python scripts (e.g., snake_rename.py) that perform file system operations like os.rename. However, these are strictly pedagogical examples. Furthermore, the skill's core logic (detailed in references/constraint-classes.md) mandates that any destructive, irreversible, or bulk action must use a 'preview' mode by default, requiring an explicit flag (e.g., --apply) or user confirmation for execution. This is a robust security pattern for preventing accidental data loss.
[DATA_EXPOSURE]: No hardcoded credentials, API keys, or sensitive file paths were found. The skill emphasizes retrieving facts from verified sources and labeling artifacts as 'drafts' until verified, reducing the risk of a user acting on unverified or sensitive-looking information.
[REMOTE_CODE_EXECUTION]: The skill does not contain any patterns for downloading or executing external scripts at runtime. All provided scripts are local examples using standard libraries (os, re, sys, math).
[PROMPT_INJECTION]: The instructions do not contain any patterns typical of prompt injection or jailbreaking. Instead, they provide structured guidance on how to classify user constraints and when to ask for clarification, which improves the agent's robustness against ambiguous or contradictory requests.
[PRIVILEGE_ESCALATION]: No commands for escalating privileges (e.g., sudo, chmod) or modifying system-level configurations were detected.

starting-point