The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes external, untrusted data (files) and interpolates a user-provided 'goal' into a system prompt without robust delimiters or sanitization.
Ingestion points: The --file and --goal arguments in scripts/look_at.py are the entry points for untrusted content.
Boundary markers: The prompt template in scripts/look_at.py uses simple labels like 'Goal:' but lacks cryptographically secure delimiters or explicit 'ignore instructions' warnings for the file content.
Capability inventory: The script can read any file accessible to the user and perform network requests to Google's API.
Sanitization: No input validation, escaping, or filtering is performed on the file content or the goal string before being sent to the LLM.
[SAFE]: The skill uses the official google-genai library to interact with Google's Gemini API, which is a well-known and trusted service. This network activity is fundamental to the skill's primary purpose.
[SAFE]: The instructions in SKILL.md for discovering and executing the script via the shell are standard for the intended environment and do not involve suspicious privilege escalation or hidden commands.

look-at