The Agent Skills Directory

[SAFE]: The skill interacts exclusively with official Google API domains (generativelanguage.googleapis.com) for generative tasks and layout optimization.
[SAFE]: Credential management follows best practices by instructing the user to utilize environment variables (GEMINI_API_KEY) rather than hardcoding sensitive tokens within the code.
[COMMAND_EXECUTION]: The skill utilizes local Bash and Python scripts to orchestrate the iterative workflow and process JSON data from API responses. These operations are confined to the skill's defined directory and do not execute untrusted remote scripts.
[PROMPT_INJECTION]: While the skill ingests user input via $ARGUMENTS to generate prompts, it employs a highly structured template to guide the model's behavior, minimizing the risk of prompt-based attacks while fulfilling its primary functional purpose.
Ingestion points: User-provided descriptions are captured in $ARGUMENTS within SKILL.md and interpolated into templates.
Boundary markers: The workflow uses a multi-stage architectural design with Claude acting as a supervisor to validate output against a fixed set of visual standards.
Capability inventory: The skill has access to Bash, Write, and curl for file management and API interaction.
Sanitization: Python's json.dump is used to sanitize content before it is included in API payloads, preventing shell-level injection.

aris-paper-illustration