The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow for ingesting and processing external datasets for evaluation purposes, which presents a surface for indirect prompt injection if the datasets contain malicious instructions.
Ingestion points: The skill reads datasets from JSON files located in src/ai/<box>/eval/datasets/ as described in SKILL.md and references/templates.md.
Boundary markers: The skill suggests using schemas for validation (e.g., input.schema in references/templates.md) but does not implement explicit delimiters or warnings to ignore instructions within the dataset content.
Capability inventory: The skill utilizes subprocess execution to run evaluations (genkit eval:flow), tests (vitest), and project configurations (tsx) as seen in SKILL.md and references/templates.md.
Sanitization: The skill relies on Genkit's internal handling and the use of Zod schemas mentioned in the 'Layout recomendado' section of SKILL.md for structure validation.
[DYNAMIC_EXECUTION]: The skill generates and executes project-specific TypeScript code and configuration files. It uses tsx to run the Genkit entry point and vitest for unit testing the logic. This is standard developer behavior for the intended use-case but involves the runtime execution of dynamically identified files.
[COMMAND_EXECUTION]: The skill provides instructions for executing various Genkit CLI commands such as genkit eval:flow, genkit start, and genkit eval:run. These commands interact with the local file system and potentially external APIs depending on the project configuration.
[EXTERNAL_DOWNLOADS]: The skill references documentation and official configuration schemas from trusted and well-known service domains including genkit.dev, google.com, and microsoft.com (via learn.microsoft.com). These references are documented neutrally as they point to official project resources.

genkit-evals