The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains eight unreferenced files in the references/ directory (e.g., architecture-patterns.md, testing-strategies.md) that contain over 1,200 repetitive, AI-generated technical sections. This 'context stuffing' is a common adversarial technique used to overwhelm an agent's context window and bypass system instructions or safety guardrails.
[PROMPT_INJECTION]: The skill is designed to ingest and process untrusted SQL code for migrations and linting, creating a surface for indirect prompt injection.
Ingestion points: transform/models/**/*.sql and migrations/*.sql (SKILL.md).
Boundary markers: None present in example configurations to distinguish instructions from data.
Capability inventory: Shell access for package management, database operations via dbt and snowsql, and cloud storage synchronization via aws s3 (SKILL.md, references/sql-linting-and-testing.md).
Sanitization: No mechanisms for filtering or escaping embedded Jinja or SQL commands are described.
[EXTERNAL_DOWNLOADS]: Several CI/CD workflow examples and reference documents (e.g., references/dbt-ci-cd-pipeline.md) rely on pip install -r transform/requirements.txt. Since the requirements.txt file is not included in the skill package, its contents are unverifiable and pose a supply chain risk.

data-pipeline-cicd