The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes architecture for a developer portal (Backstage) that ingests data from external sources, which is a known vector for indirect prompt injection.
Ingestion points: references/internal-developer-portal.md outlines how the portal fetches service metadata (catalog-info.yaml) from external GitHub repositories.
Boundary markers: The provided configuration examples do not include explicit delimiters or instructions for the agent to disregard instructions potentially embedded within the fetched external data.
Capability inventory: SKILL.md and references/internal-developer-portal.md describe workflows where the agent can initiate infrastructure provisioning, repository creation, and database setup via custom actions and templates.
Sanitization: No sanitization or validation logic is specified for the ingested metadata in the illustrative examples.
[METADATA_POISONING]: The skill folder includes eight documentation files (e.g., references/architecture-patterns.md, references/security-best-practices.md) that are not listed in the primary SKILL.md reference section and consist of approximately 150 repeated sections of filler text.
Evidence: Files such as references/deployment-pipelines.md repeat the same generic technical paragraph hundreds of times. This redundant content appears to be low-quality bloat that could exhaust an agent's context window or mislead the agent regarding its own knowledge base depth.

devops-platform-engineering