The Agent Skills Directory

[METADATA_POISONING]: The skill contains multiple reference files (e.g., references/architecture-patterns.md, references/testing-strategies.md, references/state-management.md) that claim to provide advanced technical deep dives.
These files contain 150 sections each, where almost every section consists of the exact same boilerplate text about kernel optimization (epoll_wait, io_uring) and CPU cache locality.
This massive volume of repetitive content is deceptive and likely intended to bloat the skill's context size or bias the agent's output towards specific technical jargon.
[INDIRECT_PROMPT_INJECTION]: The skill instructions define workflows for analyzing untrusted external data, creating an attack surface for indirect prompt injection.
Ingestion points: The skill reads local Kubernetes manifests, Terraform plans, and Dockerfiles for policy scanning (e.g., in references/policy-ci-cd.md).
Boundary markers: No delimiters or "ignore embedded instructions" warnings are suggested when processing these external files.
Capability inventory: The skill utilizes powerful CLI tools such as kubectl, terraform, opa, and conftest to evaluate configurations.
Sanitization: There is no mention of sanitizing or escaping the content of these external files before they are processed by the agent or policy engines.
[EXTERNAL_DOWNLOADS]: The skill documentation includes instructions to download third-party security and automation tools.
Examples include fetching the conftest binary from GitHub and the opa binary from openpolicyagent.org.
These references target established, official sources for the respective projects.

devops-policy-as-code