The Agent Skills Directory

[SAFE]: The skill demonstrates a high security posture by requiring explicit user authorization for external mutations, paid calls, and changing persistent sources. It includes quality gates that mandate the inspection of repository instructions and the use of the smallest compatible changes to preserve public contracts.
[INDIRECT_PROMPT_INJECTION]: The skill processes codebase content and external data from SandBase tools, creating a surface for indirect prompt injection.
Ingestion points: Existing codebase files and external documentation retrieved through sandbase_discover and sandbase_inspect (referenced in SKILL.md and references/sandbase-api-map.md).
Boundary markers: The workflow requires restating requested outcomes and acceptance checks before implementation to define operational boundaries.
Capability inventory: The skill can implement code changes, run repository tests, and invoke external capabilities via the sandbase_run tool.
Sanitization: The skill relies on verification steps—such as focused tests and source comparisons—and mandatory user confirmation before executing external capabilities to mitigate the risk of untrusted data influencing agent actions.

mcp-cli