The Agent Skills Directory

[SAFE]: The skill mandates the use of isolated disposable workspaces and private pipes for Blender execution, preventing cross-task data contamination.
[SAFE]: It implements strict data exfiltration preventions by prohibiting the transfer of scene files, code, paths, or private measurements to external helpers, instead requiring a 'fixed vocabulary' of symptoms.
[INDIRECT_PROMPT_INJECTION]: While the skill ingests external geometry and task specifications (ingestion point in SKILL.md), it mitigates injection risks by piping data through 'closed JSON schemas' and enforcing a 'fixed vocabulary' (sanitization/boundary markers). The capabilities are restricted to exact, caller-owned read-only probe IDs (capability inventory).

jev-blender