The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill requires the agent to write and modify Python scripts (build_previs.py) which are then executed using a Blender binary. This dynamic script generation and execution pattern is the core functionality of the skill but represents a significant security surface.
Evidence found in SKILL.md (Step 3: "Write or adapt this plan's Python") and references/blender-authoring.md (instructions for preflighting the Blender interpreter and modules).
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, including director feedback, character/location sheets, and scene JSON files. This content is interpolated into revision descriptions and used to generate the parameters for the Python scripts, creating a surface for indirect injection attacks.
Ingestion points: scene.json, director feedback (NL), and character sheets (directing-workflow.md).
Boundary markers: None explicitly mentioned for prompt interpolation, though the skill mandates a "model-neutral description" review.
Capability inventory: The agent has the ability to write files (build_previs.py), execute shell commands (renku, blender, ffmpeg), and access the local filesystem.
Sanitization: Instructions recommend an "agent editorial review" of descriptions but do not specify technical sanitization for script parameters.
[COMMAND_EXECUTION]: The skill frequently invokes external command-line interfaces and binaries, including a custom renku CLI, blender, ffmpeg, and ffprobe.
references/blender-authoring.md instructs the agent to preflight these binaries and handle output rendering/encoding via subprocess calls.
SKILL.md and references/directing-workflow.md contain specific instructions for using the renku CLI tool, including warnings against inventing syntax or bypassing permission denials.

blender-shot-planner