The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions depend on a directory structure external to the skill itself, executing setup scripts (e.g., bash "$KIT/setup.sh") and referencing tools from sibling skill folders (e.g., ../stick-figure/engine/tools/tts_kokoro.py). The setup script is indicated to download large assets (approximately 1 GB) for voice models and dependencies like Chromium.
[COMMAND_EXECUTION]: The skill relies on extensive shell command execution for its primary workflow, including project filesystem management (cp, ln, mkdir), rendering control via Node.js/Playwright, and media assembly using ffmpeg and python3 subprocesses.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied narrative scripts to generate video content. This untrusted input is used for caption generation and text-to-speech. While the rendering pipeline uses safe update methods (such as textContent in the DOM) to prevent script execution within the rendering browser, the ingestion of arbitrary text combined with the skill's extensive local execution capabilities represents an attack surface for indirect prompt injection.
Ingestion points: User-provided script content written to script.json and interpolated into CAPTIONS in shots.js via captions.py.
Boundary markers: None explicitly defined to ignore embedded instructions in the script data.
Capability inventory: Subprocess execution of ffmpeg, node, and python3 for media processing and rendering; filesystem operations for project setup.
Sanitization: captions.py escapes single quotes for JavaScript string literals, and the frontend rendering engine in main.js utilizes textContent for caption display, preventing DOM-based injection.

zach-d-films