The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill generates a central project brief (brief.md) based on user responses during an adaptive discovery phase. This brief is subsequently read by multiple agent profiles, including the 'Director' and various specialized 'Renderers', creating a vulnerability surface where malicious instructions in the brief could influence agent behavior.
Ingestion points: User-provided descriptions and requirements are collected in SKILL.md and processed by scripts/bootstrap_pipeline.py to populate assets/brief.md.tmpl.
Boundary markers: While the brief is structured using Markdown headers, it lacks explicit delimiters or system-level instructions to agents to ignore or sanitize potentially malicious instructions embedded within user-supplied text fields.
Capability inventory: Orchestrated agents are granted terminal, file, and kanban toolsets, enabling them to execute shell commands (e.g., via ffmpeg or remotion), manage local files, and create further tasks within the pipeline.
Sanitization: The bootstrap script performs basic string replacement for template interpolation. The resulting setup.sh script uses single-quoted heredocs to prevent shell-level injection during file creation, but no specific LLM-focused sanitization is applied to the brief's content.
[COMMAND_EXECUTION]: The skill generates and facilitates the execution of a local setup.sh script. This script automates project environment preparation, including the secure patching of Hermes profile configurations. The configuration logic uses Python's yaml.safe_load to prevent unsafe deserialization during the update of toolsets and skill lists.
[DYNAMIC_EXECUTION]: The skill dynamically generates a bash script (setup.sh) from pre-defined local templates. This script is used to bootstrap the multi-agent environment and includes embedded Python one-liners to handle structured configuration updates, representing a controlled use of dynamic script generation to support orchestration.

kanban-video-orchestrator