The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill supports dynamic loading of Python modules through the prompts_module configuration field.
Evidence:
The SKILL.md and references/configuration.md files describe using the prompts_module parameter to point to custom Python modules (e.g., prompts_module: "my_package.prompts_traffic").
references/configuration.md explicitly notes that the dotted path must resolve and be present on the PYTHONPATH, enabling the execution of arbitrary Python code available in the environment.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a significant surface for indirect prompt injection as it ingests and processes untrusted external data (videos and JSONL metadata) which is then used as context for VLM and LLM reasoning steps.
Ingestion points:
The skill reads video files from a directory specified by video_root.
It ingests metadata from JSONL files provided via input_jsonl_files.
Boundary markers:
Prompt templates in references/prompts_traffic.py and references/prompts_warehouse.py use structured tags such as <example>, [Global Caption], and [Detailed Video Description] to delimit data, which provides some mitigation against accidental instruction following.
Capability inventory:
The skill uses the Bash tool to execute the auto_label command, which orchestrates the pipeline (Step 0-4).
The pipeline includes steps that generate captions, descriptions, and QA pairs using external model endpoints (Gemini/OpenAI).
Sanitization:
There is no evidence of sanitization or filtering of the textual content within the input JSONL files or video metadata before it is interpolated into model prompts.

tao-generate-video-reasoning-annotations