The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the auto_label command-line tool within a TAO Toolkit container to orchestrate the video annotation pipeline, as defined in references/skill_info.yaml.
[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface due to its multi-step architecture where the output of one model (e.g., video captions) is used as input for subsequent model prompts (e.g., QA generation).
Ingestion points: Video files are processed by VLMs in Step 1, and the resulting JSONL caption files are ingested as text in Steps 2 and 3.
Boundary markers: The skill uses structured prompts with clear delimiters such as <example>, [Global Caption], and {step_2_output} to separate instructions from data.
Capability inventory: The pipeline has the capability to execute local shell commands (auto_label), perform file I/O within the results directory, and initiate network requests to configured VLM/LLM API endpoints.
Sanitization: There is no evidence of explicit sanitization or filtering of model-generated text before it is interpolated into downstream prompts, which is typical for distillation pipelines.
[EXTERNAL_DOWNLOADS]: The skill facilitates communication with external VLM and LLM services, including Google Gemini and OpenAI-compatible endpoints, to process video and text data. These connections are standard for the tool's primary purpose and target well-known or user-specified services.
[SAFE]: The documentation mentions system-level requirements such as ffmpeg and nvidia-container-toolkit, which are standard dependencies for video processing and GPU acceleration in this domain.

tao-generate-video-reasoning-annotations