DeepStream Pipeline Builder
Collects pipeline requirements interactively, then generates and validates runnable GStreamer inference pipelines for NVIDIA DeepStream.
The security notes specify no shell=True, element-name validation, shlex tokenization, input-size limits, and no credential handling; the workflow also requires explicit selection of major pipeline parameters and only writes scripts when explicitly requested. Deductions apply because it executes local gst-inspect/gst-launch and ls, may access user media, RTSP, or UDP endpoints, and lacks per-operation confirmation, isolation, and rollback guidance for sensitive media and network effects.
The scripts, retriever, validator, error recovery, and two-retry limit are broadly self-consistent, with intelligible recovery paths for failures. Deduction applies because key paths were not reproduced in this static review; multi-stream live parsing is skipped, and configuration, plugin, and platform differences may still make generated pipelines unrunnable. The score therefore remains below the static ceiling of 10.
The skill defines audiences, trigger phrases, input types, inference, tracking, sinks, platforms, extras, and several non-fit boundaries such as Linux and local DeepStream dependencies. Deductions apply because Chinese-language interaction and mainland-China environment fit are not addressed, triggering relies mainly on English phrases, and advanced configuration and custom-pipeline boundaries remain limited.
The documentation includes prerequisites, examples, parameter tables, staged workflow, assembly rules, security limitations, scripts, version, and license metadata. Deductions apply because maintenance ownership and update procedures are not sufficiently explicit, no formal changelog is provided, and the benchmark reports missing recommended Examples and author metadata plus root-structure schema warnings.
The skill clearly targets a directly copyable single-line gst-launch-1.0 command and adds retrieval, assembly, validation, and refinement guidance, giving it meaningful practical value. Deductions apply because default media and model paths usually require replacement, DeepStream may be unavailable, representative outputs were not verified here, and unusual requests may require substantial manual work; static calibration limits this score to 7 or below.
The supplied source includes implementation, dataset retrieval logic, test descriptions, benchmark material, and security documentation, providing a useful audit trail. Deductions apply because this review did not execute the tests, and the benchmark explicitly lacks Tier 3 execution evidence or independent corroboration; the score therefore remains below the static ceiling of 5.
- Generated commands depend on local DeepStream, GStreamer, plugins, model configuration, and default paths; replace and verify paths before running.
- Multi-stream pipelines skip live-parse validation, so validation status is not equivalent to full runtime assurance.
- The skill can generate RTSP inputs and UDP/RTSP outputs but does not provide per-operation network confirmation, sensitive-media handling, or rollback guidance.
- Chinese-language interaction and mainland-China environment compatibility are not documented.
What does this skill do, and when should you use it?
This skill is for developers who need NVIDIA DeepStream and GStreamer pipelines for video, image, or streaming inference. It gathers input, inference, tracking, output, platform, and transform requirements through an interactive questionnaire, then retrieves relevant examples from more than 270 verified pipelines. It assembles a gst-launch-1.0 command and validates syntax, elements, properties, and structural linking. It requires Python 3.8+, the DeepStream SDK, GStreamer, and a supported NVIDIA dGPU or Jetson/SBSA platform.
Extracts pipeline requirements from a user request; asks for unspecified input, source-count, inference, tracker, sink, platform, and extra-transform parameters; runs a Python standard-library BM25 retrieval script; boosts results using structural metadata; assembles a single-line gst-launch-1.0 pipeline; validates syntax, DeepStream elements, properties, and linking; and reports missing default sample paths when the pre-flight check cannot find them.
- A DeepStream developer needs primary object detection on one local video with display output.
- A video analytics engineer needs several RTSP sources with primary and secondary inference plus NvDCF tracking.
- A Jetson developer needs an inference pipeline for a USB camera or local image.
- An engineer needs to save inference output as JPG, PNG, MP4, or H264.
- A developer needs resize, rotation, cropping, or color-format conversion in the pipeline.
What are this skill's strengths and limitations?
- Supports local video, local images, RTSP, USB cameras, and test patterns.
- Covers primary and secondary inference, preprocessing, Triton, several trackers, and multiple sinks.
- Uses BM25 retrieval with structural metadata and no external Python dependencies.
- Includes retrieval, validation, and data-quality scripts.
- Requires NVIDIA DeepStream SDK, GStreamer, and a supported NVIDIA dGPU, Jetson, or SBSA platform.
- Depends on local scripts, datasets, configuration files, and sample paths.
- Live-parse validation is skipped for multi-stream pipelines.
- The source describes a test suite but does not provide actual test-run results for this skill.
How do you install this skill?
Install the specific skill with the repository's documented skills CLI command: npx skills add nvidia/skills --skill deepstream-generate-pipeline --yes. You can target an agent, for example: npx skills add nvidia/skills --skill deepstream-generate-pipeline --agent codex. The skill becomes available when the agent next loads skills and encounters a relevant task.
How do you use this skill?
Ask the agent for a specific or underspecified DeepStream/GStreamer pipeline, such as “give me a pipeline to infer on an image” or “detect and track on 4 rtsp streams and display on jetson.” The skill asks for missing requirements, retrieves examples, assembles a pipeline, and validates it. Running the result still requires DeepStream, GStreamer, input files, and model configuration in the local environment; the source does not specify a separate installation path for the skill's reference files.