VSS Video Analysis Reports
Generates structured reports from recorded video clips or incident time ranges.
The skill clearly separates internal VLM URLs from browser-facing links, enforces read-only behavior for Mode B, requires deployment confirmation, and forbids fabrication, reducing external-impact risk. Points are deducted because video sensitivity, credential scope, remote-VLM data flow, and access control for report links are not sufficiently specified; deployment handoff may also cause external changes.
Routing, probes, backend selection, live-config loading, malformed-response handling, and empty-result behavior are described in useful detail, with required failure reporting. Static review did not execute key paths, and the workflow depends on multiple external skills, services, and runtime assumptions, so the score remains limited by the static cap.
Audience, Mode A/Mode B scenarios, positive examples, negative triggers, output templates, and deployment profiles are clearly defined. Points are deducted because compatibility with non-NVIDIA VSS environments, Chinese-language use, and mainland-China network reachability are not addressed, while semantic boundaries depend on other skills.
The documentation has good structure and includes version 3.2.0, authorship, Apache-2.0, templates, limitations, examples, cross-references, and error handling; repository context describes synchronization and update flow. Points are deducted for no skill-specific changelog, limited explicit maintenance ownership, incomplete installation/credential guidance, and governance uncertainty because the supplied license metadata is NOASSERTION despite file-level license statements.
The skill gives explicit routing, fixed report titles and fields, directly usable Markdown templates, and policies for empty results and failures, so the core value proposition is clear. Points are deducted because no representative final report is independently verifiable, the supplied evaluation covers only 3 tasks, and Codex correctness and effectiveness were 23% and 34%; static rules cap this dimension at 7.
The files provide concrete commands, endpoints, templates, and an external evaluation summary, offering limited auditability. Points are deducted because evaluation coverage is only 3 positive tasks with no negative tasks, there is no committed test suite or CI covering key paths, and the conclusions cannot be independently reproduced from the files; therefore the score stays within the static cap of 5.
- Before execution, verify VSS, VST, VLM, and VA-MCP endpoints and video-URL reachability so internal clip URLs are not exposed to remote models or end users.
- Apply organizational access control, log redaction, and retention policies to camera video and incident reports; the skill does not define them.
- Confirm authorization for deployment handoff after probe failure because the handoff may change the runtime environment.
- Add tests for Chinese requests, empty or malformed fields, remote VLM inability to fetch internal URLs, and differing network environments.
What does this skill do, and when should you use it?
This skill is designed for the NVIDIA Video Search and Summarization Blueprint and focuses specifically on report generation. It routes requests to either clip analysis through a VLM or incident-range analysis through video analytics. It validates the relevant deployment profile, handles report-facing clip URLs, and renders a prescribed Markdown report. It is not intended for standalone summarization, real-time alerts, or ad-hoc visual questions.
The skill selects Mode A or Mode B from the request. Mode A resolves a sensor recording and clip URL, probes NIM Cosmos or RT-VLM through an OpenAI-compatible API, sends a video_url content block to chat/completions, and fills the Video Analysis Report template. Mode B resolves an ISO 8601 UTC range and optional sensor, performs a read-only video_analytics__get_incidents request through /vss-query-analytics, groups incidents by sensor or category, tallies verdicts, and fills the Incident Range Report template. Clip URLs shown in reports are rewritten to the configured public host and port when available.
- VSS operators need a timestamped analysis report for a recorded sensor video.
- Security or operations teams need a report of alerts and incidents within a defined time range.
- A team with the VSS base profile deployed needs VLM-based analysis of one recorded clip.
- A team with the VSS alerts profile deployed needs a read-only incident report with verdicts and reasoning.
What are this skill's strengths and limitations?
- Defines explicit routing between clip reports and incident-range reports.
- Enforces a read-only boundary for Mode B analytics retrieval.
- Supports both NIM Cosmos and RT-VLM backend selection with readiness checks.
- Specifies report titles, required fields, and browser-facing URL handling.
- Requires deployed and reachable VSS, VST, VLM or RT-VLM, and possibly VA-MCP services.
- The source does not provide a complete deployment tutorial, test results, or broad platform evidence.
- A browser-playable clip link cannot be produced when the required public host or port is missing.
- It does not cover ad-hoc visual Q&A, archive search, real-time alert management, or deployment changes.
How do you install this skill?
Install the named skill with the skills CLI:
npx skills add nvidia/skills --skill vss-generate-video-report --yes
The README says the CLI prompts for an installation destination. You may also target a supported client, such as Codex or Claude Code, with --agent codex or --agent claude-code.
How do you use this skill?
After installation, ask the loaded agent for a report, for example:
- "Generate a report for this video"
- "Report on incidents from 12:31Z to 12:32Z"
Single-clip or sensor-video requests generally route to Mode A. Requests naming a time range, incidents, or alerts route to Mode B. If both a sensor and a time range are mentioned, the skill asks for clarification.
How does this skill compare with similar options?
Compared with vss-ask-video, this skill renders structured reports rather than answering ad-hoc visual questions. Compared with vss-summarize-video, it additionally handles routing, deployment checks, URL rewriting, and report-template filling. Compared with vss-query-analytics, it is intended to render incident reports rather than perform a standalone read-only analytics lookup.