Skills for working with the NVIDIA Blueprint for Video Search & Summarization (VSS) — a suite of GPU-accelerated microservices for building vision agents and video-analytics applications. Each subdirectory under skills/ is a self-contained skill following the agentskills.io specification, with name, description, version, and license declared in its SKILL.md frontmatter.

New here? Read Orientation first, then Which skill do I need?. Those two sections are the fastest path from a user request to the right skill.

These skills are a developer-side tool: a coding agent (Claude Code, Codex, NemoClaw, or any agentskills.io host) loads them to deploy and operate a VSS deployment from natural language. They are distinct from the in-product VSS Agent, which runs inside a deployed VSS workflow and orchestrates microservices to answer end-user questions. The skills here drive VSS; the VSS Agent is one of the things they can drive.

VSS-based deployments are multi-layer systems. Most skills map to exactly one layer, so knowing the layers tells you which skill to reach for. Video flows top-to-bottom: raw frames are turned into features in real time, those features are published to a message bus, analyzed downstream, placed in storage or accessed by an agent to generate results (i.e. summaries, search, etc).

                         ┌──────────────────────────────────────────────┐
  VIDEO IN (files,       │  1. REAL-TIME VIDEO INTELLIGENCE             │
  RTSP, live streams) ──▶│     Extract features from video in real time │
                         │     • RT-CV     detection & tracking (2D/3D) │
                         │     • RT-Embed  semantic video embeddings    │
                         │     • RT-VLM    captions / incident detection│
                         └────────────────────────┬─────────────────────┘
                                                  │ features → message broker / DB
                         ┌────────────────────────▼─────────────────────┐
                         │  2. DOWNSTREAM ANALYTICS                     │
                         │     Turn features into insights & alerts     │
                         │     • Behavior Analytics  (events, incidents)│
                         │     • Alert Contextualization                |
                         └────────────────────────┬─────────────────────┘
                                                  │ incidents, metrics → Elasticsearch
                         ┌────────────────────────▼─────────────────────┐
                         │  3. AGENT & OFFLINE PROCESSING               │
                         │     Reason over results for users            │
                         │     • Search, Summarize, Ask, Report         │
                         │     • Query analytics  (via VA-MCP)          │
                         └──────────────────────────────────────────────┘

  MIDDLEWARE (cross-cutting): Video IO & Storage (VIOS) · API Gateway / MCP ·
  Message Broker (Kafka/Redis) · Database (Elasticsearch) · Calibration

Two ways skills are used — most workflows touch both:

• Deploy time — "Deploy VSS for video search." A skill selects the right profile or microservice, runs pre-flight checks, and brings up the Docker Compose stack. (vss-deploy-*, vss-setup-*, vss-generate-video-calibration)
• Runtime — "Add this camera," "summarize this clip," "show me today's incidents." Once VSS is running, a skill calls the live REST / MCP / VIOS APIs. (everything else)

Profiles vs. standalone microservices. A profile is a pre-assembled stack of microservices wired together for one workflow. Use vss-deploy-profile to bring up a whole workflow (base, search, lvs, alerts, warehouse, edge). Use the individual vss-deploy-* / vss-setup-* skills only when you need one microservice on its own.

Match the user's intent to a skill. Start here before opening any individual SKILL.md.

Skills chain. Skills auto-invoke each other when a prerequisite is missing — e.g. vss-deploy-detection-tracking-3d calls vss-generate-video-calibration when calibration data is absent. When a request spans layers (deploy a profile and add a camera and run a search), the agent composes several skills in sequence; the catalog below is grouped by layer so you can see what's adjacent.

Easy to confuse:

• vss-ask-video (one-off VLM question on a clip) vs. vss-search-archive (retrieval across an archive) vs. vss-query-analytics (read already-computed metrics/incidents — no live inference).
• vss-generate-video-report (formatted report from per-clip VLM or an incident range) vs. vss-generate-video-report-rag (the frag/RAG pipeline with HITL parameter collection).
• vss-deploy-profile (a whole workflow stack) vs. the vss-deploy-* / vss-setup-* skills (a single microservice).

Skills with evals/*.json specs are exercised automatically by the Skills Eval CI workflow on every PR that touches skills/**; legacy eval/*.json specs are still accepted for skills that have not moved yet. See .github/skill-eval/AGENTS.md for harness behavior.

The VSS 3.2 GA skill names replaced the pre-GA slash-command names:

Open this repository in your coding agent (Claude Code, Codex, Cursor, or any other agentskills.io-compatible host) and paste the following prompt:

Read skills/README.md and every SKILL.md file under skills/. For each skill in the catalog, install it for this host so I can invoke it from a shell or chat session. Use the host's standard skills directory:

• Claude Code: ~/.claude/skills/<name>/
• Codex: ~/.codex/skills/<name>/
• Hosts that follow the agentskills.io universal path: ~/.agents/skills/<name>/

Symlink each skill folder rather than copying it so a git pull here keeps every install up to date. Skip skills that are already installed and pointing at this checkout. When you're done, list the skills you registered and which directory you used.

The agent will read the frontmatter of each SKILL.md, create the symlinks, and confirm what's installed. The skills become invokable in the next agent session.

To install skills individually, paste the following prompt:

Install only skills/<name>/ for this host the same way.

After git pull, the symlinks already point at the updated content — nothing to do unless skills were added or renamed. To pick up new skills use the following prompt:

Re-read skills/README.md and add any new skills missing from this host's skills directory.

To uninstall skills, paste the following prompt:

Remove every VSS skill symlink you previously created under this host's skills directory.

This skills/ directory is the canonical source. Skills published to the public catalog at github.com/nvidia/skills are mirrored from here at sync time.