Skip to main content

What are Agent Skills?

Agent Skills are packaged, reusable instructions that teach an AI coding agent how to handle a specific Newton task — for example, “evaluate the assembly steps shown in this video,” “embed these vibration windows and flag the anomalies,” or “scaffold a branded dashboard for the demo front-end.” Each skill is a folder containing a SKILL.md (plus optional runnable reference scripts and tests) that the agent loads on demand. Skills give agents the institutional knowledge you’d otherwise have to repeat in every prompt: which endpoint to call, what the request body looks like, and the empirical gotchas we’ve already verified against production. They’re the difference between “explain the Direct Query API” (you do the synthesis) and “run this clip through Newton and tell me if the assembly passed” (the agent already knows how). The canonical source is the archetypeai/agent-skills repository.

Install

The skills are published on skills.sh. One command detects your coding agent (Claude Code, Cursor, Codex, Copilot, and 20+ others) and installs to the right location:

Manual install (Claude Code)

Alternatively, copy the skills folder into Claude Code’s skills directory:
The SKILL.md format is portable: the same files work with any agent runtime that loads skills from a directory — see each agent’s docs for the directory it reads from.

Invoke

Once installed, invoke any skill in your agent with a slash command:

Available skills

Four groups: Agents run a maintained pipeline server-side over the Agents API, Models call Newton directly on /lens/query and leave the orchestration to you, Integrations pair a Newton model with a third-party model, and Design covers the demo front-end. The Model skills build their reference scripts on the official python client (pip install archetypeai), declared in each skill’s references/requirements.txt. The agent skills (OSM, RED, AD, Label-Free AD, Omega Encoder, MGA and TVA) run on Agent Lifecycle 2.0 (projects, packages, deployments and sessions) through a small standard-library HTTP helper of their own, since the client doesn’t cover those calls or the Optimizations, Evals and promote endpoints yet. Seven skills call no API: the six data-prep skills (atai-newton-omega-model-data-prep, and the OSM, RED, AD, Label-Free AD and Omega Encoder -data-prep skills) are local data processing, and atai-design-system scaffolds through the ds CLI. The Integrations skill uses the official client for its Newton calls and calls the partner model on its own API (for Nemotron, NVIDIA’s hosted API with its own NVIDIA_API_KEY).

Agents

Managed end-to-end pipelines — upload an input, run a platform blueprint, poll, download the output. No model to host. OSM, RED and AD can also be fitted to your own data on the platform (Optimizations → promote → Evals → a package, deployed in batch or for serving), each with its data-prep skill in front. The four sensor agents answer different questions, and the discriminator is what you already have: every regime labelled (OSM), a handful of examples of one named fault (RED), normal operation to learn from (AD), or nothing at all — no labels and no reference period (Label-Free AD).

Models

Direct Query API — one stateless POST /lens/query per request. Use these when you want control over the pipeline, or the raw vectors to build on.

Integrations

Newton paired with a third-party model: Newton does the sensing, the other model reasons over what Newton found. For running Newton inside a partner’s stack instead, see NVIDIA VSS with Newton Fusion.

Design

Example projects

These public repos demonstrate the patterns covered by the skills above — clone them as starting templates: Anomaly discovery end to end (atai-anomaly-discovery-agent):
  • anomaly-discovery-agent-example-tep-quickstart — the Tennessee Eastman Process (a simulated chemical plant, 33 process variables every 3 minutes) → an AD agent fitted on normal operation only: an Optimizations search, the best trial promoted to a blueprint, one Evals test on faults it has never seen, a package delivered in batch sessions over 84 more plant runs and scored against their held-back labels, then the same package deployed for serving, a window per query, and checked against the delivery. Includes the prepared data and role files, so you can skip the 1.4 GB download.
Label-free anomaly discovery (atai-label-free-anomaly-discovery-agent, with atai-label-free-anomaly-discovery-agent-data-prep for the files):
  • label-free-anomaly-discovery-agent-example-metropt-quickstart — a metro train’s air compressor (MetroPT-3, 7 sensors every 10 s) → the Label-Free AD agent in batch, with no labels and no reference period: a session per three-week file, the flagged hours turned into episodes described from the sensors, then graded against the operator’s maintenance reports. Both reported leaks are found, one 13 hours before the report. Includes the prepared files, so you can skip the download.
Manual generation end to end (atai-manual-generation-agent):
  • manual-generation-agent-example — a 3-minute tire-change video (CrossTask) → an ordered, timestamped manual from the MGA agent, with nothing to fit: a copy of mga with your prompt and token budget, a package run as a batch session per video, then the same package deployed for serving, a video per query, both scored against the video’s 11 reference steps (all 11 found in batch).
Omega embeddings from a deployed agent, your own detector (atai-omega-encoder-agent, with atai-omega-encoder-agent-data-prep for the role files):
  • omega-encoder-example-aursad-quickstart — a UR3e robot’s screwdriver (AURSAD, 12 signals at 100 Hz) → each tightening embedded by the Omega Encoder agent in batch, then a client-side kNN detector fitted on normal tightenings only, with its window and channels chosen on validation, one test, a delivery, and the same embeddings served one tightening per query and checked against batch, with a live app. Missing screws are caught 93% of the time. Includes the role files, so you can skip the 6.4 GB download.
Operational state monitoring end to end (atai-operational-state-monitoring-agent, with atai-operational-state-monitoring-agent-data-prep for the role files):
  • operational-state-monitoring-agent-example-larco-quickstart — washing-machine vibration (LARCO, 9 channels at 200 Hz) → an OSM agent fitted on the platform: an Optimizations search, the best trial promoted to a blueprint, one Evals test, then a package delivered in batch sessions to a second machine and scored against its held-back labels (fill / wash / spin / drain), and the same package served a window per query.
  • operational-state-monitoring-agent-example-larco — the full study behind the quickstart: all 199 cycles of two machines, a simple vibration-level baseline compared at every stage, a 16-trial search, then a test on new settings and delivery to the second machine (106 cycles). Includes the packed role files, so you can skip the data preparation.
Rare event detection end to end (atai-rare-event-detection-agent):
  • rare-event-detection-agent-example-pump-quickstart — a water pump’s sensors (a public Kaggle recording: a reading a minute for five months, seven breakdowns) → a RED agent fitted on the platform from two labelled breakdowns: an Optimizations search over the distance metric, the best trial promoted to a blueprint, one Evals test on four breakdowns it has never seen, a package delivered in batch sessions over two of them and scored against their held-back labels, then the same package deployed for serving, a window per query, and checked against the delivery. Includes the prepared data and role files, so you can skip the download.
Task verification end to end (atai-task-verification-agent):
  • task-verification-agent-example — twelve recordings of a three-step o-ring cap assembly (some with a step skipped) and the procedure as a text SOP → a PASSED / FAILED / MISSING verdict per step from the TVA agent, with nothing to fit: a copy of tva with your frame count and token budget, a package run as a batch session per recording, then the same package deployed for serving, a recording per query. Every clip’s verdicts are shown: the skipped wrench caught 3 of 3 times, the skipped o-ring (on screen anyway) passed 5 of 6 times. Includes the recordings.
Time-series embeddings + local KNN (atai-newton-omega-model, with atai-newton-omega-model-data-prep for the reference set):
  • swat-demo-direct-query — 6-stage water-treatment-plant anomaly dashboard: per-channel Omega embeddings + client-side KNN, with PCA-2/UMAP-2 embedding views and per-stage accuracy badges (plus Direct Query API text reasoning for operator suggestions).
  • wind-turbine-demo — Penmanshiel wind-farm SCADA: Omega embeddings + local KNN against a leakage-free n-shot library precomputed offline, detecting a real frequency-converter fault against a healthy peer turbine.
  • drilling-demo — drilling state classification from 14 North Sea (Volve) wells: Omega embeddings + local KNN against a leakage-free n-shot library built from held-out reference wells, with live accuracy vs ACTC ground truth.
Vision & text reasoning (atai-newton-fusion-model):
  • traffic-demo — live Caltrans CCTV traffic analysis; samples a burst of frames per interval and sends them as one multi-frame Direct Query API clip to reason over flow, incidents, and conditions.
  • wildfire-demo — wildfire detection across 1,200+ ALERTCalifornia cameras with Newton vision (per-camera smoke/fire/haze + zone Q&A).
  • earthquake-demo — real-time USGS feed → text reasoning over structured seismic data (aftershock sequences, clustering, regional-risk ranking).
  • grid-demo — California power grid (CAISO supply/demand) → text reasoning on duck-curve dynamics, ramp, renewable share, and grid-stress risk.
  • wifi-demo — “is anyone home?” occupancy inference from residential WiFi device-telemetry (GHOST-IoT) — per-window verdict + per-device classification shown against ground truth.
Omega + NVIDIA Nemotron (atai-newton-omega-nvidia-nemotron, on top of atai-newton-omega-model):
  • swat-demo-nemotron — the SWaT water-treatment dashboard with NVIDIA Nemotron writing the operator suggestions: upstream / local / downstream action cards as JSON, checked in code against the plant topology and each stage’s equipment, cached per anomaly set.
  • wind-turbine-demo-nemotron — the Penmanshiel monitor with NVIDIA Nemotron fault briefs: when Omega flags a turbine, Nemotron compares the window with the healthy peer over the same hours. It pointed to the pitch system for a real 23-hour pitch-circuit stop, from SCADA alone.
If you build something with these skills, open a PR against agent-skills to share your patterns — the empirical gotchas in each SKILL.md came from real production runs and we’d love more.