agent-skills

Build the front-end for a Newton demo with the Archetype AI Design System — the published, versioned packages (`@archetypeai/ds-lib-tokens`, `@archetypeai/ds-ui-svelte-console`, `@archetypeai/ds-ui-svelte-labs`) and the `ds` scaffolding CLI (`@archetypeai/ds-cli`) — instead of hand-rolling tokens, components, fonts, or brand styling. Use this skill when the user wants to scaffold a new dashboard/UI, add the design system to an existing SvelteKit app, or pull in branded primitives (menubar, logo, sensor/scatter charts, video player, card, badge, table, dialog…). The CLI installs the design system's OWN agent config — `CLAUDE.md` (or `AGENTS.md` for Cursor) plus `ds-manifest.json` at the project root — which is the source of truth for component usage, fonts, fallback behavior, and styling once scaffolded; this skill's only job is to get you there. Components currently ship for Svelte 5; a React port is in progress. Do NOT use for Newton API / backend work (see the `atai-newton-*` skills).

Clean, split, and featurize multivariate time-series data before embedding it with the Omega model. Bundles three reusable building blocks: `DataPreprocessor` (diagnose timestamp regularity, gaps, and nulls; build continuous gap-aware temporal blocks with imputation), `DataSplitter` (out-of-time or random train/test split that respects temporal order), and `FeaturePreparer` (pivot per-sensor embeddings into a "joint state" (X, y, metadata) matrix with optional L2 / standardization / PCA). Use this skill when the user is preparing raw sensor CSVs for the Omega model (`atai-newton-omega-model`), when training data has gaps and they're unsure whether to drop / impute / split, or when an n-shot CSV looks noisy and they want a principled cleanup pipeline. Do NOT use this skill to run inference or produce embeddings (use `atai-newton-omega-model`). Do NOT use for video / image / text data (use `atai-newton-fusion-model`) — time-series only.

Get time-series embeddings from Archetype AI's Omega encoder (`OmegaEncoder::omega_embeddings_1_4`) over the prod `/query` endpoint — send a window of sensor readings, get back a fixed-size vector per channel, no batch job and no session. Use this skill when the user wants to embed multivariate sensor windows (vibration, flow, pressure, network, etc.) for lightweight downstream ML — KNN classification, anomaly scoring, similarity search, or 2D projection — done client-side over the embeddings. Covers the request shape (`data.numeric_array`, channel-first window), the per-channel 768-d output, the supported 16–1024 window-length range, `normalize_input`, and the "joint multi-channel state" + KNN pattern. For cleaning / splitting / windowing raw sensor CSVs first, see `atai-newton-omega-model-data-prep`. Do NOT use for text / image / video reasoning (that's the Newton fusion model on `/query`).

Call Newton's C 2.6 fusion model (`Newton::c2_6_8b_fp8_260424d7a55d5e`) on the prod `/query` endpoint with text, image, and video inputs in a single stateless request — no session, no batch job, no SSE plumbing. Use this skill when the user wants per-call multimodal reasoning (describe an image, summarize a clip, classify a state from sensor + visual context) without managing session lifecycle, or when they need the Newton C checkpoint that reasons over video frames via `/query`. Covers the request shape per modality, the two image-attachment paths (file_ids vs base64), multi-image mode, both video paths (`.mp4` + `max_frames`, and client-sampled frames + `query_metadata`), generation parameters, JSON-output prompting, latency budgets, and the C 2.6 identifier gotcha. Do NOT use for streaming / session-based activity monitoring, large multi-file batch jobs, or time-series embedding (KNN / anomaly) classification.