adamdroberts

Use when editing the Meta Watcher operator UI under `web/src/` — Astro components, `app.ts`, or `global.css` — or when rebuilding the committed static bundle at `meta_watcher/web/static/`.

Use when modifying Meta Watcher's SAM 3.1 providers — the MLX subprocess path, the Linux CUDA path, the `sam3.perflib` fp32 patch, or the `InferenceProvider` interface. Trigger on edits to `meta_watcher/inference.py`, or when the user asks about provider swapping, model IDs, GPU dtypes, or the MLX worker.

Use when modifying Meta Watcher's capture, tracking, state machine, recorder, or overlay logic. Trigger on changes to `meta_watcher/core.py`, `pipeline.py`, `overlay.py`, or `sources.py`, or when the user asks to change modes, recording behavior, detection filtering, or frame flow.

Use when adding or changing Meta Watcher HTTP endpoints, `RuntimeState` methods, MJPEG/JPEG behavior, or any API consumed by the Astro operator UI. Trigger on edits to `meta_watcher/web/server.py` or `meta_watcher/web/state.py`.

Use or modify the NeuralFn nfn CLI for train, infer, eval, dataset shortcuts, tokenizer selection, artifact paths, fine-tuning flags, CUDA harness scripts, and CLI tests. Use when the user mentions nfn, cli/, train scripts, pretraining files, cached datasets, tokenizers, or graph/weights artifacts.

Build neural network graphs programmatically using the neuralfn Python package. Use whenever the user asks to write Python code that imports neuralfn, creates neurons, builds graphs, wires edges, trains models, serializes graphs, or works with the NeuralFn graph framework directly in code. Do NOT use for MCP tool calls -- see neuralfn-mcp instead.

Build, edit, train, and inspect NeuralFn neural-network graphs using the neuralfn MCP server tools. Use whenever the user asks to create a neural network, build a model (GPT, MoE, Llama, NanoGPT, etc.), train on a dataset, add or wire neurons, execute a graph, manage datasets, or do anything with NeuralFn. The MCP server name is "neuralfn".

Build, train, and export torch-backed neural network models (GPT, Llama, MoE, Jamba, JEPA, diffusion, etc.) using the NeuralFn Python API. Use whenever the user asks to build a language model, train a transformer, use template presets, configure ModelSpec/BlockSpec, compile a torch graph, export weights, or do autoregressive inference with NeuralFn in Python code. For MCP tool operations, use neuralfn-mcp instead.

Create, expand, reorganize, and maintain repository documentation with deep coverage across README files, browsable docs, API/reference pages, llms.txt, llms-full.txt, changelogs, and repo-local agent skills. Use whenever the user asks for comprehensive project documentation, documentation refreshes, LLM-friendly docs, documentation synchronization after code changes, or agent guidance that teaches future LLMs how to use the project docs.

Enforces strict change transparency. Use when editing code, notebooks, configs, or running commands so every change is disclosed, including changes not announced in advance.

Procedural skill for adding a new TypeScript language feature or stdlib method to the typescriptc transpiler. Use when a user asks to add a built-in method, operator, syntax construct, or Node stdlib API to tsc2c.

Procedural skill for adding a new TypeScript language feature or stdlib method to the typescriptc transpiler. Use when a user asks to add a built-in method, operator, syntax construct, or Node stdlib API to tsc2c.

Carry the in-flight port of llm.c onto ThunderKittens forward. Use whenever the user asks to implement the next milestone, port a kernel, add a TK wrapper, or modify llmc/* or train_*.cu in this repo.

Drive the llm.kittens H100 validation harness — pick the right phase, set the required env vars, replay captured evidence, and avoid `goal-complete` mistakes. Use whenever the user asks to validate the goal, run a phase of `scripts/validate_goal_h100.sh`, replay an existing log, or check whether a milestone gate would pass.

Use when touching scripts/build-{deb,rpm,pacman}.sh, scripts/lib/package-common.sh, the claude-update-manager Rust crate under updater/, packaging/linux/ templates and maintainer scripts, or the /opt/claude-desktop/update-builder/ rebuild bundle. Enforces the staging-helper / REQUIRED_BUNDLE_FILES dual-update rule, the SemVer policy for the updater crate, the pkexec privilege boundary, and the maintainer-script contract across deb/rpm/pacman.

Use when adapting a new macOS Claude.dmg (or similar Electron app DMG) into the Linux pipeline in this repo, when a fresh upstream release breaks the asar patches, or when diagnosing a Claude Desktop launch failure on Linux via ~/.cache/claude-desktop/launcher.log.