development-ledger— Preserve development context from research through implementation and verification.human-interface-design— Design and review humane, platform-native interfaces using Apple HIG principles.orchestrate-long-running-tasks— Coordinate delegated Codex tasks with model-aware routing, verification, and periodic supervision. Based on this tweet.ml-research— Drive ML research as an iterative loop with experiment hygiene, a research log, and primary-source evidence, including mechanistic interpretability.ml-compute— Choose where and how to run ML training and evals (Tinker, Prime Intellect, Modal, RunPod, local): recipes, RL rewards, lm-eval and calibration, cost. Distils parts of Orchestra AI-Research-SKILLs (MIT).mech-interp— Mechanistic interpretability methods and tools: Jacobian subspaces, ablation, patching, steering, SAEs, nnsight/TransformerLens/pyvene/SAELens. Distils parts of Orchestra AI-Research-SKILLs (MIT).tool-harness-audit— Audit MCP and agent tool harnesses for agent ergonomics and reliability.
npx skills add esceptico/skills