System One

Build software around small typed AI judgments used as programming primitives — a model call returning a `choice`, an ordered `score`, or a 0–1 probability your code branches on, instead of prose you prompt-and-parse. Covers judgment design, batching questions into one call, composing answers in code, and threshold calibration. Backend is TypeSafe's System One API (model Jev; Choice / Score / Noul), whose live docs are read per task. Use when writing or debugging a TypeSafe or Jev client, service or integration; when a feature needs programmable common sense — routing, ranking, reranking, extraction, moderation, guardrails on another model's input or output, verification, ML feature extraction, knowledge-graph annotation; when an LLM prompt-and-parse step should become a structured decision; or when the user names TypeSafe or Jev. The model itself never writes prose or code. Not before the egress decision in references/data-boundary.md; not for exact lookups, arithmetic or rules code already decides.

dreamlab-ai Updated 23 repo stars
Conceptual modular agent workshop with connected workspaces

File contents

dreamlab-ai/agentbox commit aecbc0df4b

Frequently asked questions

npx get-skill add dreamlab-ai/system-one