Package repeatable procedures so ChatGPT loads the right instructions when a task matches.

Skills as folders with manifests, optional scripts, and progressive loading for automation workflows.

Anthropic’s field guide: start simple, prefer workflows with clear gates, and add autonomy only when measurement proves you need it.

How Claude emits tool calls, how to shape schemas the model can follow, and how to close the loop with tool results in context.

Constrain model generations to JSON schemas so parsers stop being the weakest link in your stack.

A practical layer for streaming chat UIs, tool calls, and swapping models without rewriting your app shell every quarter.