Level 2 of 14 · about 3 weeks

LLM APIs and structured output

Turn single model calls into dependable software: streaming, validated output and the first tool call.

Modules

1. 1

   Messages, roles and system prompts

   How a conversation is sent to a model, who keeps the state, and what a system prompt can and cannot enforce

   Core 6 parts · about 4 hours

2. 2

   Streaming

   Show tokens as they are produced, and handle cancelling, dropped connections and half-finished output

   Core 6 parts · about 4 hours

3. 3

   Structured output

   Turning model text into validated data your code can trust

   Core 6 parts · about 6 hours

4. 4

   Tool calling basics

   Describe a tool, read the model's request, run it safely, return the result

   Core 6 parts · about 6 hours

5. 5

   Reliability

   Retries with backoff, timeouts, idempotency, circuit breakers and fallbacks

   Core 5 parts · about 6 hours

6. 6

   Cost tracking and prompt caching

   Log every call's cost, cap a runaway run, and lay out prompts so the provider can reuse them

   Core 5 parts · about 5 hours

7. 7

   A thin provider layer

   One neutral interface, one adapter per provider, and a router that falls back

   Common 4 parts · about 4 hours

8. 8

   Local models

   Running a model on your own machine, sizing it, and talking to it from your app

   Specialist 5 parts · about 4 hours