Prompting & LLM APIs

What you'll learn

1. 1

   Anatomy of an LLM API Call

   Roles, messages, and the fact that models remember nothing between calls.

   • A chat is just a list
   • Watch: why chatbots "forget"
   • Find the rules
   • Complete the request
   • Build the messages list
   • Debug the forgetful bot

   Take-home template: Production LLM Client (Anthropic + OpenAI)

2. 2

   Streaming Responses

   Make AI apps feel instant: server-sent events, deltas and stop reasons.

   • Why streaming feels faster
   • Find the text
   • Parse raw SSE
   • Rebuild the message
   • Streaming in your backend

   Take-home template: Streaming Chat Endpoint (FastAPI + SSE)

3. 3

   Prompt Engineering That Works

   Clear instructions, examples, delimiters and output formats.

   • Treat the model like a brilliant new hire
   • Which prompt is better?
   • Build a few-shot prompt
   • Why delimiters?
   • The leaked system prompt

   Take-home template: Production Prompt Template

4. 4

   Reasoning Models & Chain of Thought

   When to pay for thinking, how to prompt reasoning models, and tricks that boost accuracy.

   • Models that think before they answer
   • Pick the right model
   • Separate thinking from the answer
   • Self-consistency: let answers vote
   • Prompting a reasoning model

   Take-home template: Self-Consistency Voting Wrapper

5. 5

   Structured Output (JSON)

   Get machine-readable output and handle it safely when it breaks.

   • Your code needs data, not prose
   • Watch: catch a bad JSON reply and retry
   • Parse messy model output
   • A robust JSON pipeline
   • Production incident

   Take-home template: Validated JSON with Pydantic + Retry

6. 6

   Tool Use (Function Calling)

   Let the model call your code: search, databases, APIs.

   • The model asks, your code acts
   • Read a tool call
   • Write a tool definition
   • One tool round trip
   • Dispatch a tool call
   • Designing good tools

   Take-home template: Tool Registry with Auto-Generated Schemas