GenAI 101: Building With LLMs

Syllabus outline

1. 01

   Foundations

   Text in, text out as the entire function signature, and the list of things a model cannot do without a tool. A model as an equation: architecture against weights, training as a search for the numbers, a neuron and a network of them, and the same idea scaled into a transformer. One token at a time drawn from a distribution, so identical input returns different output. Hallucination as the mechanism rather than a fault. Context as everything visible this turn, the app as distinct from the model, the two meanings of "agent", and three model tiers rather than a leaderboard.

2. 02

   Your Workbench

   Notebooks: cells, execution order, and prose beside code; Jupyter, Colab, and the one on screen. Provider-direct keys against an aggregator, and the trade in each direction. Environment variables, and why a key pasted into a cell is a key published. A first call, then the same call against a second provider. Cost arithmetic from response.usage: input tokens, output tokens, and the real price of a thousand calls.

3. 03

   Prompting as an API Skill

   The three things a chat box hides. System, user and assistant roles, with the system prompt as the largest lever a developer gains. Few-shot examples for format-sensitive work. Reasoning effort rather than temperature, swept low to high with cost printed alongside. Chain-of-thought as a judgement call now that reasoning models have internalised it. Weak and strong prompts compared on screen, with both outputs.

4. 04

   Making the Model Do Work

   Structured outputs: a schema in, valid JSON out, and an LLM callable from ordinary code without a parsing layer. Tool use and function calling as a two-turn exchange: the model emits a call, the program executes it, the result goes back, the model answers. The primitive the rest of the course is built on.

5. 05

   The Solution Shapes

   Embeddings as text reduced to numbers, where proximity means similarity. Vectors, and what "1,536 dimensions" actually states. Vector databases, and when a plain array is genuinely enough. Retrieval-augmented generation end to end: chunk, embed, retrieve, assemble the prompt, answer from grounded text. Fine-tuning against RAG against prompting as answers to three different questions. Cosine similarity over three sentences.

6. 06

   Extending Your Agent

   A skill as a folder holding a SKILL.md: frontmatter, a description, and instructions in markdown. Why a file rather than an API call: reviewable, version-controlled, portable. AGENTS.md as the standing repository context every agent reads. MCP as one agreed socket shape, so any tool reaches any service without a bespoke integration per pair. Connecting an existing server, and reading a minimal one.

7. 07

   Shipping It Safely

   Prompt injection demonstrated against an agent holding a tool, then mitigated: delimiters, instruction hierarchy, output filtering, least-privilege tools. A minimal evaluation over three cases as a pass/fail table. Rate limits and cost at production volume rather than workshop volume. What to build first, by role.