Eight topics — from assembling a prompt to production. Every topic has section quizzes and a quiz on the whole topic. Not sure where to start? Take the diagnostic: 32 questions across all topics will show what to brush up on.
How the prompt the model sees is built, and how to assemble an agent's context: message roles, the system prompt, examples, a stable prefix for the cache, compressing a long history. Separately — model behavior from the user's point of view: tokens, temperature, hallucinations, reasoning models.
How to give a model the knowledge it needs, and how an agent remembers what matters between sessions: chunking documents, embeddings and keyword search, reranking, agentic RAG, knowledge graphs, long-term memory — what to remember, how to update it and how to forget.
How the agent loop works and how an agent acts in the world: tool descriptions, MCP, skills, what to do when there are too many tools, planning, reflection, error handling.
When a rigid script is enough and when autonomy is needed, how to split work between agents, pass context along without losing it at the seams, and where a human in the loop is needed.
How to work with agents that write code without fooling yourself about the result: a spec instead of an assignment, rules files and skills, hooks as deterministic guarantees, subagents, independent verification and review.
How to tell whether an application got better after a change: evaluation datasets, LLMs as judges and their biases, agent reliability, RAG evaluation, statistical significance and observability in production.
How LLM applications and agents become vulnerable, and why protection is built with architecture rather than a prompt: trust boundaries, direct and indirect injections, data leaks, excessive agent permissions, attacks on the budget.
What an application needs after launch: input and output guardrails, choosing and routing models, caching, cost and latency, structured output, degradation tests, monitoring and working with feedback. And an architecture-level decision: a prompt, RAG or fine-tuning.