LLMs inside real systems, not demos: prompting techniques and what each costs per call, generation parameters and what they actually act on, spec-driven development with AI agents. Documented behaviour is kept separate from opinion — and the opinion is labelled.